Time Series Analysis

Chapter 14: Multivariate GARCH models

Chapter 14: Multivariate GARCH models

Daniel Traian PELE

Bucharest University of Economic Studies
IDA Institute Digital Assets
Blockchain Research Center
AI4EFin Artificial Intelligence for Energy Finance
Romanian Academy, Institute for Economic Forecasting
MSCA Digital Finance

Acronyms used in this material

Slide 1 in the PDF
  • AI: Artificial Intelligence
  • BEKK: Baba–Engle–Kraft–Kroner (multivariate GARCH model)
  • BET: Bucharest Exchange Trading (index)
  • BNR: Banca Națională a României (the National Bank of Romania)
  • CCC: Constant Conditional Correlation (model)
  • CEE: Central and Eastern Europe
  • CET: Central European Time
  • COVID-19: COronaVIrus Disease 2019
  • DCC: Dynamic Conditional Correlation
  • EODHD: EOD Historical Data (market-data provider)
  • EU: European Union
  • FHS: Filtered Historical Simulation
  • GARCH: Generalised AutoRegressive Conditional Heteroskedasticity
  • LR: Likelihood Ratio (test)
  • MGARCH: Multivariate GARCH
  • OLS: Ordinary Least Squares
  • QML: Quasi-Maximum Likelihood
  • SE: Standard Error
  • US: United States
  • UTC: Coordinated Universal Time (Temps Universel Coordonné)
  • VAR: Vector AutoRegression
  • VEC: Vector (half-vectorised) GARCH model
  • VECM: Vector Error Correction Model

Why this chapter

Slide 2 in the PDF

The question of the chapter is how the risks of several assets move together. In Chapter 5 the conditional variance of a single return series was enough to describe its risk. For a portfolio of two or more assets this is no longer enough: two portfolios with the same individual volatilities can have very different risk, depending on whether their components tend to fall on the same days or on different days. The quantity that captures this is the covariance, or its standardised version, the correlation, and the chapter treats both as variables that change from day to day.

The route goes from the general to the practical. It starts from the conditional covariance matrix \(\mathbf{H}_t\) and the two conditions it must satisfy, continues with the VEC and BEKK models, which describe all elements of \(\mathbf{H}_t\) at once, and arrives at the CCC and DCC models, which estimate volatilities and correlations separately. The last sections use these models for three applications: correlations in crises, the VaR of a portfolio and hedge ratios.

The chapter builds directly on three earlier ones. The univariate GARCH model of Chapter 5 is the building block of every multivariate model; the vector autoregression (VAR) of Chapter 6 supplies the multivariate notation and the idea of spillovers between series; cointegration, from Chapter 7, is the long-run counterpart of the short-run links studied here.

In brief

  • Question: how do the risks of several assets move together? Chapter 5 modelled the volatility of one series at a time
    • the risk of a portfolio, a hedge or the spread of a crisis depends on covariances and correlations, and these change in time
  • Route of the chapter
    • the conditional covariance matrix \(\mathbf{H}_t\) and its conditions
    • the VEC and BEKK models; the number of parameters
    • CCC and DCC: volatilities and correlations estimated in two steps
    • applications: correlations in crises, portfolio VaR 1%, hedge ratios
  • Self-study chapter
    • read the slides in order, redo the worked examples, run the notebook, then answer the self-assessment and the quiz
    • prerequisites: Chapter 5 (GARCH), Chapter 6 (VAR models), Chapter 7 (cointegration)

Learning outcomes

Slide 3 in the PDF

The first outcomes are conceptual. The reader should be able to explain why the variance of a portfolio, a hedge ratio and the question of contagion all depend on covariances that change over time, and to write the main multivariate GARCH models in matrix form. Counting the parameters of each model is part of understanding it: the count shows immediately which models can be estimated for five assets and which cannot.

The other outcomes are computational and empirical. One step of the DCC recursion can be computed by hand, and its parameters \(a\), \(b\) and \(a+b\) have a direct reading as the reaction to news, the persistence and the speed of return to the long-run correlation. A DCC model is estimated in two steps and compared with constant correlations by a likelihood ratio test. Finally, \(\mathbf{H}_t\) is turned into two practical numbers, the VaR 1% of a portfolio, judged by backtesting, and the minimum-variance hedge ratio, judged by its hedging effectiveness out of sample.

In brief

  • Explain why portfolio risk, hedging and contagion need a model for time-varying covariances
  • Write the VEC, BEKK, CCC and DCC models and count their parameters
  • Compute one step of the DCC recursion by hand and interpret \(a\), \(b\) and \(a+b\)
  • Estimate a DCC model in two steps and test it against constant correlations
  • Compute and backtest a portfolio VaR 1% from \(\mathbf{H}_t\)
  • Compute dynamic minimum-variance hedge ratios and their hedging effectiveness

Reading and tools

Slide 4 in the PDF

Huang and Petukhina (2022) give the background on GARCH models and their implementation in Python. Two surveys cover the multivariate models in more depth: Bauwens, Laurent and Rombouts (2006) classify them and discuss estimation and testing, and Silvennoinen and Teräsvirta (2009) give a compact overview with the conditions for positive definiteness. Each model of the chapter also has its original paper: Bollerslev, Engle and Wooldridge (1988) for VEC, Engle and Kroner (1995) for BEKK, Bollerslev (1990) for CCC and Engle (2002) for DCC.

The chapter's code uses the arch package for the univariate GARCH models of step 1. The package has no DCC model, so step 2 is written out in about 40 lines of numpy: the recursion for \(\mathbf{Q}_t\), the rescaling to correlations and the correlation part of the likelihood. Writing the recursion explicitly has a teaching advantage, since every element of the model is visible in the code. The lecture notebook reproduces the computations and the charts of the chapter.

In brief

Chapter 5 modelled the volatility of one return series at a time with generalised autoregressive conditional heteroskedasticity (GARCH) models. Most questions of risk management, however, concern several assets at once. The risk of a portfolio depends on how its components move together, a hedge works only as well as the link between the asset and the hedging instrument, and the spread of a crisis from one market to another is a statement about co-movement. All three questions are answered by covariances and correlations, and these change over time.

This chapter extends GARCH to \(N\) series. The object of interest is the conditional covariance matrix \(\mathbf{H}_t\), which holds the variances of all assets on its diagonal and their covariances off the diagonal. The chapter first sets out the conditions that any model for \(\mathbf{H}_t\) must respect, then presents the general VEC and BEKK models and the reason they become unusable beyond a few assets. The constant conditional correlation (CCC) and dynamic conditional correlation (DCC) models solve this by separating volatilities from correlations and estimating them in two steps.

The applications use daily data on the S&P 500, the German DAX index, the Romanian BET index, the EUR/RON exchange rate and Bitcoin. The correlation between the S&P 500 and the DAX is not constant: the likelihood ratio test against constant correlation gives 131.8, far above the 5% critical value of 5.99. Correlations rise in crises, most of all for the assets that are weakly linked to the large markets in calm times. The same conditional covariances give a portfolio Value at Risk (VaR) at 1% that adapts to the state of the market and dynamic hedge ratios that remove more risk out of sample than a static ratio.

The one-step-ahead DCC correlation of daily S&P 500 and DAX returns, the 250-day rolling correlation and the constant CCC value, 4 January 2000 – 18 September 2026.
The one-step-ahead DCC correlation of daily S&P 500 and DAX returns, the 250-day rolling correlation and the constant CCC value, 4 January 2000 – 18 September 2026.

What you will be able to do

  • Explain why portfolio risk, hedging and contagion need a model for time-varying covariances.
  • Write the VEC, BEKK, CCC and DCC models and count their parameters.
  • Compute one step of the DCC recursion by hand and interpret \(a\), \(b\) and \(a+b\).
  • Estimate a DCC model in two steps and test it against constant correlations.
  • Compute a portfolio VaR 1% from \(\mathbf{H}_t\) and backtest it with the Kupiec and Christoffersen tests.
  • Compute dynamic minimum-variance hedge ratios and their hedging effectiveness.

Chapter plan

  1. Why multivariate volatilityPortfolio variance, hedging and contagion; rolling correlations and asynchronous trading.
  2. The multivariate GARCH frameworkThe conditional covariance matrix, symmetry, positive definiteness and the likelihood.
  3. VEC and BEKK modelsThe general models, volatility spillovers and the number of parameters.
  4. Constant and dynamic conditional correlationsThe decomposition into volatilities and correlations, the DCC recursion and two-step estimation.
  5. DCC for the S&P 500 and the DAXTwo GARCH models, the estimated DCC and correlations in the crises of 2008 and 2020.
  6. Five markets: S&P 500, DAX, BET, EUR/RON, BitcoinA five-asset DCC and correlations in calm times and in the 2020 crash.
  7. Portfolio VaR 1% with DCCFrom \(\mathbf{H}_t\) to the VaR of a portfolio and its out-of-sample backtest.
  8. Dynamic hedge ratiosThe minimum-variance hedge ratio and the hedging of the BET with the DAX.
  9. Links with other chapters and limitsGARCH, VAR and cointegration; the limits of DCC and its extensions.
  10. Possible contribution of AIWhat an assistant can write and what must be checked.
  11. SummaryKey ideas, formulas and self-assessment questions.

Why multivariate volatility

The section shows why a model for a single volatility is not enough. Portfolio risk, hedging and contagion all depend on correlations, and the correlations observed in real markets move a great deal over time. The section ends with a practical trap of daily data: markets that close at different hours.

Three questions that need covariances

Slide 5 in the PDF

The first question is the risk of a portfolio. If the vector \(\mathbf{w}\) holds the weights of the assets and \(\mathbf{H}\) is the covariance matrix of their returns, the variance of the portfolio return is a quadratic form, as in Markowitz (1952):

\[\begin{gathered} \sigma_p^2 = \mathbf{w}^\top\mathbf{H}\mathbf{w}, \\ \text{for two assets: } \sigma_p^2 = w_1^2\sigma_1^2 + w_2^2\sigma_2^2 + 2w_1w_2\rho\,\sigma_1\sigma_2. \end{gathered}\]
(14.1)

The first two terms come from the individual volatilities; the third is the only place where the joint behaviour enters. With positive weights, a lower correlation \(\rho\) means a lower portfolio variance, and this is the whole mechanism of diversification.

The second question is hedging: how many units of one asset to sell in order to offset the risk of another. Section 7 shows that the answer is the ratio of a covariance to a variance, so a hedge ratio moves whenever the covariance moves. The third question is contagion: do markets move more closely together in a crisis? Longin and Solnik (2001) found that correlations between international equity markets are higher in bear markets. If correlations rise exactly when volatility rises, diversification weakens at the moment it is needed most.

A univariate GARCH model gives the volatilities \(\sigma_{1,t}\) and \(\sigma_{2,t}\) day by day. Formula (14.1) also needs the correlation, and the aim of this chapter is to add a correlation \(\rho_t\) that changes over time in the same way.

In brief

  • Portfolio risk
    • the variance of a portfolio with weights \(\mathbf{w}\) is \(\sigma_p^2 = \mathbf{w}^\top\mathbf{H}\mathbf{w}\) Markowitz (1952)
    • \(\mathbf{w}\): the vector of portfolio weights; \(\mathbf{H}\): the covariance matrix of the returns; \(\sigma_p^2\): the portfolio variance
    • two assets: \(\sigma_p^2 = w_1^2\sigma_1^2 + w_2^2\sigma_2^2 + 2w_1w_2\rho\,\sigma_1\sigma_2\), with volatilities \(\sigma_1\), \(\sigma_2\) and correlation \(\rho\)
    • the correlation \(\rho\) decides how much diversification helps
  • Hedging
    • how many units of an asset to sell to offset the risk of another
    • the answer is a ratio of a covariance to a variance (Section 7)
  • Contagion
    • do markets move more closely together in a crisis?
    • if correlations rise exactly when volatility rises, diversification fails when it is needed most Longin and Solnik (2001)
  • A univariate GARCH (Chapter 5) gives \(\sigma_{1,t}\) and \(\sigma_{2,t}\); this chapter adds the time-varying \(\rho_t\)

Worked example: correlation and portfolio risk

Slide 6 in the PDF

Take two assets with an annual volatility of 20% each and a portfolio split 50/50. With volatilities in percent, \(\sigma_1^2 = \sigma_2^2 = 400\) and \(w_1^2 = w_2^2 = 0.25\), so the first two terms of (14.1) add up to 200 whatever the correlation. With \(\rho = 0.2\) the covariance term is \(2 \cdot 0.25 \cdot 0.2 \cdot 400 = 40\), the variance is 240 and the volatility \(\sqrt{240} = 15.5\%\). With \(\rho = 0.8\) the covariance term is 160, the variance 360 and the volatility 19.0%.

Nothing changed except the correlation: the same assets, the same weights, the same individual volatilities. Yet the risk of the portfolio rises by more than a fifth. The practical consequence is that a risk model which keeps \(\rho\) at the value it had in a calm period understates the risk when correlations rise, and the data below show that they do rise in crises.

In brief

  • Two assets with annual volatility 20% each, weights 50% and 50%
  • \(\rho = 0.2\): \(\sigma_p^2 = 0.25 \cdot 400 + 0.25 \cdot 400 + 2 \cdot 0.25 \cdot 0.2 \cdot 400 = 240\), so \(\sigma_p = 15.5\%\)
  • \(\rho = 0.8\): \(\sigma_p^2 = 200 + 2 \cdot 0.25 \cdot 0.8 \cdot 400 = 360\), so \(\sigma_p = 19.0\%\)
  • The same assets and weights: the risk of the portfolio rises by more than a fifth only because the correlation rose
    • a risk model that keeps \(\rho\) fixed at its calm-period value underestimates the risk in a crisis

Portfolio volatility against the correlation

Slide 7 in the PDF

Worked example

\[\sigma_p^2=w_1^2\sigma_1^2+w_2^2\sigma_2^2+2w_1w_2\rho\,\sigma_1\sigma_2,\qquad w_1=w_2=0.5\]
      Question for the room ▶ raises ρ from 0.3 to 1 in steps of 0.1. Does σp rise by the same amount at every step?
      Answer

      No: +0.63 points from ρ = 0.2 to 0.3 (from 15.49% to 16.12%), but only +0.51 points from 0.9 to 1 (from 19.49% to 20%). σp2 is linear in ρ and σp is its square root, so the curve is concave. With ρ = −0.2 the volatility is 12.6%.

        Question for the room ▶ moves the weight of the first asset from 60% to 90% (both assets at 20%). What happens to σp for ρ = 0.8?
        Answer

        It rises: 19.02% at 60% and 19.64% at 90%, against 18.97% at 50/50. With equal volatilities the 50/50 portfolio has the lowest risk whatever ρ. With the second asset at 40% and ρ = −1, σp = |0.5 · 20 − 0.5 · 40| = 10%: the risk disappears only with w = 2/3.

        The chart plots the annual volatility of the 50/50 portfolio, in percent, on the vertical axis against the correlation \(\rho\) on the horizontal axis, from \(-1\) to \(1\), with both assets at 20% volatility. The two marked points are the cases of the worked example: 15.5% at \(\rho = 0.2\) and 19.0% at \(\rho = 0.8\).

        The curve is the square root of a linear function of \(\rho\): for this portfolio \(\sigma_p^2 = 200 + 200\rho\). It starts at zero for \(\rho = -1\), rises steeply at first and then more slowly, and reaches 20% at \(\rho = 1\). Reading the chart from right to left gives the benefit of diversification, the vertical distance between the curve and the 20% level at each correlation.

        In brief

        • A 50/50 portfolio of two assets with 20% volatility each; the points are the two cases of the worked example
        TSA_ch14_correlations

        Interpreting the diversification curve

        Slide 8 in the PDF

        Three points of the curve are worth remembering. At \(\rho = 1\) the two assets move as one, there is no diversification and the portfolio has the 20% volatility of each asset. At \(\rho = 0\) the covariance term disappears and the volatility is \(20/\sqrt{2} \approx 14.1\%\). At \(\rho = -1\) the movements of the two assets cancel exactly and the risk disappears, a case that does not occur with real equity indices.

        Because the curve is concave, the same increase of the correlation adds more volatility at low correlations than at high ones. For portfolios that combine weakly correlated assets, the very portfolios built for diversification, a rise in correlation is therefore most costly. This leads to the empirical question of the chapter: does \(\rho\) stay put in real markets, or does it move with the state of the market?

        In brief

        • \(\rho = 1\): no diversification, \(\sigma_p = 20\%\); \(\rho = 0\): \(\sigma_p = 20/\sqrt{2} \approx 14.1\%\); \(\rho = -1\): the risk disappears
        • The curve is concave: the same rise of \(\rho\) adds more volatility at low correlations than at high ones
        • The question of the chapter: does \(\rho\) stay put in real markets, or does it move with the state of the market?

        Correlations change over time

        Slide 9 in the PDF
          Question for the room ▶ sums the daily returns over k = 2, …, 10 days. What happens to the S&P 500 – BET correlation?
          Answer

          It rises: 0.32 daily, 0.40 over 2 days and 0.50 over 10 days, above the weekly correlation of 0.46: news that reaches Bucharest a day later falls into the same k-day return. Of the ten pairs, only the S&P 500 and Bitcoin have a weekly correlation below the daily one (0.25 → 0.19): the weekly Bitcoin return also contains the weekend, and the 610 weeks give a noisier estimate than the 2,815 days.

            Question for the room ▶ lengthens the window up to 500 days. What happens to the minimum of the S&P 500 – DAX correlation (0.21 at 250 days)?
            Answer

            It rises to 0.30 (November 2025): in a long window the asynchronous days of April 2025 weigh less. With 60 days the correlation goes down to −0.27 (July 2025) and up to 0.86 (January 2012), with a standard deviation of 0.147 against 0.122 at 250 days.

            A first answer comes from rolling correlations (rolling windows): the sample correlation of daily log returns computed over the last 250 trading days, about one year, and dated at the end of the window. The chart plots this correlation on the vertical axis against time on the horizontal axis for three pairs: the S&P 500 and the DAX, from 2001; the BET and the DAX, from 2001; the S&P 500 and Bitcoin, from 2015. The shaded bands mark September 2008 – March 2009 and February – April 2020, and a dashed line marks zero.

            Each correlation uses only the days on which both markets trade, so that each pair of returns refers to the same calendar day. A rolling correlation is a descriptive tool, not a model: it gives every day in the window the same weight and forgets everything older than 250 days. It is useful here because it shows, without any assumption about dynamics, whether the correlation is stable.

            In brief

            • Correlation of daily log returns over rolling windows of 250 trading days (about one year), dated at the window end
              • returns on the days on which both markets trade; shaded: September 2008–March 2009 and February–April 2020
            TSA_ch14_correlations

            Interpreting the rolling correlations

            Slide 10 in the PDF

            None of the three correlations is stable. The S&P 500 and the DAX range between 0.21 in April 2026 and 0.80 in October 2012, around a full-sample value of 0.60. The highest values do not come in 2008 but during the euro-area debt crisis of 2010–2012, which tied the two markets even more closely.

            The BET and the DAX show a slower change. Before 2007 the correlation was close to zero, with a minimum of \(-0.08\); after Romania joined the European Union (EU) it rose in steps, up to 0.70 in August 2020. The Romanian market became integrated with European markets, and its value for diversifying a European portfolio fell accordingly. The S&P 500 and Bitcoin went from \(-0.16\) in December 2019 to 0.57 in February 2023: after 2020 Bitcoin behaved more like a risk asset correlated with equities.

            The conclusion is that correlations wander far from any constant value, over years through integration and within months through crises. A model must let them move.

            In brief

            • S&P 500 and DAX: between 0.21 (April 2026) and 0.80 (October 2012); full-sample value 0.60
              • the euro-area debt crisis of 2010–2012 tied the two markets even more closely than 2008
            • BET and DAX: close to zero before 2007 (minimum \(-\)0.08), up to 0.70 in August 2020: the Romanian market became integrated with Europe after EU accession
            • S&P 500 and Bitcoin: from \(-\)0.16 (December 2019) to 0.57 (February 2023); Bitcoin became a risk asset correlated with equities after 2020
            • Correlations wander far from any constant value: a model must let them move

            Asynchronous trading and the bias of daily correlations

            Slide 11 in the PDF

            Daily returns of different markets do not cover the same hours. Bucharest closes at 16:45 Central European Time (CET), 17:45 local time, Frankfurt at 17:30 CET and New York at 22:00 CET, while Bitcoin trades all day, every day. News that arrives after the European close enters European prices only on the next trading day. This is asynchronous trading (non-synchronous trading), and it splits the reaction to one event across two daily returns.

            The tariff announcement of 9 April 2025 shows the effect. The pause of the United States (US) tariffs was announced in the afternoon in New York, after the European close: on 9 April the S&P 500 rose by 9.1% and the DAX fell by 3.1%; on 10 April the S&P 500 fell by 3.5% and the DAX rose by 4.4%. The same news produced returns of opposite sign on the same calendar day, and this pulls the daily correlation towards zero.

            Weekly returns are much less affected, because a one-day delay matters little over five days. Since 2015 the correlation of daily against weekly returns is 0.32 against 0.46 for the S&P 500 and the BET, 0.56 against 0.68 for the S&P 500 and the DAX, and 0.42 against 0.47 for the BET and the DAX. The gap is small for the last pair, whose markets close almost together. The remedies are weekly returns or returns measured between synchronous times; the daily models of this chapter keep the bias in mind when reading transatlantic correlations.

            In brief

            • Markets close at different hours: Bucharest at 16:45 CET (17:45 local time), Frankfurt at 17:30 CET, New York at 22:00 CET; Bitcoin trades all day, every day
              • news that arrives after the European close shows up in the European prices only the next day
            • Example: 9 April 2025, the pause of the US tariffs announced in the afternoon in New York
              • 9 April: S&P 500 +9.1%, DAX \(-\)3.1%; 10 April: S&P 500 \(-\)3.5%, DAX +4.4%
              • the same news, on two different days: the daily correlation is biased towards zero
            • Correlations of daily and weekly returns since 2015: S&P 500 and BET 0.32 against 0.46; S&P 500 and DAX 0.56 against 0.68; BET and DAX 0.42 against 0.47
              • the gap is small for BET and DAX, which close almost together; remedies: weekly returns or returns between synchronous times

            Recap: Why multivariate volatility

            Slide 12 in the PDF

            The portfolio variance \(\mathbf{w}^\top\mathbf{H}\mathbf{w}\), the hedge ratio and the question of contagion all depend on covariances. Rolling correlations of real markets move a great deal, through market integration, through crises and through the arrival of new asset classes such as Bitcoin. Before computing any daily correlation, the trading hours must be checked, because asynchronous closes bias daily correlations towards zero.

            In brief

            • Portfolio variance \(\mathbf{w}^\top\mathbf{H}\mathbf{w}\), hedge ratios and contagion all depend on covariances
            • Rolling correlations of real markets move a lot: integration, crises, new asset classes
            • Check the trading hours: asynchronous closes bias daily correlations towards zero

            Section summary

            The section has motivated a model in which correlations change over time; rolling windows show the change but do not explain it and depend on an arbitrary window length. The next section sets up the general framework, a conditional covariance matrix that is updated every day from past shocks, and the conditions it must satisfy.

            The multivariate GARCH framework

            The section defines the multivariate GARCH model in general form: the vector of returns, the conditional covariance matrix \(\mathbf{H}_t\), the conditions of symmetry and positive definiteness, and estimation by maximum likelihood.

            The model for \(N\) return series

            Slide 13 in the PDF

            The model for \(N\) returns has the same structure as the univariate GARCH model, written with vectors and matrices:

            \[\begin{gathered} \mathbf{r}_t = \boldsymbol{\mu} + \boldsymbol{\varepsilon}_t, \qquad \boldsymbol{\varepsilon}_t = \mathbf{H}_t^{1/2}\boldsymbol{\eta}_t, \\ \mathbf{H}_t = \Var(\mathbf{r}_t \mid \mathcal{F}_{t-1}). \end{gathered}\]
            (14.2)

            Here \(\mathbf{r}_t\) is the vector of the \(N\) returns on day \(t\), \(\boldsymbol{\mu}\) the vector of means (or the forecast of a VAR model, as in Chapter 6), and \(\boldsymbol{\eta}_t\) a vector of independent and identically distributed (i.i.d.) innovations with mean zero and covariance matrix \(\mathbf{I}_N\). The matrix \(\mathbf{H}_t^{1/2}\) is a square root of \(\mathbf{H}_t\), for instance the Cholesky factor \(\mathbf{L}_t\) with \(\mathbf{L}_t\mathbf{L}_t^\top = \mathbf{H}_t\); any such root gives the shocks \(\boldsymbol{\varepsilon}_t\) the conditional covariance matrix \(\mathbf{H}_t\).

            The conditional covariance matrix (conditional covariance matrix) \(\mathbf{H}_t\) is computed from the information \(\mathcal{F}_{t-1}\) available up to day \(t-1\). Its diagonal holds the conditional variances \(h_{ii,t} = \sigma_{i,t}^2\), exactly the quantities of Chapter 5. Off the diagonal are the conditional covariances \(h_{ij,t}\), and the conditional correlation follows from them:

            \[\rho_{ij,t} = \frac{h_{ij,t}}{\sqrt{h_{ii,t}\,h_{jj,t}}}.\]
            (14.3)

            A multivariate GARCH (MGARCH) model is a rule that computes \(\mathbf{H}_t\) from the past shocks \(\boldsymbol{\varepsilon}_{t-1}\) and from \(\mathbf{H}_{t-1}\). The models of the following sections differ only in this rule.

            In brief

            • \(\mathbf{r}_t = \boldsymbol{\mu} + \boldsymbol{\varepsilon}_t\),   \(\boldsymbol{\varepsilon}_t = \mathbf{H}_t^{1/2}\boldsymbol{\eta}_t\),   \(\boldsymbol{\eta}_t\) i.i.d. with mean \(\mathbf{0}\) and covariance \(\mathbf{I}_N\)
              • \(\mathbf{r}_t\): the vector of \(N\) returns on day \(t\); \(\boldsymbol{\mu}\): the means (or a VAR model for the mean, Chapter 6)
              • \(\mathbf{H}_t^{1/2}\): a square root of \(\mathbf{H}_t\), e.g. the Cholesky factor \(\mathbf{L}_t\) with \(\mathbf{L}_t\mathbf{L}_t^\top = \mathbf{H}_t\)
            • \(\mathbf{H}_t = \Var(\mathbf{r}_t \mid \mathcal{F}_{t-1})\): the conditional covariance matrix, given the information \(\mathcal{F}_{t-1}\) up to day \(t-1\)
              • diagonal: the conditional variances \(h_{ii,t} = \sigma_{i,t}^2\) (Chapter 5)
              • off-diagonal: the conditional covariances \(h_{ij,t}\); the conditional correlation is \(\rho_{ij,t} = h_{ij,t}/\sqrt{h_{ii,t}h_{jj,t}}\)
            • A multivariate GARCH (MGARCH) model is a rule that computes \(\mathbf{H}_t\) from past shocks \(\boldsymbol{\varepsilon}_{t-1}\) and past \(\mathbf{H}_{t-1}\)

            Two conditions on \(\mathbf{H}_t\)

            Slide 14 in the PDF

            The first condition is symmetry: the covariance of asset \(i\) with asset \(j\) is the covariance of \(j\) with \(i\), so \(h_{ij,t} = h_{ji,t}\) and only the \(N(N+1)/2\) elements on and below the diagonal are distinct. The operator \(\mathrm{vech}\) stacks these elements column by column in a vector; for two assets, \(\mathrm{vech}(\mathbf{H}) = (h_{11}, h_{12}, h_{22})^\top\). The number of distinct elements grows quickly: 3 for \(N = 2\), 55 for \(N = 10\) and 1 275 for \(N = 50\).

            The second condition is positive definiteness (positive definiteness): \(\mathbf{w}^\top\mathbf{H}_t\mathbf{w} > 0\) for every non-zero vector \(\mathbf{w}\). By (14.1) the left side is the variance of the portfolio with weights \(\mathbf{w}\), so the condition says that every portfolio has a positive variance. For two assets it reduces to \(h_{11,t} > 0\) and \(h_{11,t}h_{22,t} - h_{12,t}^2 > 0\), which is the same as \(|\rho_{12,t}| < 1\).

            The example on the slide shows what goes wrong without this condition. With \(h_{11} = 4\), \(h_{22} = 1\) and \(h_{12} = 2.5\) the determinant is \(4 - 6.25 < 0\), and (14.3) gives a "correlation" of 1.25, which no pair of random variables can have. A model that computes each element of \(\mathbf{H}_t\) separately can produce such matrices. A good MGARCH model guarantees both conditions for every \(t\), whatever the shocks, and does so with few parameters.

            In brief

            • Symmetry
              • \(h_{ij,t} = h_{ji,t}\), so only \(N(N+1)/2\) elements are distinct
              • the operator \(\mathrm{vech}\) stacks the lower triangle: \(\mathrm{vech}\begin{pmatrix} h_{11} & h_{12}\\ h_{12} & h_{22}\end{pmatrix} = (h_{11}, h_{12}, h_{22})^\top\)
              • \(N = 2\): 3 elements; \(N = 10\): 55; \(N = 50\): 1 275
            • Positive definiteness
              • \(\mathbf{w}^\top\mathbf{H}_t\mathbf{w} > 0\) for every \(\mathbf{w} \neq \mathbf{0}\), i.e. every portfolio has a positive variance
              • \(N = 2\): \(h_{11,t} > 0\) and \(h_{11,t}h_{22,t} - h_{12,t}^2 > 0\), i.e. \(|\rho_{12,t}| < 1\)
              • example: \(h_{11} = 4\), \(h_{22} = 1\), \(h_{12} = 2.5\) gives \(4 - 6.25 < 0\): an impossible “correlation” of 1.25
            • A good MGARCH model guarantees both conditions for every \(t\), with few parameters

            Estimation by maximum likelihood

            Slide 15 in the PDF

            If the innovations \(\boldsymbol{\eta}_t\) follow the multivariate Normal distribution, the shocks \(\boldsymbol{\varepsilon}_t\) are conditionally Normal with covariance matrix \(\mathbf{H}_t\), and the log-likelihood of the sample is

            \[\ell(\theta) = -\frac12\sum_{t=1}^{T}\left(N\log 2\pi + \log|\mathbf{H}_t| + \boldsymbol{\varepsilon}_t^\top\mathbf{H}_t^{-1}\boldsymbol{\varepsilon}_t\right),\]
            (14.4)

            where \(|\mathbf{H}_t|\) is the determinant and \(\theta\) collects all parameters. For \(N = 1\) the determinant is \(\sigma_t^2\), the quadratic form is \(\varepsilon_t^2/\sigma_t^2\), and (14.4) becomes the GARCH likelihood of Chapter 5. The two terms play the same roles as there: the determinant penalises large variances, the quadratic form penalises shocks that are large relative to the predicted covariance.

            Daily returns have fat tails, so the Normal assumption is not literally true. The Normal likelihood is then used as a quasi-maximum likelihood (QML): the estimates of the parameters of \(\mathbf{H}_t\) remain consistent as long as \(\mathbf{H}_t\) is correctly specified, but the usual standard errors are wrong and need a robust (sandwich) correction.

            The practical difficulty is computational. Each evaluation of \(\ell(\theta)\) requires the inverse and the determinant of \(\mathbf{H}_t\) for every day of the sample, and a numerical optimiser evaluates \(\ell(\theta)\) many times. The cost grows quickly with \(N\) and, above all, with the number of parameters, which is the subject of the next section.

            In brief

            • If \(\boldsymbol{\eta}_t\) follows the multivariate Normal distribution, the log-likelihood is
              • \(\ell(\theta) = -\frac12\sum_{t=1}^{T}\left(N\log 2\pi + \log|\mathbf{H}_t| + \boldsymbol{\varepsilon}_t^\top\mathbf{H}_t^{-1}\boldsymbol{\varepsilon}_t\right)\)
              • \(|\mathbf{H}_t|\) is the determinant; for \(N = 1\) this is the GARCH likelihood of Chapter 5
            • Returns have fat tails, so the Normal likelihood is used as quasi-maximum likelihood (QML)
              • the estimates stay consistent; the standard errors need a robust (sandwich) correction
            • Every evaluation of \(\ell(\theta)\) needs \(\mathbf{H}_t^{-1}\) and \(|\mathbf{H}_t|\) for all \(t\): the cost grows fast with \(N\) and with the number of parameters

            Recap: The multivariate GARCH framework

            Slide 16 in the PDF

            The conditional covariance matrix \(\mathbf{H}_t\) has the conditional variances on its diagonal and the covariances off it, and the conditional correlation follows from (14.3). A valid model must keep \(\mathbf{H}_t\) symmetric, with \(N(N+1)/2\) distinct elements, and positive definite on every day. Estimation uses the multivariate Normal likelihood (14.4), as a quasi-likelihood when returns have fat tails.

            In brief

            • \(\mathbf{H}_t\): conditional variances on the diagonal, covariances off the diagonal; \(\rho_{ij,t} = h_{ij,t}/\sqrt{h_{ii,t}h_{jj,t}}\)
            • Conditions: symmetry (\(N(N+1)/2\) distinct elements) and positive definiteness
            • Estimation by (quasi-)maximum likelihood with the multivariate Normal density

            Section summary

            The framework leaves open the rule that updates \(\mathbf{H}_t\). The next section presents the two classic answers, VEC and BEKK, which model all elements of the matrix at once, and shows the price they pay in parameters.

            VEC and BEKK models

            The section presents the most general linear MGARCH model, VEC, and the BEKK model, which guarantees positive definiteness by construction. Counting their parameters shows why neither can be used for more than a handful of assets.

            The VEC model

            Slide 17 in the PDF

            Bollerslev, Engle and Wooldridge (1988) wrote the most general linear extension of GARCH(1,1): every distinct element of \(\mathbf{H}_t\) depends on every past squared shock and cross-product and on every element of \(\mathbf{H}_{t-1}\). With the \(\mathrm{vech}\) operator, the VEC(1,1) model is

            \[\begin{aligned} \mathrm{vech}(\mathbf{H}_t) = \mathbf{c} &+ \mathbf{A}\,\mathrm{vech}(\boldsymbol{\varepsilon}_{t-1}\boldsymbol{\varepsilon}_{t-1}^\top) \\ &+ \mathbf{B}\,\mathrm{vech}(\mathbf{H}_{t-1}). \end{aligned}\]
            (14.5)

            The matrix \(\boldsymbol{\varepsilon}_{t-1}\boldsymbol{\varepsilon}_{t-1}^\top\) holds yesterday's squared shocks on its diagonal and their cross-products off it. The vector \(\mathbf{c}\) holds constants, \(\mathbf{A}\) measures the reaction to shocks and \(\mathbf{B}\) the persistence, as \(\omega\), \(\alpha\) and \(\beta\) do in a univariate GARCH.

            The generality is expensive. With \(k = N(N+1)/2\) distinct elements, \(\mathbf{c}\) has \(k\) entries and \(\mathbf{A}\) and \(\mathbf{B}\) are \(k \times k\), so the model has \(k + 2k^2\) parameters: 21 for two assets and 6 105 for ten. Worse, nothing in (14.5) keeps \(\mathbf{H}_t\) positive definite; the restrictions that would do so are complicated and hard to impose during estimation.

            The diagonal VEC keeps only the diagonal of \(\mathbf{A}\) and \(\mathbf{B}\), so that each element follows its own GARCH(1,1):

            \[h_{ij,t} = c_{ij} + a_{ij}\varepsilon_{i,t-1}\varepsilon_{j,t-1} + b_{ij}h_{ij,t-1}.\]

            This leaves \(3k\) parameters, 9 for \(N = 2\), but positive definiteness is still not guaranteed, because each element is updated without regard to the others.

            In brief

            • Bollerslev, Engle and Wooldridge (1988): every element of \(\mathbf{H}_t\) depends on every past squared shock, cross-product and element of \(\mathbf{H}_{t-1}\)
              • VEC(1,1): \(\mathrm{vech}(\mathbf{H}_t) = \mathbf{c} + \mathbf{A}\,\mathrm{vech}(\boldsymbol{\varepsilon}_{t-1}\boldsymbol{\varepsilon}_{t-1}^\top) + \mathbf{B}\,\mathrm{vech}(\mathbf{H}_{t-1})\)
              • \(\boldsymbol{\varepsilon}_{t-1}\boldsymbol{\varepsilon}_{t-1}^\top\): the matrix of yesterday's squared shocks and cross-products; \(\mathbf{c}\): a vector of constants; \(\mathbf{A}\): the reaction to shocks; \(\mathbf{B}\): the persistence
              • with \(k = N(N+1)/2\): \(\mathbf{c}\) has \(k\) elements, \(\mathbf{A}\) and \(\mathbf{B}\) are \(k \times k\): in total \(k + 2k^2\) parameters
            • \(N = 2\): \(k = 3\) and 21 parameters; \(N = 10\): 6 105
              • and positive definiteness of \(\mathbf{H}_t\) is not guaranteed: it needs complicated restrictions
            • Diagonal VEC
              • \(h_{ij,t} = c_{ij} + a_{ij}\varepsilon_{i,t-1}\varepsilon_{j,t-1} + b_{ij}h_{ij,t-1}\)
              • each element is a GARCH(1,1) of its own; \(3k\) parameters (9 for \(N = 2\)); still no guarantee of positive definiteness

            The BEKK model

            Slide 18 in the PDF

            Engle and Kroner (1995) proposed a form named after Baba, Engle, Kraft and Kroner (BEKK) that is positive definite by construction:

            \[\begin{aligned} \mathbf{H}_t = \mathbf{C}\mathbf{C}^\top &+ \mathbf{A}^\top\boldsymbol{\varepsilon}_{t-1}\boldsymbol{\varepsilon}_{t-1}^\top\mathbf{A} \\ &+ \mathbf{B}^\top\mathbf{H}_{t-1}\mathbf{B}, \end{aligned}\]
            (14.6)

            with \(\mathbf{C}\) lower triangular and \(\mathbf{A}\), \(\mathbf{B}\) full \(N \times N\) matrices. \(\mathbf{A}\) describes the reaction to yesterday's shocks and \(\mathbf{B}\) the persistence of \(\mathbf{H}_{t-1}\); the elements of the constant matrix \(\mathbf{C}\mathbf{C}^\top\) are written \(c_{ij}^*\).

            The guarantee comes from the form of each term, a "square" \(\mathbf{X}^\top\mathbf{M}\mathbf{X}\) of a positive semidefinite matrix \(\mathbf{M}\): the identity in the first term (with \(\mathbf{X} = \mathbf{C}^\top\)), the outer product \(\boldsymbol{\varepsilon}_{t-1}\boldsymbol{\varepsilon}_{t-1}^\top\) in the second and \(\mathbf{H}_{t-1}\) in the third. For any vector \(\mathbf{w}\),

            \[\mathbf{w}^\top\mathbf{X}^\top\mathbf{M}\mathbf{X}\mathbf{w} = (\mathbf{X}\mathbf{w})^\top\mathbf{M}(\mathbf{X}\mathbf{w}) \ge 0,\]

            and a constant term \(\mathbf{C}\mathbf{C}^\top\) of full rank makes the sum strictly positive. The number of parameters is \(N(N+1)/2 + 2N^2\): 11 for two assets, 255 for ten.

            The off-diagonal elements of a full \(\mathbf{A}\) have an economic meaning: they measure volatility spillovers (volatility spillovers), the effect of a shock to asset 1 on the variance of asset 2. Restricting \(\mathbf{A}\) and \(\mathbf{B}\) to be diagonal gives the diagonal BEKK, in which the covariance follows

            \[\begin{aligned} h_{12,t} = c_{12}^* &+ a_{11}a_{22}\,\varepsilon_{1,t-1}\varepsilon_{2,t-1} \\ &+ b_{11}b_{22}\,h_{12,t-1}. \end{aligned}\]

            The scalar BEKK goes further, with \(\mathbf{A} = a\mathbf{I}\) and \(\mathbf{B} = b\mathbf{I}\), so that all elements share the same dynamics. Both restrictions save parameters and remove the spillovers.

            In brief

            • Engle and Kroner (1995): \(\mathbf{H}_t = \mathbf{C}\mathbf{C}^\top + \mathbf{A}^\top\boldsymbol{\varepsilon}_{t-1}\boldsymbol{\varepsilon}_{t-1}^\top\mathbf{A} + \mathbf{B}^\top\mathbf{H}_{t-1}\mathbf{B}\)
              • BEKK: Baba, Engle, Kraft and Kroner; \(\mathbf{C}\) lower triangular, \(\mathbf{A}\) and \(\mathbf{B}\) are \(N \times N\)
              • \(\mathbf{A}\): the reaction to yesterday's shocks; \(\mathbf{B}\): the persistence of \(\mathbf{H}_{t-1}\); \(c_{ij}^*\): the elements of \(\mathbf{C}\mathbf{C}^\top\)
            • Positive definite by construction
              • each term is a “square” \(\mathbf{X}^\top\mathbf{M}\mathbf{X}\) of a positive (semi)definite matrix \(\mathbf{M}\)
              • \(\mathbf{M}\): \(\mathbf{I}\), \(\boldsymbol{\varepsilon}_{t-1}\boldsymbol{\varepsilon}_{t-1}^\top\) or \(\mathbf{H}_{t-1}\); \(\mathbf{X}\): \(\mathbf{C}^\top\), \(\mathbf{A}\) or \(\mathbf{B}\)
              • parameters: \(N(N+1)/2 + 2N^2\), i.e. 11 for \(N = 2\) and 255 for \(N = 10\)
            • Diagonal BEKK (\(\mathbf{A}\), \(\mathbf{B}\) diagonal)
              • \(h_{12,t} = c_{12}^* + a_{11}a_{22}\,\varepsilon_{1,t-1}\varepsilon_{2,t-1} + b_{11}b_{22}\,h_{12,t-1}\)
              • scalar BEKK: \(\mathbf{A} = a\mathbf{I}\), \(\mathbf{B} = b\mathbf{I}\), the same dynamics for all elements
            • The off-diagonal elements of a full \(\mathbf{A}\) measure volatility spillovers: a shock to asset 1 raises the variance of asset 2

            Worked example: one step of a diagonal BEKK

            Slide 19 in the PDF

            The covariance equation of a diagonal BEKK can be computed by hand. The given values are \(c_{12}^* = 0.02\), \(a_{11} = 0.30\), \(a_{22} = 0.25\), \(b_{11} = 0.94\), \(b_{22} = 0.95\), yesterday's covariance \(h_{12,t-1} = 0.5\) and yesterday's shocks \(\varepsilon_{1,t-1} = -2\) and \(\varepsilon_{2,t-1} = -1.5\). The coefficients of the covariance equation are products: \(a_{11}a_{22} = 0.075\) for the reaction to the cross-product of shocks and \(b_{11}b_{22} = 0.893\) for persistence.

            The cross-product of the shocks is \((-2)(-1.5) = 3.000\), positive because both assets fell. Today's covariance is

            \[\begin{aligned} h_{12,t} &= 0.020 + 0.075 \cdot 3.000 + 0.893 \cdot 0.5 \\ &= 0.020 + 0.225 + 0.4465 = 0.6915. \end{aligned}\]

            A joint fall raises the covariance from 0.5 to 0.6915; a shock with opposite signs would have given a negative cross-product and lowered it.

            The example also shows what the diagonal model cannot do. Because \(a_{12} = a_{21} = 0\), a shock to asset 1 affects the variance of asset 2 only through the covariance, never directly: the diagonal BEKK has no volatility spillovers.

            In brief

            • Given: \(c_{12}^* = 0.02\), \(a_{11} = 0.30\), \(a_{22} = 0.25\), \(b_{11} = 0.94\), \(b_{22} = 0.95\), \(h_{12,t-1} = 0.5\); yesterday's shocks \(\varepsilon_{1,t-1} = -2\), \(\varepsilon_{2,t-1} = -1.5\)
            • Step 1: the coefficients \(a_{11}a_{22} = 0.075\) and \(b_{11}b_{22} = 0.893\)
            • Step 2: the cross-product of shocks \(\varepsilon_{1,t-1}\varepsilon_{2,t-1} = 3.000\) (both negative: a joint fall)
            • Step 3: \(h_{12,t} = 0.020 + 0.075 \cdot 3.000 + 0.893 \cdot 0.5 = 0.020 + 0.225 + 0.4465 = 0.6915\)
            • A joint fall raises the covariance from 0.5 to 0.6915; a shock of opposite signs would have lowered it
              • note: a diagonal BEKK has no spillovers, because \(a_{12} = a_{21} = 0\)

            The curse of dimensionality

            Slide 20 in the PDF

            The table counts the parameters of the variance equation of each model for 2, 5, 10 and 50 assets; the means \(\boldsymbol{\mu}\) are not counted. For CCC and DCC, introduced in the next section, the count is 3 GARCH parameters per asset plus the \(N(N-1)/2\) correlations, and DCC adds 2 more.

            The rows differ by orders of magnitude. The full VEC needs 21 parameters for two assets, 465 for five, 6 105 for ten and 3 252 525 for fifty: for fifty assets it has more parameters than any sample of daily data has observations. The full BEKK is much more economical but still reaches 6 275 parameters for fifty assets. The diagonal BEKK, CCC and DCC stay close to each other, with 1 375 to 1 377 parameters for fifty assets, and the scalar BEKK is the cheapest, with 1 277. This growth of the number of parameters with the dimension is called the curse of dimensionality (curse of dimensionality).

            In brief

            Model\(N = 2\)\(N = 5\)\(N = 10\)\(N = 50\)
            VEC(1,1)214656 1053 252 525
            Diagonal VEC9451653 825
            BEKK(1,1)11652556 275
            Diagonal BEKK725751 375
            Scalar BEKK517571 277
            CCC725751 375
            DCC927771 377
            • Number of parameters of the variance equation (the means \(\boldsymbol{\mu}\) are not counted); CCC and DCC: 3 GARCH parameters per asset plus the \(N(N-1)/2\) correlations, and 2 more for DCC
            • A full VEC for 50 assets has more parameters than any sample has observations

            Parameters against the number of assets

            Slide 21 in the PDF

            Worked example

            \[\mathbf{w}^\top\mathbf{H}\mathbf{w}>0\ \ \forall\,\mathbf{w}\neq\mathbf{0}\iff h_{11}>0,\ \ h_{11}h_{22}-h_{12}^2>0,\qquad \rho_{12}=\frac{h_{12}}{\sqrt{h_{11}h_{22}}}\]
                Question for the room ▶ lowers h12 from 2.75 to −3 (h11 = 4, h22 = 1). For which values is H positive definite?
                Answer

                Only for |h12| < 2, i.e. |ρ| < 1. At h12 = ±2 the determinant and one eigenvalue are 0 (a riskless combination); for |h12| > 2 some portfolios have a negative variance, for example −0.416 for h12 = ±2.5. The sign of the covariance does not matter, only its size against √(h11h22).

                Worked example

                \[\begin{gathered}h_{12,t}=c_{12}^*+a_{11}a_{22}\,\varepsilon_{1,t-1}\varepsilon_{2,t-1}+b_{11}b_{22}\,h_{12,t-1}\\ c_{12}^*=0.02,\ a_{11}=0.30,\ a_{22}=0.25,\ b_{11}=0.94,\ b_{22}=0.95,\ h_{12,t-1}=0.5\end{gathered}\]
                    Question for the room Set ε2,t−1 = +1.5 (the opposite sign of ε1,t−1 = −2). What is h12,t?
                    Answer

                    0.2415 = 0.020 + 0.075 · (−3) + 0.4465: shocks of opposite signs lower the covariance. With ε2,t−1 = 0, h12,t = 0.4665, still below 0.5, although the long-run level is 0.625: that level accounts for the cross-product of the shocks being equal to h12 on average, and a day without a joint shock brings only the constant and the persistence.

                      Question for the room For N = 30, BEKK(1,1) has 2,265 parameters. How many does DCC have?
                      Answer

                      527: 90 GARCH parameters estimated series by series, 435 sample correlations and only 2 parameters, a and b, found by numerical optimisation. VEC(1,1) has 432,915.

                        Question for the room ▶ raises the largest N up to 100. By what factor does the number of VEC(1,1) parameters grow from N = 50 to N = 100?
                        Answer

                        15.7 times (from 3,252,525 to 51,010,050), like N4: 24 = 16. DCC grows 3.8 times (from 1,377 to 5,252), BEKK(1,1) 4 times (from 6,275 to 25,050).

                        The chart plots the number of parameters of the variance equation, on a logarithmic vertical axis, against the number of assets \(N\), from 2 to 50, for six models: VEC(1,1), BEKK(1,1), diagonal VEC, diagonal BEKK, DCC and CCC. The scalar BEKK, from the table, is not drawn.

                        On a logarithmic scale a constant vertical distance means a constant ratio. The VEC curve lies far above all others and keeps moving away, passing \(10^6\) at \(N = 50\). The full BEKK and the diagonal VEC lie in between, and the last three curves (diagonal BEKK, DCC and CCC) almost coincide, because their counts differ by at most two parameters.

                        In brief

                        • Number of parameters of the variance equation for \(N = 2, \dots, 50\) assets, log scale
                        TSA_ch14_parameters

                        Interpreting the parameter counts

                        Slide 22 in the PDF

                        The leading terms of the formulas give the growth rates. For VEC, \(2k^2\) with \(k \approx N^2/2\) grows like \(N^4/2\); for the full BEKK the count grows like \(2N^2\). Both become impossible to estimate beyond a handful of assets, because a numerical optimiser cannot search reliably over hundreds or thousands of parameters with a likelihood like (14.4).

                        CCC and DCC grow like \(N^2/2\), but the count hides an important difference: the correlations are not found by numerical optimisation. As the next section shows, they are sample correlations of standardised residuals, computed directly. The GARCH parameters are estimated one series at a time, with three parameters each. The only joint optimisation in DCC concerns two parameters, \(a\) and \(b\), whatever the number of assets. This is why DCC can be used where VEC and BEKK cannot.

                        In brief

                        • VEC grows like \(N^4/2\), full BEKK like \(2N^2\): both become impossible to estimate beyond a handful of assets
                        • CCC and DCC grow like \(N^2/2\), but the correlations are not found by numerical optimisation: they are sample correlations (Section 4)
                          • the DCC optimiser works on only 2 parameters, \(a\) and \(b\), whatever \(N\)

                        Recap: VEC and BEKK

                        Slide 23 in the PDF

                        VEC is the most general linear model, with \(k + 2k^2\) parameters and no guarantee of positive definiteness. BEKK is positive definite by construction, and the off-diagonal elements \(a_{ij}\) of its matrix \(\mathbf{A}\) measure volatility spillovers. The diagonal and scalar versions cut the number of parameters but remove the spillovers.

                        In brief

                        • VEC: the most general linear model; \(k + 2k^2\) parameters, no positive-definiteness guarantee
                        • BEKK: positive definite by construction; off-diagonal \(a_{ij}\) measure volatility spillovers
                        • Diagonal and scalar versions cut the parameters but remove the spillovers

                        Section summary

                        VEC and BEKK model the whole matrix \(\mathbf{H}_t\) in one equation, and this is the source of their cost. The next section takes a different route: it splits \(\mathbf{H}_t\) into volatilities, which univariate GARCH models already handle, and a correlation matrix with very few dynamic parameters.

                        Constant and dynamic conditional correlations

                        The section decomposes the covariance matrix into volatilities and correlations, presents the constant conditional correlation model and Engle's dynamic conditional correlation model, explains their estimation in two steps and the test of DCC against CCC.

                        Splitting volatilities and correlations

                        Slide 24 in the PDF

                        Any covariance matrix can be written as volatilities times correlations:

                        \[\begin{gathered} \mathbf{H}_t = \mathbf{D}_t\mathbf{R}_t\mathbf{D}_t, \\ \mathbf{D}_t = \mathrm{diag}(\sigma_{1,t}, \dots, \sigma_{N,t}), \end{gathered}\]
                        (14.7)

                        where \(\mathbf{R}_t\) is the conditional correlation matrix, with ones on the diagonal. Element by element, \(h_{ij,t} = \rho_{ij,t}\,\sigma_{i,t}\sigma_{j,t}\), which is (14.3) read the other way. Each volatility follows a univariate GARCH(1,1) model (Bollerslev, 1986):

                        \[\sigma_{i,t}^2 = \omega_i + \alpha_i\varepsilon_{i,t-1}^2 + \beta_i\sigma_{i,t-1}^2.\]

                        Dividing each shock by its own volatility gives the standardised residuals (standardised residuals) \(z_{i,t} = \varepsilon_{i,t}/\sigma_{i,t}\). Each has conditional variance one, and the conditional covariance of \(z_{i,t}\) and \(z_{j,t}\) is \(h_{ij,t}/(\sigma_{i,t}\sigma_{j,t}) = \rho_{ij,t}\). The conditional covariance matrix of the vector \(\mathbf{z}_t\) is therefore exactly \(\mathbf{R}_t\): once the volatilities are removed, only the correlations are left to model.

                        The simplest choice is the constant conditional correlation (CCC) model of Bollerslev (1990), with \(\mathbf{R}_t = \mathbf{R}\) for all \(t\). \(\mathbf{H}_t\) is then positive definite whenever \(\mathbf{R}\) is and the GARCH variances are positive, and \(\mathbf{R}\) is estimated by the sample correlation matrix of the \(z_{i,t}\). In CCC the covariances still move, but only through the volatilities. The rolling correlations of the first section contradict this assumption.

                        In brief

                        • Decomposition: \(\mathbf{H}_t = \mathbf{D}_t\mathbf{R}_t\mathbf{D}_t\),   \(\mathbf{D}_t = \mathrm{diag}(\sigma_{1,t}, \dots, \sigma_{N,t})\)
                          • \(\mathbf{R}_t\): the conditional correlation matrix (ones on the diagonal); element by element: \(h_{ij,t} = \rho_{ij,t}\,\sigma_{i,t}\sigma_{j,t}\)
                          • each \(\sigma_{i,t}\) is a univariate GARCH(1,1) Bollerslev (1986): \(\sigma_{i,t}^2 = \omega_i + \alpha_i\varepsilon_{i,t-1}^2 + \beta_i\sigma_{i,t-1}^2\)
                        • Standardised residuals
                          • \(z_{i,t} = \varepsilon_{i,t}/\sigma_{i,t}\)
                          • their conditional covariance matrix is exactly \(\mathbf{R}_t\)
                        • CCC (constant conditional correlation, Bollerslev (1990))
                          • \(\mathbf{R}_t = \mathbf{R}\) for all \(t\)
                          • \(\mathbf{H}_t\) is positive definite whenever \(\mathbf{R}\) is; \(\mathbf{R}\) is estimated by the sample correlation of the \(z_{i,t}\)
                          • covariances still move, but only through the volatilities; the rolling correlations of Section 1 reject this

                        Robert Engle and the DCC model

                        Slide 25 in the PDF

                        Robert F. Engle, born in 1942, received the Nobel Prize in economics in 2003 for methods of analysing economic time series with time-varying volatility. Engle (2002) proposed letting the correlations follow a recursion of the same form as GARCH. An auxiliary matrix \(\mathbf{Q}_t\), with elements \(q_{ij,t}\), is updated each day from yesterday's vector of standardised residuals:

                        \[\begin{aligned} \mathbf{Q}_t &= (1 - a - b)\,\bar{\mathbf{Q}} + a\,\mathbf{z}_{t-1}\mathbf{z}_{t-1}^\top \\ &\quad + b\,\mathbf{Q}_{t-1}, \\ \rho_{ij,t} &= \frac{q_{ij,t}}{\sqrt{q_{ii,t}\,q_{jj,t}}}. \end{aligned}\]
                        (14.8)

                        The rescaling in the second part is needed because \(\mathbf{Q}_t\) is a weighted average of matrices whose diagonals are not exactly one; dividing by the square roots of the diagonal turns it into a proper correlation matrix with ones on the diagonal.

                        \(\bar{\mathbf{Q}}\) is the unconditional correlation matrix of \(\mathbf{z}_t\), and it is set to its sample value before the optimisation. This is correlation targeting (correlation targeting): the recursion is anchored at the observed long-run correlation, and only \(a\) and \(b\) remain to be estimated. The parameter \(a \ge 0\) is the reaction to yesterday's joint shock, \(b \ge 0\) the persistence, and \(a + b < 1\) ensures that \(\mathbf{Q}_t\) reverts to \(\bar{\mathbf{Q}}\).

                        With \(a = b = 0\) the recursion gives \(\mathbf{Q}_t = \bar{\mathbf{Q}}\) on every day, which is CCC. DCC therefore contains CCC as a special case, and this nesting is what makes a likelihood ratio test between the two possible.

                        In brief

                        Robert Engle and the DCC model Robert F. Engle (b. 1942), Nobel Prize 2003 Photo: Jon Demske (2022); CC BY-SA 4.0; Wikimedia Commons
                        • Engle (2002): let the correlations follow a GARCH-like recursion
                          • \(\mathbf{Q}_t = (1 - a - b)\,\bar{\mathbf{Q}} + a\,\mathbf{z}_{t-1}\mathbf{z}_{t-1}^\top + b\,\mathbf{Q}_{t-1}\)
                          • \(\mathbf{Q}_t\): an auxiliary matrix with elements \(q_{ij,t}\); \(\mathbf{z}_{t-1}\): yesterday's vector of standardised residuals
                          • \(\rho_{ij,t} = q_{ij,t}/\sqrt{q_{ii,t}\,q_{jj,t}}\): the rescaling puts ones on the diagonal
                        • \(\bar{\mathbf{Q}}\): the unconditional correlation of \(\mathbf{z}_t\), set to its sample value (correlation targeting)
                          • \(a \ge 0\): reaction to yesterday's joint shock; \(b \ge 0\): persistence; \(a + b < 1\): mean reversion to \(\bar{\mathbf{Q}}\)
                        • \(a = b = 0\) gives back CCC

                        Two-step estimation

                        Slide 26 in the PDF

                        Substituting (14.7) into the log-likelihood (14.4) splits it into two parts, \(\ell = \ell_V(\theta_1) + \ell_C(\theta_1, a, b)\). The volatility part \(\ell_V\) depends only on \(\theta_1\), the GARCH parameters of all series, and is the sum of the univariate GARCH likelihoods. The correlation part \(\ell_C\) depends on \(\theta_1\) through the standardised residuals and on \(a\) and \(b\).

                        The split suggests estimating in two steps. In step 1 a GARCH(1,1) is fitted to each series separately, as in Chapter 5, and the fitted volatilities \(\hat\sigma_{i,t}\) and standardised residuals \(\hat z_{i,t} = \hat\varepsilon_{i,t}/\hat\sigma_{i,t}\) are kept. In step 2, \(\bar{\mathbf{Q}}\) is set to the sample correlation of \(\hat{\mathbf{z}}_t\) and the correlation log-likelihood is maximised over \((a, b)\) only:

                        \[\ell_C(a, b) = -\frac12\sum_{t=1}^{T}\left(\log|\mathbf{R}_t| + \hat{\mathbf{z}}_t^\top\mathbf{R}_t^{-1}\hat{\mathbf{z}}_t - \hat{\mathbf{z}}_t^\top\hat{\mathbf{z}}_t\right).\]
                        (14.9)

                        The estimator is consistent, fast and works for large \(N\). Its weak point is inference: the standard errors of step 2 treat the step-1 estimates as known and ignore their estimation error, so they are too small.

                        The test against CCC compares the maximised correlation log-likelihoods: \(LR = 2(\ell_C^{DCC} - \ell_C^{CCC})\), compared with the \(\chi^2(2)\) distribution, whose 5% critical value is 5.99. A large value rejects constant correlation. Because the null hypothesis \(a = b = 0\) lies on the boundary of the parameter space, the \(\chi^2(2)\) critical value is conservative: a rejection with it is a safe rejection. Tse (2000) proposed another test of constant correlations, based on the Lagrange multiplier principle.

                        In brief

                        • The log-likelihood splits into a volatility part and a correlation part: \(\ell = \ell_V(\theta_1) + \ell_C(\theta_1, a, b)\)
                          • \(\theta_1\): the GARCH parameters of all series; \(\ell_V\) depends only on them, \(\ell_C\) also on \(a\) and \(b\)
                        • Step 1
                          • fit a GARCH(1,1) to each series separately (Chapter 5)
                          • keep \(\hat\sigma_{i,t}\) and \(\hat z_{i,t} = \hat\varepsilon_{i,t}/\hat\sigma_{i,t}\)
                        • Step 2
                          • set \(\bar{\mathbf{Q}}\) to the sample correlation of \(\hat{\mathbf{z}}_t\) and maximise over \((a, b)\) only
                          • \(\ell_C(a, b) = -\frac12\sum_t\left(\log|\mathbf{R}_t| + \hat{\mathbf{z}}_t^\top\mathbf{R}_t^{-1}\hat{\mathbf{z}}_t - \hat{\mathbf{z}}_t^\top\hat{\mathbf{z}}_t\right)\)
                        • Consistent, fast, works for large \(N\); the step-2 standard errors ignore the estimation error of step 1 (they are too small)
                        • Test against CCC
                          • likelihood ratio \(LR = 2(\ell_C^{DCC} - \ell_C^{CCC})\), compared with \(\chi^2(2)\) (5% critical value 5.99)
                          • also Tse (2000)
                          • \(\ell_C^{DCC}\), \(\ell_C^{CCC}\): the maximised correlation log-likelihoods of the two models; a large \(LR\) rejects constant correlation
                          • \(a = b = 0\) lies on the boundary of the parameter space, so the \(\chi^2(2)\) critical value is conservative

                        Worked example: one step of the DCC recursion

                        Slide 27 in the PDF

                        The recursion (14.8) can be followed by hand for two assets. The given values are \(a = 0.05\), \(b = 0.93\), a long-run correlation \(\bar q_{12} = 0.5\) (with \(\bar q_{11} = \bar q_{22} = 1\)), yesterday's matrix \(\mathbf{Q}_{t-1}\) with ones on the diagonal and 0.5 off it, and yesterday's standardised residuals \(z_{1,t-1} = -2\) and \(z_{2,t-1} = -2.5\), a large joint fall. The weight of the long-run matrix is \(1 - a - b = 0.020\).

                        The off-diagonal element and, because the squared residuals enter them, the diagonal elements are

                        \[\begin{aligned} q_{12,t} &= 0.020 \cdot 0.5 + 0.05 \cdot (-2)(-2.5) \\ &\quad + 0.93 \cdot 0.5 \\ &= 0.010 + 0.250 + 0.465 = 0.725, \\ q_{11,t} &= 0.020 + 0.05 \cdot 4 + 0.93 = 1.150, \\ q_{22,t} &= 0.020 + 0.05 \cdot 6.25 + 0.93 = 1.2625. \end{aligned}\]

                        The correlation is \(\rho_{12,t} = 0.725/\sqrt{1.150 \cdot 1.2625} = 0.602\).

                        The large joint fall raises tomorrow's correlation from 0.50 to 0.602. The rescaling matters: without it, the value 0.725 of \(q_{12,t}\) would be read as the correlation and overstate the reaction. Large shocks raise the diagonal of \(\mathbf{Q}_t\) as well, and the division by \(\sqrt{q_{11,t}q_{22,t}}\) takes this into account.

                        In brief

                        • Given: \(a = 0.05\), \(b = 0.93\), \(\bar q_{12} = 0.5\) (and \(\bar q_{11} = \bar q_{22} = 1\)), \(\mathbf{Q}_{t-1} = \begin{pmatrix}1 & 0.5\\ 0.5 & 1\end{pmatrix}\), yesterday \(z_{1,t-1} = -2\), \(z_{2,t-1} = -2.5\)
                        • Step 1: \(1 - a - b = 0.020\)
                        • Step 2: \(q_{12,t} = 0.020 \cdot 0.5 + 0.05 \cdot (-2)(-2.5) + 0.93 \cdot 0.5 = 0.010 + 0.250 + 0.465 = 0.725\)
                        • Step 3: \(q_{11,t} = 0.020 + 0.05 \cdot 4 + 0.93 = 1.150\);   \(q_{22,t} = 0.020 + 0.05 \cdot 6.25 + 0.93 = 1.2625\)
                        • Step 4: \(\rho_{12,t} = 0.725/\sqrt{1.150 \cdot 1.2625} = 0.602\)
                          • a large joint fall raises tomorrow's correlation from 0.50 to 0.602; without step 3 the “correlation” \(q_{12,t}\) would be 0.725

                        Reading \(a\) and \(b\)

                        Slide 28 in the PDF

                        The parameter \(a\) is the news coefficient: it measures how much one day's joint shock moves the correlation. Typical daily estimates lie between 0.01 and 0.05, so a single day moves the correlation only a little, and a large change requires a run of joint shocks.

                        The sum \(a + b\) is the persistence. Taking the long-run value away from both sides of (14.8) and replacing \(\mathbf{z}_{t-1}\mathbf{z}_{t-1}^\top\) by its conditional mean, approximately \(\mathbf{Q}_{t-1}\), shows that a deviation of \(q_{ij,t}\) from \(\bar q_{ij}\) is expected to shrink by the factor \(a + b\) each day. After \(n\) days it is multiplied by \((a+b)^n\), and it is halved when \((a+b)^n = 0.5\), which gives the half-life (half-life) \(\ln 0.5/\ln(a + b)\) days. With \(a + b = 0.98\) the half-life is 34 days; with \(0.99\) it is 69 days. Near one, small changes in \(a + b\) change the half-life a great deal.

                        The reading is the same as for \(\alpha\) and \(\alpha + \beta\) in a GARCH(1,1) model in Chapter 5. Two variants address limits of the basic model: the asymmetric DCC of Cappiello, Engle and Sheppard (2006) lets joint falls move correlations more than joint rises, and the corrected DCC of Aielli (2013) modifies the recursion so that the two-step estimator remains consistent.

                        In brief

                        • \(a\) is the news coefficient
                          • how much one day's joint shock moves the correlation
                          • typical daily values 0.01–0.05
                        • \(a + b\) is the persistence
                          • a deviation of \(q_{ij,t}\) from \(\bar q_{ij}\) shrinks by the factor \(a + b\) each day
                          • half-life: \(\ln 0.5/\ln(a + b)\) days; \(a + b = 0.98\) gives 34 days, \(0.99\) gives 69 days
                        • The same reading as \(\alpha\) and \(\alpha + \beta\) in a GARCH(1,1) (Chapter 5)
                        • Variants: asymmetric DCC (joint falls move correlations more than joint rises) Cappiello, Engle and Sheppard (2006); corrected DCC Aielli (2013)

                        Recap: CCC and DCC

                        Slide 29 in the PDF

                        The decomposition (14.7) writes \(\mathbf{H}_t\) as GARCH volatilities times a correlation matrix. In CCC the correlation matrix is constant; in DCC the auxiliary matrix \(\mathbf{Q}_t\) follows a GARCH-like recursion with two parameters, \(a\) and \(b\). Estimation proceeds in two steps, univariate GARCH models first and \((a, b)\) with correlation targeting second, and a likelihood ratio test compares DCC with CCC.

                        In brief

                        • \(\mathbf{H}_t = \mathbf{D}_t\mathbf{R}_t\mathbf{D}_t\): GARCH volatilities times a correlation matrix
                        • CCC: \(\mathbf{R}\) constant; DCC: \(\mathbf{Q}_t\) follows a GARCH-like recursion with parameters \(a\), \(b\)
                        • Two steps: univariate GARCH, then \((a, b)\) with correlation targeting; LR test against CCC

                        Section summary

                        The section has built the model; the next applies it to the pair of the S&P 500 and the DAX, following both steps of the estimation and asking whether the correlation is constant and how it behaves in the crises of 2008 and 2020.

                        DCC for the S&P 500 and the DAX

                        The section estimates a DCC model for daily returns of the S&P 500 and the DAX from January 2000 to September 2026, step by step, and uses it to study correlations in the crises of 2008 and 2020.

                        Step 1: two univariate GARCH models

                        Slide 30 in the PDF

                        The data are daily log returns in percent on the 6 605 days on which both markets traded, from 4 January 2000 to 18 September 2026, from the data provider EOD Historical Data (EODHD). Each series receives a GARCH(1,1) with a constant mean, estimated by Gaussian QML. The table gives the mean \(\mu\), the three GARCH parameters and the persistence \(\alpha + \beta\), together with the excess kurtosis of the raw returns and of the standardised residuals.

                        The GARCH estimates have the usual shape for equity indices. For the S&P 500, \(\alpha = 0.122\) and \(\beta = 0.857\), with persistence 0.980; for the DAX, \(\alpha = 0.098\) and \(\beta = 0.886\), with persistence 0.984. The reaction to news is moderate and volatility shocks fade slowly.

                        The kurtosis columns show what step 1 achieves. The excess kurtosis of the S&P 500 falls from 10.5 for the returns to 1.8 for the standardised residuals, and that of the DAX from 6.2 to 1.4: most of the fat tails come from changing volatility. What remains is still above zero, so the tails of \(z_t\) are fatter than those of the Normal distribution. The correlation of the returns is 0.599 and that of the residuals is \(\bar q_{12} = 0.575\); the latter is the CCC estimate and the anchor of the DCC recursion.

                        In brief

                        Index\(\mu\)\(\omega\)\(\alpha\)\(\beta\)\(\alpha + \beta\)kurt. \(r_t\)kurt. \(z_t\)
                        S&P 5000.0630.0270.1220.8570.98010.51.8
                        DAX0.0670.0300.0980.8860.9846.21.4
                        • Daily log returns in %, 6 605 common trading days, 4 January 2000–18 September 2026 (EODHD); GARCH(1,1) with a constant mean, Gaussian QML; kurt.: excess kurtosis
                        • Standardising by \(\hat\sigma_{i,t}\) removes most of the excess kurtosis, but not all: the tails of \(z_t\) are still fatter than Normal
                        • Correlation of the returns 0.599; correlation of the residuals \(\bar q_{12} = 0.575\), the CCC estimate

                        Step 1: conditional volatilities

                        Slide 31 in the PDF

                        Worked example

                        \[\begin{gathered}\mathbf{Q}_t=(1-a-b)\,\bar{\mathbf{Q}}+a\,\mathbf{z}_{t-1}\mathbf{z}_{t-1}^\top+b\,\mathbf{Q}_{t-1},\qquad \rho_{12,t}=\frac{q_{12,t}}{\sqrt{q_{11,t}\,q_{22,t}}}\\ \bar q_{12}=0.5,\ \ \mathbf{Q}_{t-1}=\begin{pmatrix}1&0.5\\0.5&1\end{pmatrix},\ \ z_{1,t-1}=-2\end{gathered}\]
                            Question for the room ▶ moves z2,t−1 up to +3, with z1,t−1 = −2. What is ρ12,t for z2,t−1 = +2.5?
                            Answer

                            0.187: q12,t = 0.010 − 0.250 + 0.465 = 0.225, while q11,t = 1.150 and q22,t = 1.2625 stay the same. With z2,t−1 = 0 the correlation is 0.454, still below 0.5: a move of one market alone raises a variance in the denominator without adding anything to the covariance.

                            Worked example

                            \[q_{ij,t+k}-\bar q_{ij}\approx(a+b)^k\,(q_{ij,t}-\bar q_{ij}),\qquad \text{half-life}=\frac{\ln 0.5}{\ln(a+b)}\]
                                Question for the room ▶ raises a + b up to 0.995. How does the half-life change from a + b = 0.99 to 0.995?
                                Answer

                                It doubles, from 69 to 138 days: ln(a + b) ≈ −(1 − a − b), so the half-life is about 0.69/(1 − a − b). With a + b = 0.90 it is 6.6 days, and after 20 days 12% of the deviation is left.

                                  Question for the room Which index loses the larger share of its excess kurtosis after the standardisation?
                                  Answer

                                  The S&P 500: 1 − 1.8/10.5 = 83%, against the DAX: 1 − 1.4/6.2 = 78%. Time-varying volatility explains most of the fat tails, but the residuals zt keep an excess kurtosis of 1.4–1.8; this is why a VaR with the Normal quantile will have too many violations.

                                    Question for the room ▶ raises the multiplier of α up to 2 (same persistence). What happens to the 2020 peak of the S&P 500 volatility?
                                    Answer

                                    It rises to 139% (17 March 2020), from 109%: a larger α reacts more to each day’s shock. With the multiplier 0.25 the peak falls to 64% and moves to 25 March 2020. These are extra runs of the filter, not estimates.

                                    The chart plots the one-step-ahead GARCH(1,1) volatility of each index, annualised as \(\hat\sigma_{i,t}\sqrt{252}\) and expressed in percent, on the vertical axis against time from 2000 to 2026. The two lines are the S&P 500 and the DAX, and the shaded bands mark the crises of 2008 and 2020.

                                    In calm years the two volatilities stay low, with the DAX usually somewhat above the S&P 500. The largest peaks are in the autumn of 2008 and in March 2020, and smaller peaks appear in 2001–2003, 2011, 2022 and April 2025. The two lines rise and fall together.

                                    In brief

                                    • One-step-ahead GARCH(1,1) volatility \(\hat\sigma_{i,t}\sqrt{252}\), annualised, in %; shaded: the crises of 2008 and 2020
                                    TSA_ch14_dcc_estimation

                                    Interpreting the volatilities

                                    Slide 32 in the PDF

                                    Both volatilities peak in the same weeks: the S&P 500 reaches 109% on 17 March 2020 and the DAX 80% on 25 March 2020. Volatility clustering happens jointly. This alone makes the covariance \(h_{12,t} = \rho_{12,t}\sigma_{1,t}\sigma_{2,t}\) rise sharply in a crisis, even if the correlation did not change at all.

                                    The persistence of 0.980 and 0.984 means that volatility shocks fade slowly, as in Chapter 5. Step 2 asks a different question from step 1: once the volatilities have been divided out, is the correlation of what remains constant? This separation is the point of DCC: it distinguishes a rise in covariance caused by higher volatility from a rise caused by a closer link.

                                    In brief

                                    • Both volatilities peak in the same weeks: S&P 500 109% on 17 March 2020, DAX 80% on 25 March 2020
                                      • volatility clustering happens jointly: the covariance rises even with a constant correlation
                                    • Persistence 0.980 and 0.984: volatility shocks fade slowly, as in Chapter 5
                                    • Step 2 asks a different question: once the volatilities are removed, is the remaining correlation constant?

                                    Step 2: the estimated DCC

                                    Slide 33 in the PDF

                                    Step 2 estimates the two DCC parameters on the standardised residuals of step 1. The news coefficient is \(a = 0.0152\) with a standard error (SE) of 0.0035, and the persistence parameter is \(b = 0.9766\) with an SE of 0.0065. The sum \(a + b = 0.9918\) implies a half-life of a correlation shock of 85 days.

                                    The estimate of \(a\) is small, at the low end of the typical range, but more than four times its standard error, so the correlation does react to news. The sum close to one means that a deviation of the correlation from its long-run value lasts for months. The SEs come from the Hessian of the step-2 likelihood and ignore the estimation error of step 1, so they are too small; even allowing for this, \(a\) is clearly separated from zero.

                                    The likelihood ratio statistic against CCC is 131.8, far above the 5% critical value of 5.99. Constant correlation is rejected: the correlation between the S&P 500 and the DAX changes over time even after the changes in volatility have been removed.

                                    In brief

                                    QuantityEstimateSE
                                    \(a\)0.01520.0035
                                    \(b\)0.97660.0065
                                    \(a + b\)0.9918
                                    Half-life of a correlation shock (days)85
                                    LR against CCC131.8
                                    \(\chi^2(2)\) critical value, 5%5.99
                                    • \(a\) is small but well determined; \(a + b\) is close to 1: correlation shocks last for months
                                    • LR = 131.8 is far above 5.99: constant correlation is rejected
                                    • SE: standard errors from the step-2 Hessian (too small, they ignore step 1)

                                    DCC correlation of the S&P 500 and the DAX

                                    Slide 34 in the PDF
                                      Question for the room ▶ holds b at 0.98 and then at 0.99. With b = 0.99, is constant correlation still rejected?
                                      Answer

                                      Yes: the best a on the grid, 0.0074, gives LR = 121.0, far above 5.99. With b = 0.90 the best a is 0.0375 and LR = 102.1: the likelihood has a ridge along which a and b partly replace each other; the maximum, 131.8, is at the estimate (0.0152, 0.9766).

                                        Question for the room ▶ raises a up to 0.022 (b fixed). What happens to the standard deviation of the DCC correlation (0.083 at the estimate)?
                                        Answer

                                        It rises to 0.120, close to the 250-day rolling correlation (0.122), and the correlation ranges from 0.19 to 0.82. The LR falls to 100.1: the extra variation is noise that the likelihood penalises. With a = 0 the correlation is constant, 0.575, and LR = 0.

                                        The chart plots three estimates of the correlation on the vertical axis against time, daily from 4 January 2000 to 18 September 2026: the one-step-ahead DCC correlation \(\hat\rho_{12,t}\), the 250-day rolling correlation and the constant CCC value \(\bar q_{12} = 0.575\) as a dashed horizontal line. The shaded bands mark the crises of 2008 and 2020.

                                        The rolling correlation starts only in 2001, after its first window is complete, whereas the DCC correlation is available from the start of the sample. The two dynamic estimates follow the same long swings: high values in 2002–2003 and 2010–2012, lower values in 2005–2006 and 2018–2019.

                                        In brief

                                        • One-step-ahead DCC correlation \(\hat\rho_{12,t}\), the 250-day rolling correlation and the CCC value \(\bar q_{12}\); daily, 4 January 2000–18 September 2026
                                        TSA_ch14_dcc_estimation

                                        Interpreting the DCC correlation

                                        Slide 35 in the PDF

                                        The DCC correlation ranges from 0.29 in March 2000 to 0.75 in January 2010, around a mean of 0.57. It follows the same swings as the rolling correlation, but it reacts within days instead of a year, and it needs no choice of window. Its weights decline geometrically with the age of the observation, at rate \(b\), rather than being equal inside a window and zero outside it.

                                        The rolling estimate is more erratic: its standard deviation over the sample is 0.122, against 0.083 for DCC. It jumps when an extreme day enters the window and jumps again, in the opposite direction, exactly 250 days later when the day leaves it. The fall of the rolling correlation in 2025–2026 is an example. The asynchronous days of April 2025, described in the first section, enter the window and pull it down to about 0.21 for exactly 250 days, after which it jumps back. The DCC correlation registers the same days but forgets them gradually.

                                        In brief

                                        • The DCC correlation ranges from 0.29 (March 2000) to 0.75 (January 2010), around a mean of 0.57
                                        • It follows the rolling correlation, but reacts within days instead of a year and has no window to choose
                                          • the rolling estimate is more erratic (standard deviation 0.122 against 0.083) and jumps when an extreme day enters or leaves the window
                                        • The rolling estimate falls sharply in 2025–2026: the asynchronous days of April 2025 (Section 1) pull it down for exactly 250 days

                                        Correlations in crises

                                        Slide 36 in the PDF

                                        The table compares three periods: a calm period, 2017–2019, and the two crises, September 2008 – March 2009 (the months that followed the bankruptcy of Lehman Brothers, shown in the photograph) and February – April 2020. For each period it gives the average DCC correlation between the S&P 500 and the DAX and the average annualised GARCH volatility of the S&P 500. In the calm period the correlation averages 0.56 and the volatility 13%; in 2008–2009, 0.63 and 51%; in 2020, 0.70 and 58%.

                                        Correlations are higher in crises, but the change is moderate compared with the change in volatility, which roughly quadruples. Most of the rise in covariance during a crisis therefore comes from volatility, a smaller part from the correlation.

                                        The comparison needs care. Forbes and Rigobon (2002) showed that the sample correlation of returns rises mechanically when volatility rises, even if the underlying link between markets is unchanged: a larger common shock dominates the idiosyncratic noise. A higher correlation of raw returns in a crisis is therefore not by itself evidence of contagion. DCC reduces this effect because it measures correlation on the standardised residuals \(z_t\), not on the returns \(r_t\), so the rise in volatility is removed before the correlation is computed.

                                        In brief

                                        Correlations in crises Lehman Brothers, New York, 15 September 2008 Photo: Robert Scoble (2008); CC BY 2.0; Wikimedia Commons
                                        Period\(\bar\rho^{DCC}\)\(\bar\sigma^{S\&P}\) (%)
                                        Calm, 2017–20190.5613
                                        Sept 2008–March 20090.6351
                                        Feb–April 20200.7058
                                        • Average DCC correlation and average annualised S&P 500 volatility in each period
                                        • Correlations are higher in crises, but volatility rises far more
                                        • Caution Forbes and Rigobon (2002): the sample correlation rises mechanically when volatility rises, even if the link is unchanged; DCC reduces this effect because it works with \(z_t\), not with \(r_t\)

                                        DCC correlation in the crises of 2008 and 2020

                                        Slide 37 in the PDF
                                          Question for the room ▶ raises a up to 0.022. What happens to the gap between the mean DCC correlation of 2020 and that of the calm period (0.70 − 0.56 = 0.14)?
                                          Answer

                                          It widens to 0.18 (0.73 against 0.55); with a = 0 all three means are 0.575. The sample correlation of the returns (choose it in the box) gives 0.74 and 0.52: it rises mechanically with the volatility (Forbes and Rigobon), while DCC works with the standardised residuals.

                                            Question for the room ▶ raises the multiplier of α up to 2. Does the 2020 correlation still peak before the volatility?
                                            Answer

                                            No: with α doubled the correlation peak, 0.71, is on 12 June 2020, after the volatility peak (139% on 17 March 2020). The larger volatilities of March divide the shocks of those days, so the standardised residuals that drive the DCC recursion are smaller. With the multiplier 0.25 the 2008 volatility peaks only on 2 December 2008 (60%).

                                            The chart has two panels, one for each crisis: June 2008 – June 2009 on the left and December 2019 – August 2020 on the right. Each panel plots the DCC correlation of the S&P 500 and the DAX on the left axis, from 0 to 1, and the annualised GARCH volatility of the S&P 500, in percent, on the right axis, against time.

                                            The two axes make it possible to compare the timing of the two quantities, not their levels. In both panels the volatility line rises and falls steeply, while the correlation line moves in a much narrower range.

                                            In brief

                                            • DCC correlation of the S&P 500 and the DAX (left axis) and the GARCH volatility of the S&P 500 (right axis)
                                            TSA_ch14_dcc_estimation

                                            Interpreting the two crises

                                            Slide 38 in the PDF

                                            The two crises unfolded differently. In 2008 the volatility of the S&P 500 peaked at 90% on 16 October 2008, while the correlation rose slowly, from 0.57 to its maximum of 0.68, reached only on 23 June 2009, months after the volatility peak. The crisis started in the US financial system and spread to Europe gradually.

                                            In 2020 the order was reversed. The correlation jumped to 0.73 on 28 February 2020, before the volatility peak of 109% on 17 March 2020. A global shock, the COronaVIrus Disease 2019 (COVID-19) pandemic, hit both markets on the same days, and the run of joint falls drove the DCC recursion upwards quickly.

                                            Correlations and volatilities need not peak together. A model that ties them, such as a single rolling covariance, would mix the two stories; modelling them separately, as DCC does, shows the difference between a crisis that spreads and a shock that strikes everywhere at once.

                                            In brief

                                            • 2008: volatility peaks at 90% on 16 October 2008, while the correlation drifts up from 0.57 to 0.68 only by 23 June 2009
                                            • 2020: the correlation jumps to 0.73 on 28 February 2020, before the volatility peak (109% on 17 March 2020)
                                              • a global shock hit both markets on the same days: joint falls drive the DCC recursion
                                            • Correlations and volatilities need not peak together: modelling them separately, as DCC does, shows the difference

                                            Recap: S&P 500 and DAX

                                            Slide 39 in the PDF

                                            Step 1 gives persistent GARCH volatilities, with persistence 0.980 and 0.984, which peak together. Step 2 gives \(a + b = 0.9918\) and a likelihood ratio of 131.8 against CCC: the correlation is not constant. Correlations are higher in 2008 and 2020, but the largest change during a crisis is in volatility.

                                            In brief

                                            • Step 1: persistent GARCH volatilities that peak together
                                            • Step 2: \(a + b = 0.9918\); LR 131.8: the correlation is not constant
                                            • Higher correlations in 2008 and 2020, but the largest change is in volatility

                                            Section summary

                                            Two assets show the method; a portfolio usually has more. The next section estimates a single DCC model for five markets of very different kinds, including the Romanian equity market, the leu and Bitcoin, and compares their correlations in calm times and in the 2020 crash.

                                            Five markets: S&P 500, DAX, BET, EUR/RON, Bitcoin

                                            The section fits one DCC model to five assets, two large equity indices, the Romanian BET, the EUR/RON exchange rate and Bitcoin, and compares the average correlations in a calm period and in the COVID-19 crash.

                                            A five-asset DCC

                                            Slide 40 in the PDF

                                            The data are daily log returns on the 2 815 days on which all five markets trade, from 6 January 2015. The BET is the index of the Bucharest Stock Exchange, shown in the photograph. For EUR/RON the reference rate of the National Bank of Romania (BNR), fixed at 13:00 Bucharest time, is used, and for Bitcoin the close of the day in Coordinated Universal Time (UTC). The closing times differ even more than in the first section, so daily correlations of these series carry the asynchronous-trading bias discussed there.

                                            Step 1 gives five GARCH(1,1) models with persistence 0.961 for the S&P 500, 0.965 for the DAX, 0.941 for the BET, 0.985 for EUR/RON and 0.958 for Bitcoin. Step 2 gives \(\hat a = 0.0088\) (SE 0.0019) and \(\hat b = 0.9782\) (SE 0.0062), so the half-life of a correlation shock is 53 days. The likelihood ratio against CCC is 88.5, again far above the 5% critical value of 5.99.

                                            The model has 10 correlation pairs but only one pair of dynamics parameters \((a, b)\) for all of them. The 10 long-run correlations come from \(\bar{\mathbf{Q}}\), the sample correlation of the standardised residuals, so the optimiser of step 2 still works on two parameters only.

                                            In brief

                                            • Daily log returns on the 2 815 days on which all five markets trade, from 6 January 2015
                                              • EUR/RON: the BNR reference rate (fixed at 13:00 Bucharest time); BET: Bucharest Stock Exchange; Bitcoin: close of the day (UTC)
                                            • Step 1: GARCH persistence S&P 500 0.961, DAX 0.965, BET 0.941, EUR/RON 0.985, Bitcoin 0.958
                                            • Step 2: \(\hat a = 0.0088\) (SE 0.0019), \(\hat b = 0.9782\) (SE 0.0062); half-life 53 days; LR against CCC 88.5
                                              • 10 correlation pairs, one pair of dynamics parameters \((a, b)\) for all of them

                                            Dynamic correlations of four pairs

                                            Slide 41 in the PDF
                                              Question for the room ▶ shows each asset in turn. Which one has the highest volatility peak, and when?
                                              Answer

                                              Bitcoin: 281% on 13 March 2020. The S&P 500 reaches 121% on 17 March 2020, the DAX 84% on 13 March 2020, the BET 98% on 20 December 2018 (the day on which the DCC hedge ratio reaches 1.00), EUR/RON 13% on 21 May 2025.

                                                Question for the room ▶ raises a up to 0.02. What happens to the range of the S&P 500 – Bitcoin correlation (−0.05 … 0.42)?
                                                Answer

                                                It widens: from −0.33 to 0.62, with the mean still 0.21. With a = 0 the four correlations become constant, equal to Q̄: 0.55, 0.32, 0.21 and −0.07; this is the CCC model, rejected by LR = 88.5.

                                                The chart plots the one-step-ahead DCC correlations of four of the ten pairs on the vertical axis against time, from 2015 to 2026: the S&P 500 and the DAX, the BET and the DAX, the S&P 500 and Bitcoin, and the BET and EUR/RON. A dashed line marks zero, and the shaded band marks the COVID-19 crash, from mid-February to June 2020.

                                                The four lines are clearly separated in level: the S&P 500 and the DAX at the top, the BET and the DAX and the S&P 500 and Bitcoin in the middle, and the BET and EUR/RON around or below zero. Their movements over time are the subject of the next frame.

                                                In brief

                                                • One-step-ahead DCC correlations from the five-asset model; shaded: the COVID-19 crash, mid-February–June 2020
                                                TSA_ch14_dcc_panel

                                                Interpreting the four pairs

                                                Slide 42 in the PDF

                                                The S&P 500 and the DAX have the strongest link, with a mean correlation of 0.54 over 2015–2026. The BET and the DAX have a mean of 0.32 and move between 0.15 and 0.53, the maximum reached in April 2022. The jump of this pair at the start of the 2020 crash is the largest of the four.

                                                The S&P 500 and Bitcoin range from \(-0.05\) to 0.42, the maximum in May 2022, with a mean of 0.21. Before 2020 the correlation stayed low; after 2020 it stayed clearly higher. The diversification benefit of Bitcoin for an equity portfolio has shrunk.

                                                The BET and EUR/RON have a correlation close to zero, with a mean of \(-0.06\). The exchange rate is quoted as lei per euro, so a rise of EUR/RON is a weaker leu, and the small negative correlation means that a weaker leu tends to come with a weaker BET. One feature of the chart is a consequence of the model rather than of the data: all four lines share the same \((a, b)\), so pairs that might adjust at different speeds are forced to the same rhythm. This is the price of the parsimony of DCC.

                                                In brief

                                                • S&P 500 and DAX: the strongest link, mean 0.54; BET and DAX: mean 0.32, between 0.15 and 0.53 (April 2022)
                                                • S&P 500 and Bitcoin: from \(-\)0.05 to 0.42 (May 2022); mean 0.21
                                                  • the diversification benefit of Bitcoin shrank after 2020
                                                • BET and EUR/RON: close to zero (mean \(-\)0.06); a weaker leu tends to come with a weaker BET, a small negative link
                                                • One \((a, b)\) for all pairs: pairs with different speeds are forced to the same rhythm, the price of DCC's parsimony

                                                Average correlations: calm against crisis

                                                Slide 43 in the PDF
                                                  Question for the room ▶ raises a up to 0.02. What happens to the average correlation in the 2020 crash (0.19)?
                                                  Answer

                                                  It rises to 0.23, while the calm-period average falls to 0.10. With a = 0 the two matrices are identical (average 0.135): with constant correlations no rise in the crisis can be seen.

                                                  The chart shows two heatmaps of the average one-step-ahead DCC correlation matrix of the five assets. The left one averages over the calm period 2017–2019 and the right one over the COVID-19 crash, from mid-February to June 2020, 89 common trading days. Rows and columns are the five assets, each cell holds the average correlation, and the colour scale runs from \(-1\) (blue) to \(1\) (dark red).

                                                  The diagonals are ones by construction. The cells for EUR/RON are close to white in both matrices, while the cells linking the BET and Bitcoin with the large equity markets become clearly darker in the crisis panel.

                                                  In brief

                                                  • Average one-step-ahead DCC correlation matrix: 2017–2019 and mid-February–June 2020 (89 common trading days)
                                                  TSA_ch14_dcc_panel

                                                  Interpreting the two matrices

                                                  Slide 44 in the PDF

                                                  The average of the ten pairwise correlations rises from 0.12 in the calm period to 0.19 in the crash. The rise is uneven. For the BET and the DAX the average correlation goes from 0.28 to 0.48, for the S&P 500 and the BET from 0.22 to 0.38, and for the DAX and Bitcoin from 0.10 to 0.27. The S&P 500 and the DAX, already strongly linked, go from 0.55 to 0.64.

                                                  The largest rises concern the BET and Bitcoin, the assets weakly linked to the large markets in calm times, which are exactly the assets that were expected to diversify an equity portfolio. When a global shock arrives, these weak links become stronger, and the benefit of diversification shrinks when it matters most.

                                                  EUR/RON stays close to zero or slightly negative with all equity markets in both periods, from \(-0.03\) to \(-0.10\). In this set of five assets it is the only one that kept its diversifying role in the crash.

                                                  In brief

                                                  • The average pairwise correlation rises from 0.12 to 0.19
                                                    • BET and DAX: 0.28 \(\to\) 0.48; S&P 500 and BET: 0.22 \(\to\) 0.38; DAX and Bitcoin: 0.10 \(\to\) 0.27
                                                  • The largest rises are for the BET and for Bitcoin, which are weakly linked to the large markets in calm times: exactly the assets that were supposed to diversify
                                                  • EUR/RON stays near zero or slightly negative with all equity markets: the only diversifier in this set

                                                  Recap: Five markets

                                                  Slide 45 in the PDF

                                                  One DCC model describes the ten correlations of five assets with only two dynamics parameters, and the likelihood ratio of 88.5 rejects constant correlation. The correlations of the Romanian market with Europe and of Bitcoin with equities move a great deal. In the 2020 crash the weak links strengthened most, so diversification weakened in the crisis.

                                                  In brief

                                                  • One DCC for 5 assets: 2 dynamics parameters for 10 correlations
                                                  • Correlations of the Romanian market with Europe and of Bitcoin with equities move a lot
                                                  • In the 2020 crash the weak links strengthened most: diversification weakens in crises

                                                  Section summary

                                                  Correlations and volatilities are inputs; risk managers need numbers that can be checked against outcomes. The next section turns the conditional covariance matrix into the VaR of a portfolio and tests it on data that were not used for estimation.

                                                  Portfolio VaR 1% with DCC

                                                  The section computes the Value at Risk of a portfolio from \(\mathbf{H}_t\), presents the Kupiec and Christoffersen backtests and compares DCC, CCC and a static covariance matrix on an S&P 500 and DAX portfolio from 2015 to 2026.

                                                  From \(\mathbf{H}_t\) to the VaR of a portfolio

                                                  Slide 46 in the PDF

                                                  A portfolio with weights \(\mathbf{w}\) has the return \(r_{p,t} = \mathbf{w}^\top\mathbf{r}_t\), with conditional mean \(\mu_p = \mathbf{w}^\top\boldsymbol{\mu}\) and conditional volatility \(\sigma_{p,t} = \sqrt{\mathbf{w}^\top\mathbf{H}_t\mathbf{w}}\), by (14.1) applied to the conditional covariance matrix. The VaR 1% is the loss that is exceeded with probability 1%, minus the 1% quantile of the portfolio return:

                                                  \[\mathrm{VaR}_{0.01,t} = -q_{0.01}(r_{p,t}) = -\left(\mu_p + q\,\sigma_{p,t}\right).\]
                                                  (14.10)

                                                  The value of \(q\) depends on the assumed distribution of the standardised portfolio return. Under the Normal distribution \(q = z_{0.01} = -2.326\). Filtered historical simulation (FHS) uses instead the empirical 1% quantile of the past standardised portfolio returns \((r_{p,t} - \mu_p)/\sigma_{p,t}\), which takes the fat tails of the data into account while keeping the volatility dynamics of the model.

                                                  Three choices of \(\mathbf{H}_t\) are compared: DCC; CCC, which has GARCH volatilities but a constant correlation; and a static matrix, a single sample covariance matrix used for all days. The static VaR is the same every day, the CCC VaR moves with the volatilities only, and the DCC VaR moves with both volatilities and correlation.

                                                  In brief

                                                  • Portfolio return \(r_{p,t} = \mathbf{w}^\top\mathbf{r}_t\); conditional mean \(\mu_p = \mathbf{w}^\top\boldsymbol{\mu}\) and volatility \(\sigma_{p,t} = \sqrt{\mathbf{w}^\top\mathbf{H}_t\mathbf{w}}\)
                                                  • VaR 1%
                                                    • the loss exceeded with probability 1%, \(\mathrm{VaR}_{0.01,t} = -q_{0.01}(r_{p,t}) = -(\mu_p + q\,\sigma_{p,t})\)
                                                    • Normal: \(q = z_{0.01} = -2.326\)
                                                    • filtered historical simulation (FHS): \(q\) = the empirical 1% quantile of the past standardised portfolio returns \((r_{p,t} - \mu_p)/\sigma_{p,t}\)
                                                  • Three choices of \(\mathbf{H}_t\): DCC; CCC (GARCH volatilities with a constant correlation); static (one sample covariance matrix for all days)

                                                  Worked example: VaR 1% of two assets

                                                  Slide 47 in the PDF

                                                  Today's forecasts are daily volatilities \(\sigma_1 = 1.5\%\) and \(\sigma_2 = 1.2\%\), the weights are 50% and 50% and the mean is zero. In a calm state with \(\rho_t = 0.3\), (14.1) gives

                                                  \[\begin{aligned} \sigma_p^2 &= 0.25 \cdot 2.25 + 0.25 \cdot 1.44 \\ &\quad + 2 \cdot 0.25 \cdot 0.3 \cdot 1.8, \end{aligned}\]

                                                  where \(1.8 = 1.5 \cdot 1.2\) is the product of the volatilities. The portfolio volatility is \(\sigma_p = 1.092\%\) and the Normal VaR 1% is \(2.326 \cdot 1.092 = 2.54\%\).

                                                  In a crisis state with \(\rho_t = 0.9\) and the same volatilities, \(\sigma_p = 1.316\%\) and the VaR 1% is 3.06%. The jump of the correlation alone adds about a fifth to the VaR. In a real crisis the volatilities rise as well, so the actual increase is larger.

                                                  For a portfolio of 1 million lei, the VaR 1% is about 25.4 thousand lei in the calm case and 30.6 thousand lei in the crisis case. A model with a fixed correlation would report the first number on both days.

                                                  In brief

                                                  • Today's forecasts: \(\sigma_1 = 1.5\%\), \(\sigma_2 = 1.2\%\) (daily), weights 50% and 50%, mean 0
                                                  • Calm, \(\rho_t = 0.3\): \(\sigma_p^2 = 0.25 \cdot 2.25 + 0.25 \cdot 1.44 + 2 \cdot 0.25 \cdot 0.3 \cdot 1.8\), so \(\sigma_p = 1.092\%\) and VaR 1% \(= 2.326 \cdot 1.092 = 2.54\%\)
                                                  • Crisis, \(\rho_t = 0.9\) (the same volatilities): \(\sigma_p = 1.316\%\) and VaR 1% \(= 3.06\%\)
                                                    • the jump of the correlation alone adds a fifth to the VaR; in a real crisis the volatilities rise too
                                                  • For a portfolio of 1 million lei: VaR 1% of about 25.4 thousand lei in calm times and 30.6 thousand lei in the crisis case

                                                  Backtesting a VaR (1/2): frequency

                                                  Slide 48 in the PDF

                                                  A VaR model is judged by backtesting: comparing its forecasts with the returns that actually followed. A violation (violation) occurs on day \(t\) when \(r_{p,t} < -\mathrm{VaR}_{0.01,t}\), that is when the loss exceeds the VaR. A correct VaR 1% has violations on 1% of the days, and these days occur at random, not in clusters. The two properties are tested separately.

                                                  The test of Kupiec (1995) checks the frequency. With \(x\) violations in \(n\) days and observed rate \(\hat p = x/n\), the statistic

                                                  \[LR_{uc} = -2\log\frac{(1-0.01)^{n-x}\,0.01^{x}}{(1-\hat p)^{n-x}\,\hat p^{x}}\]
                                                  (14.11)

                                                  compares the likelihood of the observed violations under the nominal rate of 1% with their likelihood under the observed rate. The subscript \(uc\) stands for unconditional coverage. Under a correct model \(LR_{uc}\) follows approximately a \(\chi^2(1)\) distribution, and the model is rejected at 5% if \(LR_{uc} > 3.84\). The test rejects both too many violations, a VaR that is too low, and too few, a VaR that is too high.

                                                  In brief

                                                  • A violation on day \(t\)
                                                    • \(r_{p,t} < -\mathrm{VaR}_{0.01,t}\)
                                                    • a correct VaR 1% has violations on 1% of the days, at random times
                                                  • Kupiec test Kupiec (1995) (frequency)
                                                    • \(x\) violations in \(n\) days, \(\hat p = x/n\)
                                                    • \(LR_{uc} = -2\log\dfrac{(1-0.01)^{n-x}\,0.01^{x}}{(1-\hat p)^{n-x}\,\hat p^{x}} \sim \chi^2(1)\); reject at 5% if \(LR_{uc} > 3.84\)
                                                    • \(uc\): unconditional coverage; the ratio compares the likelihood of the violations under the rate 1% with that under the observed rate \(\hat p\)

                                                  Backtesting a VaR (2/2): independence

                                                  Slide 49 in the PDF

                                                  The test of Christoffersen (1998) checks independence: is a violation more likely on the day after a violation? It compares \(\pi_{11}\), the probability of a violation given a violation yesterday, with \(\pi_{01}\), the probability of a violation given no violation yesterday. For a correct model the two are equal, because a VaR that uses all available information leaves nothing predictable in the violations.

                                                  The statistic is built from counts: \(n_{ij}\) is the number of days in state \(j\) that follow a day in state \(i\), where state 1 is a violation and 0 is none. The estimates are \(\hat\pi_{01} = n_{01}/(n_{00}+n_{01})\) and \(\hat\pi_{11} = n_{11}/(n_{10}+n_{11})\), and \(\hat\pi\) is the overall violation rate. The statistic

                                                  \[LR_{ind} = -2\log\frac{(1-\hat\pi)^{n_{00}+n_{10}}\,\hat\pi^{\,n_{01}+n_{11}}}{(1-\hat\pi_{01})^{n_{00}}\,\hat\pi_{01}^{\,n_{01}}\,(1-\hat\pi_{11})^{n_{10}}\,\hat\pi_{11}^{\,n_{11}}}\]

                                                  has a numerator that assumes \(\pi_{01} = \pi_{11}\), no clustering, and a denominator that lets the two probabilities differ. Under independence it follows approximately \(\chi^2(1)\); a small p-value rejects independence.

                                                  A backtest is meaningful only out of sample. All parameters are estimated on 2000–2014, 3 712 days, and the filters are then run forward on 2015–2026, 2 893 days, without re-estimating. Each day's VaR uses only information available on the previous day, as it would in practice.

                                                  In brief

                                                  • Christoffersen test Christoffersen (1998) (independence)
                                                    • is a violation more likely the day after a violation?
                                                    • compares \(\pi_{11} = P(\text{violation} \mid \text{violation yesterday})\) with \(\pi_{01} = P(\text{violation} \mid \text{no violation yesterday})\)
                                                  • The statistic
                                                    • \(LR_{ind} = -2\log\dfrac{(1-\hat\pi)^{n_{00}+n_{10}}\,\hat\pi^{\,n_{01}+n_{11}}}{(1-\hat\pi_{01})^{n_{00}}\,\hat\pi_{01}^{\,n_{01}}\,(1-\hat\pi_{11})^{n_{10}}\,\hat\pi_{11}^{\,n_{11}}} \sim \chi^2(1)\)
                                                    • \(n_{ij}\): the number of days in state \(j\) after a day in state \(i\) (1: violation, 0: none); \(\hat\pi_{01} = n_{01}/(n_{00}+n_{01})\), \(\hat\pi_{11} = n_{11}/(n_{10}+n_{11})\)
                                                    • \(\hat\pi\): the overall violation rate; the numerator assumes \(\pi_{01} = \pi_{11}\) (no clustering); a small \(p\)-value rejects independence
                                                  • Out-of-sample design
                                                    • estimate all parameters on 2000–2014 (3 712 days), then run the filters forward on 2015–2026 (2 893 days) without re-estimating

                                                  VaR 1% of an S&P 500 and DAX portfolio, 2015–2026

                                                  Slide 50 in the PDF

                                                  Worked example

                                                  \[\mathbf{H}_t=\mathbf{D}_t\mathbf{R}_t\mathbf{D}_t,\quad \sigma_{p,t}=\sqrt{\mathbf{w}^\top\mathbf{H}_t\mathbf{w}},\quad \mathrm{VaR}_{0.01,t}=-(\mu_p+z_{0.01}\,\sigma_{p,t}),\ \ z_{0.01}=-2.326\]
                                                      Question for the room ▶ raises ρt up to 1. What is the VaR 1% for ρt = 1?
                                                      Answer

                                                      3.14%: σp = 0.5 · 1.5 + 0.5 · 1.2 = 1.35%, the largest possible value with these volatilities. With ρt = −0.5 the VaR 1% is 1.60%, about 16.0 thousand lei for 1 million lei.

                                                        Question for the room ▶ moves to the levels 2.5% and 5%. At 5%, does any model pass the Christoffersen independence test?
                                                        Answer

                                                        No: all four have p ≤ 0.0004; the violations come in clusters on the turbulent days. At 2.5% only DCC with FHS passes the Kupiec test (79 violations, a rate of 2.73%, p = 0.433).

                                                          Question for the room ▶ moves to the levels 2.5% and 5%. At 5%, what is the violation rate of the static VaR?
                                                          Answer

                                                          2.83% (82 violations against 145 expected): too few, LRuc = 33.6, p < 0.001. At 1% the static VaR passed the test (0.90%, p = 0.577): it is too high in calm years, which at 1% offsets the violations of the crises and at 5% does not. DCC with FHS has a rate of 4.60% at 5% (p = 0.314).

                                                          The chart plots, against time from 2015 to 2026, the daily return in percent of a portfolio with 50% in the S&P 500 and 50% in the DAX, together with minus the VaR 1% of two models: DCC with filtered historical simulation, a line that moves, and the static covariance matrix, a flat line. Red triangles mark the 48 violations of the DCC-FHS model and green triangles the 26 violations of the static model.

                                                          The DCC line follows the volatility of the returns: it stays close to the returns in calm years and drops far below them in March 2020, when the returns themselves swing widely. The static line stays at the same level throughout. The green triangles of the static model are concentrated in a few episodes, most visibly in early 2020; the red triangles of DCC are spread across the whole period.

                                                          In brief

                                                          • Daily return of a 50/50 portfolio and minus VaR 1% from DCC with filtered historical simulation and from a static covariance matrix; triangles: violations
                                                          TSA_ch14_portfolio_var

                                                          Interpreting the backtest

                                                          Slide 51 in the PDF

                                                          With 2 893 out-of-sample days, a correct VaR 1% should have about 29 violations. The table gives, for each model, the number and rate of violations, the Kupiec statistic \(LR_{uc}\) and its p-value, the p-value of the Christoffersen independence test and the number of violations in February – April 2020. DCC with Normal quantiles has 65 violations, a rate of 2.25%, \(LR_{uc} = 33.6\) with a p-value below 0.001 and an independence p-value of 0.003. DCC with FHS has 48 violations, 1.66%, \(LR_{uc} = 10.6\), p-value 0.001, independence p-value 0.008. CCC with Normal quantiles has 57 violations, 1.97%, \(LR_{uc} = 21.4\), p-value below 0.001, independence p-value 0.030. The static model has 26 violations, 0.90%, \(LR_{uc} = 0.3\), p-value 0.577, independence p-value 0.001.

                                                          The dynamic models with Normal quantiles have too many violations, because the standardised residuals have fatter tails than the Normal distribution, as the kurtosis of step 1 showed and as Chapter 5 found for single series. FHS replaces \(-2.326\) by the empirical quantile \(q = -2.59\) and brings the rate down to 1.66%, still above 1% but much closer. The remaining violations are also less clustered than for the static model: in February – April 2020 every dynamic model has 4 violations.

                                                          The static VaR passes the frequency test, and this is the instructive case. Its 26 violations are close to the expected 29, but 10 of them fall in February – April 2020, and after a violation the next day is also a violation with probability 11.5%, so independence is rejected. The static VaR is 2.95% every day: too high in calm years and far too low in a crash. Its correct average frequency comes from errors in both directions that offset each other. A dynamic VaR adapts in both directions; its remaining problem is the quantile, not the covariance.

                                                          In brief

                                                          ModelViolationsRate\(LR_{uc}\)\(p\)\(p\) indep.Feb–Apr 2020
                                                          DCC, Normal652.25%33.6\(<\) 0.0010.0034
                                                          DCC, FHS481.66%10.60.0010.0084
                                                          CCC, Normal571.97%21.4\(<\) 0.0010.0304
                                                          Static, Normal260.90%0.30.5770.00110
                                                          • Expected: about 29 violations in 2 893 days
                                                          • Dynamic models with Normal quantiles have too many violations: the residuals have fat tails (Chapter 5); FHS (\(q = -2.59\)) brings the rate down to 1.66%
                                                          • The static VaR passes the frequency test, but its violations cluster: 10 in February–April 2020 alone, and after a violation the next day is a violation with probability 11.5%
                                                          • The static VaR is too high in calm years (2.95% every day) and far too low in a crash; a dynamic VaR adapts in both directions

                                                          Recap: Portfolio VaR

                                                          Slide 52 in the PDF

                                                          The portfolio volatility is \(\sigma_{p,t} = \sqrt{\mathbf{w}^\top\mathbf{H}_t\mathbf{w}}\) and the VaR 1% is \(-(\mu_p + q\,\sigma_{p,t})\). A VaR is backtested out of sample with the Kupiec test, for the frequency of violations, and the Christoffersen test, for their independence. Dynamic covariances remove most of the clustering of violations; fat tails require a better quantile than the Normal one.

                                                          In brief

                                                          • \(\sigma_{p,t} = \sqrt{\mathbf{w}^\top\mathbf{H}_t\mathbf{w}}\); VaR 1% \(= -(\mu_p + q\,\sigma_{p,t})\)
                                                          • Backtest out of sample: Kupiec (frequency) and Christoffersen (independence)
                                                          • Dynamic covariances fix the clustering of violations; fat tails need a better quantile than the Normal

                                                          Section summary

                                                          The VaR uses the whole matrix \(\mathbf{H}_t\) through one quadratic form. The next application uses its elements separately: the ratio of a covariance to a variance gives the number of units of a hedging asset to sell, and DCC lets this number change every day.

                                                          Dynamic hedge ratios

                                                          The section derives the minimum-variance hedge ratio and its hedging effectiveness, makes the ratio dynamic with DCC, and compares static, rolling and DCC hedges of the BET with the DAX out of sample.

                                                          The minimum-variance hedge ratio

                                                          Slide 53 in the PDF

                                                          An investor holds one unit of an asset \(s\), for example Romanian shares represented by the BET, and sells \(h\) units of a hedging asset \(f\), for example a futures contract on the DAX. The hedged return is \(r_{s,t} - h\,r_{f,t}\), and its variance is \(\sigma_s^2 - 2h\,\sigma_{sf} + h^2\sigma_f^2\), where \(\sigma_s^2\) and \(\sigma_f^2\) are the variances of the two returns and \(\sigma_{sf}\) their covariance. Setting the derivative with respect to \(h\) to zero gives the minimum-variance hedge ratio (hedge ratio) of Ederington (1979):

                                                          \[h^* = \frac{\sigma_{sf}}{\sigma_f^2} = \rho\,\frac{\sigma_s}{\sigma_f}.\]
                                                          (14.12)

                                                          The static version of \(h^*\) is the slope of the ordinary least squares (OLS) regression of \(r_s\) on \(r_f\), because the OLS slope is exactly a covariance divided by a variance. The dynamic version, proposed by Kroner and Sultan (1993), uses the elements of \(\mathbf{H}_t\) from an MGARCH model, \(h_t^* = h_{sf,t}/h_{ff,t}\), and therefore changes with the forecast covariance and variance every day.

                                                          The hedging effectiveness (hedging effectiveness) measures the share of variance removed: \(HE = 1 - \Var(r_s - h\,r_f)/\Var(r_s)\). Substituting the constant optimal ratio into the hedged variance gives \(\sigma_s^2(1 - \rho^2)\), so \(HE = \rho^2\). A hedge is only as good as the correlation between the asset and the hedging instrument.

                                                          In brief

                                                          • Hold one unit of asset \(s\) (e.g. Romanian shares, the BET) and sell \(h\) units of a hedging asset \(f\) (e.g. a DAX futures contract)
                                                            • hedged return \(r_{s,t} - h\,r_{f,t}\), with variance \(\sigma_s^2 - 2h\,\sigma_{sf} + h^2\sigma_f^2\)
                                                            • \(\sigma_s^2\), \(\sigma_f^2\): the variances of the two returns; \(\sigma_{sf}\): their covariance
                                                          • Setting the derivative in \(h\) to zero: \(h^* = \dfrac{\sigma_{sf}}{\sigma_f^2} = \rho\,\dfrac{\sigma_s}{\sigma_f}\) Ederington (1979)
                                                            • static \(h^*\): the OLS slope of \(r_s\) on \(r_f\)
                                                            • dynamic: \(h_t^* = h_{sf,t}/h_{ff,t}\) from an MGARCH model Kroner and Sultan (1993); \(h_{sf,t}\), \(h_{ff,t}\): elements of \(\mathbf{H}_t\)
                                                          • Hedging effectiveness
                                                            • \(HE = 1 - \Var(r_s - h\,r_f)/\Var(r_s)\), the share of the variance removed
                                                            • with \(h = h^*\) constant: \(HE = \rho^2\); a hedge is only as good as the correlation

                                                          Worked example: a hedge ratio

                                                          Slide 54 in the PDF

                                                          Take daily volatilities \(\sigma_s = 1.2\%\) and \(\sigma_f = 1.0\%\) and a correlation \(\rho = 0.8\). The covariance is \(\sigma_{sf} = \rho\,\sigma_s\sigma_f = 0.8 \cdot 1.2 \cdot 1.0 = 0.96\), and by (14.12) the hedge ratio is \(h^* = 0.96/1 = 0.96\): the investor sells 0.96 units of \(f\) for each unit of \(s\).

                                                          The hedging effectiveness is \(\rho^2 = 0.64\), so the hedge removes 64% of the variance. The volatility of the hedged position is \(\sigma_s\sqrt{1 - \rho^2} = 0.72\%\) instead of 1.2%. Volatility falls less than variance, by 40% instead of 64%, because it is the square root.

                                                          If tomorrow the DCC model forecasts \(\rho_t = 0.5\) and \(\sigma_{f,t} = 2\%\), with \(\sigma_s\) unchanged, the ratio becomes \(0.5 \cdot 1.2/2 = 0.3\). Both the lower correlation and the higher volatility of the hedging asset reduce the number of units to sell, and the position must be adjusted.

                                                          In brief

                                                          • Given: \(\sigma_s = 1.2\%\), \(\sigma_f = 1.0\%\) (daily), \(\rho = 0.8\)
                                                          • Step 1: \(\sigma_{sf} = \rho\,\sigma_s\sigma_f = 0.8 \cdot 1.2 \cdot 1.0 = 0.96\)
                                                          • Step 2: \(h^* = \sigma_{sf}/\sigma_f^2 = 0.96/1 = 0.96\): sell 0.96 units of \(f\) per unit of \(s\)
                                                          • Step 3: \(HE = \rho^2 = 0.64\); the hedged volatility is \(\sigma_s\sqrt{1 - \rho^2} = 0.72\%\) instead of 1.2%
                                                          • If tomorrow DCC forecasts \(\rho_t = 0.5\) and \(\sigma_{f,t} = 2\%\), the hedge ratio becomes \(0.5 \cdot 1.2/2 = 0.3\): the position must be adjusted

                                                          Hedging the BET with the DAX, 2015–2026

                                                          Slide 55 in the PDF

                                                          Worked example

                                                          \[h^*=\frac{\sigma_{sf}}{\sigma_f^2}=\rho\,\frac{\sigma_s}{\sigma_f},\qquad HE=1-\frac{\mathrm{Var}(r_s-h^*r_f)}{\mathrm{Var}(r_s)}=\rho^2\]
                                                              Question for the room Set ρ = −0.8. What are h* and HE?
                                                              Answer

                                                              h* = −0.96: buy 0.96 units of f per unit of s; HE = ρ2 = 0.64, as for ρ = 0.8. With ρ = 1 the volatility of the hedged position is 0 (h* = 1.2).

                                                                Question for the room Is there a window (60–500 days) for which rolling OLS removes more risk than DCC?
                                                                Answer

                                                                No: the best result is 18.6%, with 80 days, below 19.1% for DCC; 300 days give only 17.0%. The choice of the window moves the result by more than a percentage point, and DCC needs no such choice.

                                                                  Question for the room ▶ lengthens the OLS window up to 500 days. Does rolling OLS then remove more risk than DCC (19.1%)?
                                                                  Answer

                                                                  No: between 17.0% (300 days) and 18.0% (500 days). With a = 0 in the DCC the effectiveness falls to 15.5%, still above the static ratio (13.6%): the correlation is then constant, but the ratio q̄ σBET,t/σDAX,t still follows the GARCH volatilities.

                                                                  The chart plots three hedge ratios, in units of DAX sold per unit of BET, against time from 2015 to 2026: the DCC ratio \(h_t = \hat\rho_t\hat\sigma_{BET,t}/\hat\sigma_{DAX,t}\), the rolling OLS slope over the previous 250 days and the static OLS ratio estimated on 2005–2014, a dashed horizontal line. The data are daily returns on common trading days of the two markets since 2005, and all parameters of the DCC model are estimated on 2005–2014.

                                                                  The DCC ratio is the most volatile line, with short spikes, the highest at the end of 2018. The rolling OLS ratio moves in steps, and its large jump in early 2020 lasts about a year, the length of its window. The static line lies above both dynamic ratios for most of the period.

                                                                  In brief

                                                                  • Daily returns on common trading days since 2005; all parameters estimated on 2005–2014; DCC: \(h_t = \hat\rho_t\hat\sigma_{BET,t}/\hat\sigma_{DAX,t}\); rolling OLS: the slope over the previous 250 days
                                                                  TSA_ch14_hedging

                                                                  Interpreting the hedge ratios

                                                                  Slide 56 in the PDF

                                                                  The static ratio of 0.51 reflects the period 2005–2014, which includes the 2008 crisis, when the BET and the DAX were closely linked. After 2015 the DCC ratio averages 0.29 and ranges from 0.07 to 1.00, the maximum reached on 20 December 2018. The static ratio thus over-hedges in most of the later period.

                                                                  The hedging effectiveness out of sample, over 2 893 days, is 13.6% for the static ratio, 18.0% for the rolling OLS ratio and 19.1% for the DCC ratio. The dynamic ratios remove more risk, and DCC does so without the choice of a window.

                                                                  All three values are modest. The DCC correlation between the BET and the DAX averages 0.32 since 2015, so \(\rho^2\) is about a tenth, and no ratio can remove much of the variance of the BET with the DAX. A daily DCC ratio also implies daily rebalancing and trading costs, which are not counted in these figures; after costs, the advantage of DCC over the static ratio would be smaller.

                                                                  In brief

                                                                  • The static ratio 0.51 reflects 2005–2014, which includes the 2008 crisis; after 2015 the DCC ratio averages 0.29, from 0.07 to 1.00 (20 December 2018)
                                                                  • Hedging effectiveness out of sample (2 893 days): static 13.6%, rolling OLS 18.0%, DCC 19.1%
                                                                    • the dynamic ratios remove more risk, and DCC does so without choosing a window
                                                                  • All effectiveness values are modest: the BET–DAX DCC correlation averages 0.32 since 2015, so \(\rho^2\) is small
                                                                    • a daily DCC ratio also implies daily trading costs, which are not counted here

                                                                  Recap: Hedging

                                                                  Slide 57 in the PDF

                                                                  The dynamic minimum-variance hedge ratio is \(h_t^* = h_{sf,t}/h_{ff,t} = \rho_t\sigma_{s,t}/\sigma_{f,t}\), and the hedging effectiveness is the share of variance removed. For the BET hedged with the DAX, DCC removes 19.1% of the variance out of sample, against 13.6% for the static ratio. A hedge must be judged out of sample and after trading costs.

                                                                  In brief

                                                                  • \(h_t^* = h_{sf,t}/h_{ff,t} = \rho_t\sigma_{s,t}/\sigma_{f,t}\); \(HE = 1 - \Var(\text{hedged})/\Var(\text{unhedged})\)
                                                                  • BET with DAX: DCC 19.1% against static 13.6% out of sample
                                                                  • Judge a hedge out of sample and after trading costs

                                                                  Section summary

                                                                  The applications have shown what the conditional covariance matrix adds in practice. The next section places the chapter among the earlier ones and lists what the DCC model cannot do.

                                                                  Links with other chapters and limits

                                                                  The section links multivariate GARCH models with univariate GARCH, VAR models and cointegration, and lists the main limits of DCC and the extensions that address them.

                                                                  Every MGARCH model contains univariate GARCH models: the diagonal of \(\mathbf{H}_t\) consists of conditional variances, as in Chapter 5. In DCC the link is literal, since the univariate GARCH models are step 1 of the estimation.

                                                                  A VAR model of Chapter 6 describes spillovers in the mean, for example whether yesterday's DAX return helps to forecast today's BET return. A BEKK model with a full matrix \(\mathbf{A}\) describes spillovers in the variance, whether yesterday's shock to the DAX raises today's variance of the BET. The two combine naturally: a VAR for the mean \(\boldsymbol{\mu}_t\) and an MGARCH model for \(\mathbf{H}_t\). Diebold and Yilmaz (2009) built spillover indices from the variance decompositions of a VAR.

                                                                  Cointegration, from Chapter 7, is a long-run link between price levels; correlation is a short-run link between returns. The two are distinct. Two prices can be highly correlated from day to day and still drift apart for ever, or be cointegrated with a low daily correlation. A vector error correction model (VECM) can carry DCC errors, with equilibrium in the mean and dynamic correlation in the shocks.

                                                                  In brief

                                                                  • Chapter 5
                                                                    • every MGARCH model contains univariate GARCH models
                                                                    • in DCC they are literally step 1
                                                                  • Chapter 6
                                                                    • a VAR models spillovers in the mean (e.g. whether yesterday's DAX return helps to forecast today's BET return)
                                                                    • BEKK with a full \(\mathbf{A}\) models spillovers in the variance
                                                                    • the two combine: a VAR for \(\boldsymbol{\mu}_t\) and an MGARCH for \(\mathbf{H}_t\); spillover indices from variance decompositions Diebold and Yilmaz (2009)
                                                                  • Chapter 7
                                                                    • cointegration is a long-run link between price levels
                                                                    • correlation is a short-run link between returns
                                                                    • two prices can be highly correlated day by day and drift apart for ever, or cointegrated with low daily correlation
                                                                    • a VECM can carry a DCC error term: equilibrium in the mean, dynamic correlation in the shocks

                                                                  Limits and extensions

                                                                  Slide 59 in the PDF

                                                                  DCC buys its parsimony with restrictions. A single pair \((a, b)\) drives all correlations, so pairs that adjust at different speeds are forced to the same rhythm, as the five-market model showed. The volatilities come from separate univariate GARCH models, so DCC has no spillovers between volatilities. And the Normal quasi-likelihood says nothing about the tails: for risk measures, Student \(t\) innovations or filtered historical simulation are needed, as the backtest showed.

                                                                  Correlations also react asymmetrically: they rise more after joint falls than after joint rises, which the asymmetric DCC of Cappiello, Engle and Sheppard (2006) captures. For hundreds of assets even DCC becomes hard to handle, because the sample matrix \(\bar{\mathbf{Q}}\) is noisy and the step-2 likelihood requires inverting a large matrix every day; factor models, composite likelihood and shrinkage of \(\bar{\mathbf{Q}}\) are the usual answers.

                                                                  Finally, correlation measures linear dependence only. Two assets can have a moderate correlation and still crash together much more often than a Normal model implies. Joint extremes are described by copulas and by tail dependence; Longin and Solnik (2001) studied the correlation of international equity markets in the tails directly.

                                                                  In brief

                                                                  • One \((a, b)\) for all pairs; no spillovers between volatilities in DCC
                                                                  • Normal QML underestimates the tails: use Student \(t\) innovations or FHS for risk
                                                                  • Asymmetry: correlations react more to joint falls (asymmetric DCC, Cappiello, Engle and Sheppard (2006))
                                                                  • Hundreds of assets: factor models, composite likelihood and shrinkage of \(\bar{\mathbf{Q}}\)
                                                                  • Correlation is linear dependence; for joint extremes see copulas and tail dependence Longin and Solnik (2001)

                                                                  Section summary

                                                                  DCC is the workhorse of multivariate volatility because it splits a hard problem into univariate GARCH models and a correlation recursion with two parameters. Its limits, a common speed for all pairs, no volatility spillovers and Normal tails, are known and have standard extensions.

                                                                  Possible contribution of AI

                                                                  The section describes what an artificial intelligence (AI) assistant can do in multivariate volatility modelling and what must be checked every time.

                                                                  Possible contribution of AI

                                                                  Slide 60 in the PDF

                                                                  An AI assistant can quickly write a first version of the code the chapter needs: the DCC correlation likelihood, a BEKK filter or the Kupiec test. It can also give a second explanation, phrased differently, of the ideas that are harder on first reading, such as the decomposition \(\mathbf{D}_t\mathbf{R}_t\mathbf{D}_t\), correlation targeting or the formula of the hedge ratio.

                                                                  The greatest benefit comes in exploration: DCC models for many pairs at once, such as the markets of Central and Eastern Europe (CEE), sectors or currencies, and a first look at how their correlations behaved in crises. The prompt on the slide is an example of a precise request: it names the data, the alignment on common trading days, the two steps of estimation, correlation targeting and the out-of-sample comparison with static OLS. The more of these choices the request fixes, the easier the result is to check.

                                                                  In brief

                                                                  • Code
                                                                    • a first draft of the DCC likelihood, of a BEKK filter, of a Kupiec test
                                                                  • Explanation
                                                                    • a second explanation of \(\mathbf{D}_t\mathbf{R}_t\mathbf{D}_t\), of correlation targeting, of the hedge-ratio formula
                                                                  • Exploration
                                                                    • DCC for many pairs (CEE markets, sectors, currencies) and their behaviour in crises
                                                                  • Example prompt
                                                                    • Write Python code that fits a GARCH(1,1) to daily BET and DAX log returns on common trading days, estimates a DCC(1,1) model by two-step maximum likelihood with correlation targeting, plots the dynamic hedge ratio and compares its out-of-sample hedging effectiveness with static OLS.

                                                                  Checks you must run

                                                                  Slide 61 in the PDF

                                                                  The first check is the library. The arch package has no DCC model, so an answer that calls arch_model(..., vol='DCC') has invented a function. The second is timing: \(\mathbf{Q}_t\) must use \(\mathbf{z}_{t-1}\), not \(\mathbf{z}_t\); otherwise the VaR and the hedge ratio use tomorrow's information, and the backtest looks better than any real forecast could. The third is the rescaling (14.8): every \(\mathbf{R}_t\) must have ones on the diagonal and correlations below one in absolute value.

                                                                  The other checks concern data and evaluation. Prices must be aligned on common trading days before returns are computed, and asynchronous closes must be kept in mind. Models must be estimated on one period and tested on another. A simple test of any DCC code is a simulation: generate data from a DCC with known \((a, b)\) and check that the code recovers them. Every reference suggested by an assistant must exist, and its digital object identifier (DOI) must be checked.

                                                                  In brief

                                                                  • The library: arch has no DCC model; an AI answer that calls arch_model(..., vol='DCC') has invented it
                                                                  • The timing: \(\mathbf{Q}_t\) must use \(\mathbf{z}_{t-1}\), not \(\mathbf{z}_t\); otherwise the VaR and the hedge use tomorrow's information
                                                                  • The rescaling: \(\rho_{ij,t} = q_{ij,t}/\sqrt{q_{ii,t}q_{jj,t}}\), and every \(\mathbf{R}_t\) must have ones on the diagonal and \(|\rho| < 1\)
                                                                  • The calendars: align prices on common trading days before computing returns; watch asynchronous closes
                                                                  • The evaluation: estimate on one period and test on another; simulate a DCC with known \((a, b)\) and check that the code recovers it
                                                                  • Every cited reference: it must exist; check the DOI

                                                                  Section summary

                                                                  An assistant shortens the writing of code but does not replace the checks specific to multivariate models: the timing of the recursion, the validity of each correlation matrix, the alignment of calendars and an honest out-of-sample evaluation.

                                                                  Summary

                                                                  The summary gathers the chapter's main ideas, the formulas to remember and the self-assessment questions.

                                                                  Key takeaways

                                                                  Slide 62 in the PDF

                                                                  Portfolio risk, hedging and contagion depend on the conditional covariance matrix \(\mathbf{H}_t\), which changes over time. VEC is the most general linear model but has far too many parameters; BEKK is positive definite by construction and measures volatility spillovers. CCC and DCC split \(\mathbf{H}_t = \mathbf{D}_t\mathbf{R}_t\mathbf{D}_t\) and are estimated in two steps, and step 2 of DCC needs only \((a, b)\) whatever the number of assets.

                                                                  The data show that real correlations move: through the integration of the Romanian market with Europe, through crises and through the arrival of Bitcoin as a risk asset. Constant correlation is rejected for the S&P 500 and the DAX and for the five-market model. The applications turn \(\mathbf{H}_t\) into a portfolio VaR 1%, judged by backtesting, and into dynamic hedge ratios, judged out of sample.

                                                                  In brief

                                                                  • Portfolio risk, hedging and contagion depend on time-varying covariances \(\mathbf{H}_t\)
                                                                  • VEC is general but has too many parameters; BEKK is positive definite by construction and measures spillovers
                                                                  • CCC and DCC split \(\mathbf{H}_t = \mathbf{D}_t\mathbf{R}_t\mathbf{D}_t\) and are estimated in two steps; DCC needs only \((a, b)\) in step 2
                                                                  • Real correlations move: integration, crises, new assets; constant correlation is rejected
                                                                  • Applications: portfolio VaR 1% with backtesting; dynamic hedge ratios judged out of sample

                                                                  The table gathers the chapter's formulas: the conditional correlation (14.3); the BEKK model (14.6); the decomposition (14.7) used by CCC and DCC; the DCC recursion with its rescaling (14.8); the half-life \(\ln 0.5/\ln(a + b)\); the portfolio VaR 1% (14.10); and the hedge ratio (14.12) with the hedging effectiveness.

                                                                  The formulas form a chain. The decomposition separates what univariate GARCH models already handle from the correlations; the DCC recursion makes the correlations dynamic; the half-life summarises their persistence; and the VaR and the hedge ratio turn the resulting \(\mathbf{H}_t\) into decisions.

                                                                  In brief

                                                                  QuantityFormula
                                                                  Correlation\(\rho_{ij,t} = h_{ij,t}/\sqrt{h_{ii,t}h_{jj,t}}\)
                                                                  BEKK(1,1)\(\mathbf{H}_t = \mathbf{C}\mathbf{C}^\top + \mathbf{A}^\top\boldsymbol{\varepsilon}_{t-1}\boldsymbol{\varepsilon}_{t-1}^\top\mathbf{A} + \mathbf{B}^\top\mathbf{H}_{t-1}\mathbf{B}\)
                                                                  CCC, DCC\(\mathbf{H}_t = \mathbf{D}_t\mathbf{R}_t\mathbf{D}_t\),   \(\mathbf{D}_t = \mathrm{diag}(\sigma_{i,t})\)
                                                                  DCC\(\mathbf{Q}_t = (1 - a - b)\bar{\mathbf{Q}} + a\,\mathbf{z}_{t-1}\mathbf{z}_{t-1}^\top + b\,\mathbf{Q}_{t-1}\),   \(\rho_{ij,t} = q_{ij,t}/\sqrt{q_{ii,t}q_{jj,t}}\)
                                                                  Half-life\(\ln 0.5/\ln(a + b)\)
                                                                  Portfolio VaR 1%\(-(\mathbf{w}^\top\boldsymbol{\mu} + q_{0.01}\sqrt{\mathbf{w}^\top\mathbf{H}_t\mathbf{w}})\)
                                                                  Hedge ratio\(h_t^* = h_{sf,t}/h_{ff,t}\),   \(HE = 1 - \Var(r_s - h r_f)/\Var(r_s)\)

                                                                  Self-assessment (1/2)

                                                                  Slide 64 in the PDF

                                                                  The first question applies the parameter count of the diagonal BEKK: for \(N = 3\) the constant matrix has \(N(N+1)/2 = 6\) distinct elements and the diagonals of \(\mathbf{A}\) and \(\mathbf{B}\) have \(2N = 6\), so 12 in total. The second applies (14.3) and the positive-definiteness condition: with \(h_{11} = 1\), \(h_{22} = 4\) and \(h_{12} = 1.2\) the correlation is \(1.2/\sqrt{4} = 0.6\), and the determinant \(1 \cdot 4 - 1.44\) is positive, so the matrix is positive definite.

                                                                  The third question is one step of (14.8) with \(a = 0.02\), \(b = 0.97\), \(\bar q_{12} = 0.4\) and \(z_{t-1} = (1, -1)\). The weight of the long-run term is 0.01 and the cross-product of the residuals is \(-1\), so

                                                                  \[\begin{aligned} q_{12,t} &= 0.01 \cdot 0.4 + 0.02 \cdot (-1) + 0.97 \cdot 0.4 \\ &= 0.372: \end{aligned}\]

                                                                  moves in opposite directions lower the correlation. The fourth computes the half-life for \(a + b = 0.95\): \(\ln 0.5/\ln 0.95 \approx 13.5\) days.

                                                                  In brief

                                                                  • Question: how many parameters does a diagonal BEKK(1,1) have for \(N = 3\)?
                                                                    • Answer: \(N(N+1)/2 + 2N = 6 + 6 = 12\)
                                                                  • Question: \(h_{11} = 1\), \(h_{22} = 4\), \(h_{12} = 1.2\). What is the correlation, and is the matrix positive definite?
                                                                    • Answer: \(\rho = 1.2/\sqrt{4} = 0.6\); yes, \(1 \cdot 4 - 1.44 > 0\)
                                                                  • Question: DCC with \(a = 0.02\), \(b = 0.97\), \(\bar q_{12} = 0.4\), \(q_{11,t-1} = q_{22,t-1} = 1\), \(q_{12,t-1} = 0.4\) and \(z_{t-1} = (1, -1)\). What is \(q_{12,t}\)?
                                                                    • Answer: \(0.01 \cdot 0.4 + 0.02 \cdot (-1) + 0.97 \cdot 0.4 = 0.372\); opposite moves lower the correlation
                                                                  • Question: what is the half-life of correlation shocks when \(a + b = 0.95\)?
                                                                    • Answer: \(\ln 0.5/\ln 0.95 \approx 13.5\) days

                                                                  Self-assessment (2/2)

                                                                  Slide 65 in the PDF

                                                                  The first question applies (14.12): with \(\sigma_s = 2\%\), \(\sigma_f = 1\%\) and \(\rho = 0.5\) the hedge ratio is \(0.5 \cdot 2/1 = 1\) and the hedging effectiveness \(\rho^2 = 0.25\). The second concerns backtesting: 26 violations in 2 500 days are close to the expected 25, so the frequency test is passed, but 10 violations in a single month are a cluster that the independence test rejects, and they show a VaR that does not adapt to crises, like the static VaR of the S&P 500 and DAX portfolio.

                                                                  The third question returns to asynchronous trading: news from New York reaches Bucharest only on the next trading day, so the daily returns of the same calendar day do not match and the daily correlation is lower than the weekly one. The fourth concerns contagion: a higher correlation of returns in a crisis does not prove contagion by itself, because higher volatility raises the sample correlation mechanically (Forbes and Rigobon, 2002); the comparison must use correlations of standardised residuals. Chapter 15 reviews the whole course.

                                                                  In brief

                                                                  • Question: \(\sigma_s = 2\%\), \(\sigma_f = 1\%\), \(\rho = 0.5\). What are the hedge ratio and the hedging effectiveness?
                                                                    • Answer: \(h^* = 0.5 \cdot 2/1 = 1\); \(HE = 0.25\)
                                                                  • Question: a VaR 1% has 26 violations in 2 500 days, 10 of them in one month. Is it a good VaR?
                                                                    • Answer: the frequency is fine (expected 25), but the violations cluster: it fails the independence test and does not adapt to crises
                                                                  • Question: why can the daily correlation of the S&P 500 and the BET be lower than the weekly one?
                                                                    • Answer: asynchronous trading: news from New York reaches Bucharest the next day, so daily returns of the same day do not match
                                                                  • Question: does a higher correlation in a crisis prove contagion?
                                                                    • Answer: not by itself: higher volatility raises the sample correlation mechanically Forbes and Rigobon (2002); compare correlations of standardised residuals
                                                                  • Next: Chapter 15, review and exam preparation; check yourself with the quiz of this chapter

                                                                  Section summary

                                                                  The chapter extended volatility modelling from one series to several. The central object is the conditional covariance matrix, and the central idea of DCC is to model volatilities and correlations separately, with univariate GARCH models for the first and a two-parameter recursion for the second. On real data the correlations move, rise in crises most for the weakest links, and change the risk of a portfolio and the size of a hedge.

                                                                  The daily log returns (in %) of the chapter’s samples, on the common trading days (S&P 500, DAX, BET, EUR/RON, Bitcoin; data/market, EODHD), are not revised and reproduce the numbers of the lecture exactly. The estimates come from the chapter’s code: the GARCH(1,1) models of step 1 (arch package) and the parameters (a, b) of step 2 with their standard errors, for the S&P 500 and the DAX (2000–2026), for the five-asset panel (since 2015), for the 2000–2014 estimation sample of the VaR and for the 2005–2014 sample of the hedge. The browser runs the recursions again: the GARCH variance filter with the given parameters, the DCC recursion, the rolling correlations and slopes, the correlation log-likelihood, the VaR and the Kupiec and Christoffersen tests. Variants that do not appear in the lecture (other windows, other values of a and b, another α of step 1 with the same persistence, other VaR levels, k-day returns, the forecasts without new shocks, the quarterly means) are extra runs of the chapter’s code on the same data; the charts of the “Worked example” boxes, “Asynchronous trading and the bias of daily correlations”, “Two conditions on Hₜ”, “The curse of dimensionality”, “Reading a and b”, “Step 1: two univariate GARCH models”, “Step 2: the estimated DCC”, “Correlations in crises”, “A five-asset DCC”, “Interpreting the backtest” and “Interpreting the hedge ratios” are extra charts for slides that give only numbers.

                                                                  References