Blockchain Applications

Lecture 6: Stablecoins II – Systemic Depegging Pressure and Portfolios

Lecture 6: Stablecoins II – Systemic Depegging Pressure and Portfolios

Daniel Traian PELE

Bucharest University of Economic Studies
IDA Institute Digital Assets
Blockchain Research Center
Master Blockchain for Business (MBB), FABIZ
25 November 2026

Lecture Question and Outline

Slide 2 in the PDF

Lecture 5 watched one pool and one pair of coins. The question now is whether the pressure on the whole stablecoin system can be measured in one number, and whether that number is useful for forecasting. The two parts of the question are answered separately, because a good measure of the present is not necessarily a good predictor of the future.

The route has five steps. Tokens with different anchors, the dollar, the euro and gold, are first made comparable through a reference adjustment. Quantile regression and the Lasso are then introduced, together with the idea of the Financial Risk Meter (FRM), which reads the penalty selected by a Lasso as a risk measure. The De-pegging Pressure Index (DPI) is built from these pieces, and its regimes and responses to market events are examined. The fourth step looks inside the model at which coins are linked to which in their lower tails, and the last one separates measuring from forecasting.

The main source is the working paper Reference-adjusted Quantile Lasso reveals regime-dependent stablecoin depegging pressure by Fan, Sha, Härdle & Pele (2026), which is based on an MSc thesis (Fan, 2026). The data are daily closes of 10 stablecoins from End-of-Day Historical Data (EODHD), from 1 January 2020 to 23 October 2026. Stablecoins trade every day, while gold and the euro–dollar exchange rate (EUR/USD) trade only on weekdays, so Friday's close of the reference is carried over the weekend.

In brief

  • Question: Lecture 5 watched one pool. Can we measure pressure on the whole stablecoin system in one number, and is that number useful for forecasting?
  • Outline
    • comparing tokens with different anchors: reference adjustment
    • quantile regression and the Lasso; the Financial Risk Meter (FRM) idea
    • the De-pegging Pressure Index (DPI): construction, regimes, events
    • inside the tail links: who is linked to whom
    • measure or forecast: two different tasks
  • Main source: Fan, Sha, Härdle & Pele (2026), Reference-adjusted Quantile Lasso reveals regime-dependent stablecoin depegging pressure, working paper (2026); based on Fan (2026), MSc thesis, Humboldt-Universität zu Berlin
  • Data: daily closes of 10 stablecoins from End-of-Day Historical Data (EODHD), 1 January 2020 – 23 October 2026
  • Gold and EUR/USD trade Monday–Friday, while stablecoins trade every day: Friday's close of the reference is carried over the weekend

From Lecture 5: Observed Depeg, Pressure Index and Breach Probability

Slide 3 in the PDF

The Silicon Valley Bank (SVB) weekend of March 2023 can be described by three different numbers, which answer three different questions. The observed depeg of Lecture 5 is a fact about prices, in basis points (bp, one bp is 0.01%): the daily close of USDC on 11 March 2023 was \(d = -285\) bp, 2.85% below one dollar.

The contemporaneous pressure of this lecture is the DPI, a penalty implied by a model, with no unit. Computed from the 90 days up to 11 March, the DPI on 12 March 2023 was 0.52. On its own the number has no meaning; it is read against its own history, where 0.52 lies in the fourth of five quintiles, between 0.43 and 0.55. The future probability of Lecture 5 is the estimate of an alarm that a breach will occur within the next 24 hours, a number between 0 and 1.

The first number says what happened, the second how strongly the coins move together in their lower tails now, and the third what may happen next. The DPI is therefore not a probability of a depeg. Whether it helps to forecast one is a separate question, tested at the end of the lecture.

In brief

  • The SVB weekend, March 2023, seen through three different numbers
  • Observed depeg (Lecture 5): a price fact, in basis points (bp)
    • USDC daily close on 11 March 2023: \(d = -285\) bp against the dollar
  • Contemporaneous pressure (this lecture): the De-pegging Pressure Index (DPI), a model-implied penalty with no unit
    • DPI on 12 March 2023, from the 90 days to 11 March: 0.52 (fourth of five quintiles, 0.43–0.55)
  • Future probability (Lecture 5): an alarm's estimate of \(P(\text{breach within the next 24 hours})\), a number in \([0, 1]\)
  • The three answer different questions: what happened, how strongly coins move together in their lower tails now, what may happen next
  • The DPI is not a depeg probability; whether it helps to forecast is tested at the end

Learning Outcomes

Slide 4 in the PDF
  • Compute a reference-adjusted peg deviation for dollar, euro and gold tokens
  • Write down the check loss, the 5% Quantile Lasso and the generalised approximate cross-validation (GACV) criterion, and explain what the selected penalty measures
  • Construct and interpret the DPI, its regimes and event responses, and judge their robustness
  • Distinguish in-sample fit from out-of-sample forecasting
  • Optional (appendix): judge whether the DPI is useful for building a portfolio

Abbreviations

Slide 5 in the PDF
bp basis point, 0.01%
BUSD Binance USD (stablecoin)
CVIX crypto volatility index
DAI Dai (MakerDAO stablecoin)
DD drawdown (max DD: maximum drawdown)
df degrees of freedom
DPI De-pegging Pressure Index
DXY US dollar index
EODHD End-of-Day Historical Data (data provider)
ES expected shortfall
EUR euro
EURS STASIS Euro (stablecoin)
FABIZ Faculty of Business Administration in Foreign Languages, ASE
FRM Financial Risk Meter
FTX FTX crypto exchange (collapsed November 2022)
GACV generalised approximate cross-validation
GMV global minimum variance (portfolio)
GUSD Gemini Dollar (stablecoin)
HAC heteroskedasticity and autocorrelation consistent
HHI Herfindahl–Hirschman index
IDRT Rupiah Token (stablecoin)
L1 L1 norm, \(\|\beta\|_1 = \sum_k |\beta_k|\) (Lasso penalty)
Lasso least absolute shrinkage and selection operator
MAD median absolute deviation
n.s. not significant
OOS out-of-sample
PAXG Pax Gold (gold-backed token)
pp percentage points
S&P 500 Standard & Poor's 500 stock index
SD standard deviation
SE standard error
sUSD Synthetix USD (stablecoin)
SVB Silicon Valley Bank (failed March 2023)
TUSD TrueUSD (stablecoin)
USD US dollar
USDC USD Coin (stablecoin)
USDP Pax Dollar (stablecoin)
USDT Tether (stablecoin)
VIX CBOE volatility index
XAU gold, one troy ounce

Reading and Tools

Slide 6 in the PDF

The main reading is the working paper by Fan, Sha, Härdle & Pele (2026) and the MSc thesis on which it is based, Large Quantile Lasso FRM for Stable Coin Depegging Risks (Fan, 2026), together with the keynote Stable Coin Risks by Fan, Härdle, Pele and Sha of 16 September 2026; the slides that quote its results name it in their source line. The Financial Risk Meter is described by Mihoci, Althof, Chen & Härdle (2020) and Yu, Härdle, Borke & Benschop (2023). Two related approaches to links between the tails of financial institutions are the tail-event network of Härdle, Wang & Yu (2016) and the CoVaR of Adrian & Brunnermeier (2016).

The methods come from three classic papers: quantile regression from Koenker & Bassett (1978), the Lasso from Tibshirani (1996), and the Lasso for quantile regression with many predictors from Belloni & Chernozhukov (2011). The arbitrage that keeps stablecoins near their peg is studied by Lyons & Viswanath-Natraj (2023). The Quantinar video unit FRM for Cryptos introduces the FRM. Each Quantile Lasso fit is a linear programme, so the index is computationally heavy: a daily index for 10 coins over 2,397 days needs 23,970 equations, and each is fitted for 25 values of the penalty, about 600,000 linear programmes.

In brief

We propose a measure of systemic risk, ΔCoVaR, defined as the change in the value at risk of the financial system conditional on an institution being under distress relative to its median state.

— Adrian & Brunnermeier (2016)

Lecture 5 watched one stablecoin pair, USD Coin (USDC) priced in Tether (USDT), and asked whether a depeg of that pair could be seen coming. A supervisor, an exchange or a treasury that holds several stablecoins needs a wider view: is the stablecoin system as a whole under pressure today, and do the coins tend to lose their pegs together? This lecture asks whether that pressure can be measured in one number, and whether the number helps to forecast.

The number studied is the De-pegging Pressure Index (DPI) of a recent working paper by Fan, Sha, Härdle & Pele (2026), which builds on the idea of the Financial Risk Meter (FRM). For each stablecoin and each day, a quantile regression with a Lasso penalty relates the coin's lower tail to the other coins and to a few macro variables; the penalty that the data select is read as a measure of how strongly the coins move together in their lower tails, and the DPI averages it over the coins. The lecture explains the two tools, quantile regression and the Lasso, and then asks what the index measures, how robust it is, and whether it forecasts anything.

The empirical results use daily prices of 10 stablecoins from the data provider End-of-Day Historical Data (EODHD), from 1 January 2020 to 23 October 2026. They include dollar tokens, a euro token and a gold token, so the first step is to put tokens with different anchors on one scale.

The De-pegging Pressure Index (DPI) of 10 stablecoins, daily, April 2020 – October 2026, on a log scale; the colour of each point gives its quintile, from low (quintile 1) to severe (quintile 5). The dotted lines mark six market events, and the dashed red line the end of the paper's sample on 31 May 2026. The index moves in regimes, with its highest values in late 2025.
The De-pegging Pressure Index (DPI) of 10 stablecoins, daily, April 2020 – October 2026, on a log scale; the colour of each point gives its quintile, from low (quintile 1) to severe (quintile 5). The dotted lines mark six market events, and the dashed red line the end of the paper's sample on 31 May 2026. The index moves in regimes, with its highest values in late 2025.

What you will be able to do

  • Compute a reference-adjusted peg deviation for dollar, euro and gold tokens.
  • Write down the check loss, the 5% Quantile Lasso and the generalised approximate cross-validation (GACV) criterion, and explain what the selected penalty measures.
  • Construct and interpret the DPI, its regimes and event responses, and judge their robustness.
  • Distinguish in-sample fit from out-of-sample forecasting.
  • Optional (Appendix): judge whether the DPI is useful for building a portfolio.

The Appendix holds optional material on bad prints in daily prices, a decomposition of the model's fit and stablecoin portfolios.

Lecture plan

  1. Comparing Different AnchorsReference adjustment for dollar, euro and gold tokens, peg quality across ten coins, and a persistent depeg.
  2. Quantile Regression and the LassoThe check loss, the 5% quantile in small samples, the Quantile Lasso and the selected penalty.
  3. The De-pegging Pressure IndexConstruction, the index from 2020 to 2026, its trend, and its responses to six market events.
  4. Tail Dependence between StablecoinsWhich coins are linked to which, sparser models under pressure, fit against forecast, and concentration.
  5. SummaryMeasure or forecast, the main numbers of the lecture and four interpretation exercises.
  6. Appendix: Further MaterialBad prints in daily prices, a Shapley decomposition of the fit, and four long-only stablecoin portfolios.

Comparing Different Anchors

Not every stablecoin promises a dollar. Before any model can compare tokens, their deviations must be measured against their own anchors. The section defines the reference adjustment and checks what it removes, compares the peg quality of ten tokens, and follows one token, sUSD, whose depeg did not close.

Reference Adjustment

Slide 7 in the PDF

Some stablecoins track other assets: STASIS Euro (EURS) tracks the euro, Pax Gold (PAXG) one troy ounce of gold (XAU), and Rupiah Token (IDRT) the Indonesian rupiah. If \(P_{i,t}\) is the dollar price of token \(i\) on day \(t\) and \(B_{i,t}\) the dollar value of its reference asset, the reference-adjusted deviation and the daily reference-adjusted return, both in bp, are

\[d_{i,t} = 10^4\left(\frac{P_{i,t}}{B_{i,t}} - 1\right), \qquad r_{i,t} = 10^4\,\Delta \ln\frac{P_{i,t}}{B_{i,t}}\]
(6.1)

where \(\Delta\) is the change from the previous day. The reference is \(B = 1\) for dollar tokens, the EUR/USD exchange rate for EURS and the gold price XAU/USD for PAXG.

In the worked example EURS trades at 1.1194 USD while EUR/USD is 1.1202, so \(d = 10^4\,(1.1194/1.1202 - 1) = -7.1\) bp: the token is slightly below its euro peg. Without the adjustment, every 1% move of the euro against the dollar would look like a 100 bp "depeg" of EURS. The adjustment puts all tokens on one economic scale, the distance from their own promise, before any model sees them.

In brief

  • Not every stablecoin targets the dollar: EURS tracks the euro, PAXG one ounce of gold, IDRT the rupiah
  • Let \(P_{i,t}\) be the dollar price of token \(i\) and \(B_{i,t}\) the dollar value of its reference asset:
    \[ d_{i,t} = 10^4\left(\frac{P_{i,t}}{B_{i,t}} - 1\right) \text{ bp}, \qquad r_{i,t} = 10^4\,\Delta \ln\frac{P_{i,t}}{B_{i,t}} \]
    • \(B = 1\) for dollar tokens; EUR/USD for EURS; XAU/USD for PAXG
  • Worked example: EURS at 1.1194 USD while EUR/USD = 1.1202
    \[ d = 10^4\left(\frac{1.1194}{1.1202} - 1\right) = -7.1 \text{ bp} \]
    • without adjustment, a 1% move of the euro would look like a 100 bp “depeg” of EURS
  • The adjustment puts all tokens on one economic scale before any model sees them

Effect of the Reference Adjustment

Slide 8 in the PDF

The adjustment should remove the part of a token's dollar price that comes from its reference asset. The chart checks how much it removes for the two non-dollar tokens of the sample, EURS and PAXG. For each, it plots the standard deviation over the previous 90 days of three daily returns, in bp per day on a log scale, from January 2021 to May 2026: the token's dollar return \(\Delta \ln P_{i,t}\) (blue), the dollar return of its reference \(\Delta \ln B_{i,t}\) (gold, dashed), and the reference-adjusted return \(r_{i,t}\) (red).

For PAXG the adjustment works as intended. Its dollar return has a standard deviation of 96 bp a day, almost the same as gold's 94 bp, and the two are correlated 0.87. After the adjustment the standard deviation falls to 48 bp: half of the day-to-day movement of PAXG in dollars was the price of gold.

For EURS it does not. The token's dollar return has a standard deviation of 146 bp a day, while the euro moves only 39 bp, and the correlation between the two is 0.10. The adjusted return has a standard deviation of 147 bp, as large as before: almost all of the movement of EURS is its own noise, not the euro. The largest euro move of the sample makes the point: on 10 April 2025 EUR/USD rose 227 bp, EURS rose 159 bp in dollars, and so it fell 68 bp against the euro.

The adjustment therefore puts each token's target in the right unit, so that a move of the euro is not counted as a depeg, but it cannot remove noise that does not come from the reference. A thinly traded token stays noisy after any adjustment, and this noise enters every model that uses it.

In brief

Effect of the Reference Adjustment
  • 90-day standard deviation of daily returns in bp, January 2021 – May 2026, log scale; blue: token in USD, gold: reference in USD, red: reference-adjusted
  • PAXG: 96 bp a day in USD, gold 94 bp, correlation of the two 0.87; adjusted 48 bp: the adjustment halves the noise
  • EURS: 146 bp a day in USD, the euro only 39 bp, correlation 0.10; adjusted 147 bp: the token's own noise dominates
  • 10 April 2025: EUR/USD \(+227\) bp; EURS \(+159\) bp in USD, \(-68\) bp against the euro
  • The adjustment puts the target in the right unit; it cannot remove noise that does not come from the reference
  • Source: own computation on EODHD daily prices of PAXG and EURS, gold and EUR/USD, January 2021 – May 2026
BA_L6_refadj

Peg Quality, January 2020 – May 2026

Slide 9 in the PDF

The table measures peg quality for ten tokens with three statistics of the reference-adjusted deviation (6.1), in bp, computed from EODHD daily prices: the median absolute deviation from the peg, med \(|d|\); the 5% quantile \(q_{0.05}\), which describes the lower tail; and the percentage of days at least 100 bp below the reference, % \(d \leq -100\). The sample has 2,343 days per coin, except EURS with 2,315, because EODHD has no prices for it on 43 days in November–December 2025. For six coins, a column gives the median absolute deviation that the paper reports from CoinGecko prices.

The tokens fall into two tiers. The large dollar tokens stay within a few bp: median absolute deviations of 1.3 bp for USDC, 2.8 for DAI, 3.6 for USDT and 3.7 for Binance USD (BUSD), with fewer than 1% of days 100 bp below the peg. The small, non-dollar or crypto-collateralised tokens are tens of bp away and have deep lower tails: the Gemini Dollar (GUSD) has a median of 17.7 bp, PAXG 43.5, EURS 50.9 and Synthetix USD (sUSD) 56.3, and sUSD spent 26.0% of days at least 100 bp below its peg, with a 5% quantile of −1,883 bp. Pax Dollar (USDP) and TrueUSD (TUSD) lie in between, with medians of 7.3 and 7.7 bp.

The price source changes the levels but not the ranking. The medians from EODHD are lower than those from CoinGecko: 1.3 against 4.11 bp for USDC. The downside rates agree closely, 0.09% of days for USDT in both sources and 26.0% against 25.0% for sUSD. sUSD, a crypto-collateralised token, traded at 0.95 USD in March 2025, 0.76 in March 2026 and 0.09 on 23 October 2026, a persistent depeg.

In brief

EODHD pricesCoinGecko (paper)EODHD prices
coinmed \(|d|\)\(q_{0.05}\)% \(d \leq -100\)med \(|d|\)coinmed \(|d|\)\(q_{0.05}\)% \(d \leq -100\)
USDC1.3\(-5.0\)0.174.11GUSD17.7\(-110.0\)6.32
DAI2.8\(-9.0\)0.687.06PAXG43.5\(-124.4\)9.13
USDT3.6\(-11.3\)0.095.37EURS50.9\(-251.9\)17.11
BUSD3.7\(-8.6\)0.09sUSD56.3\(-1\,883.4\)25.99
USDP7.3\(-47.5\)0.217.63
TUSD7.7\(-37.8\)2.09
  • \(d\) in bp; % \(d \leq -100\): share of days at least 100 bp below the reference; 2 343 days per coin, EURS 2 315 (no EODHD prices for 43 days, November–December 2025)
  • Two tiers: large dollar tokens within a few bp; small, non-dollar or crypto-collateralised tokens tens of bp off, with deep lower tails
  • The price source changes levels, not the ranking: USDC median 1.3 bp on EODHD, 4.11 bp on CoinGecko
  • Downside rates agree across sources: USDT 0.09% of days in both; sUSD 26.0% vs 25.0%
  • sUSD (crypto-collateralised): 0.95 USD in March 2025, 0.76 in March 2026, 0.09 on 23 October 2026, a persistent depeg
  • Source: own computation on EODHD daily prices; CoinGecko column: Fan et al. (2026), working paper

Two Tiers of Peg Quality

Slide 10 in the PDF
    Question for the room With the sample to 31 May 2026, the 95% quantile of |d| is 1,883 bp for sUSD and its median 56.3 bp. The ▶ button moves the end of the sample. With the sample extended to 23 October 2026, how does the 95% quantile of |d| of sUSD change?
    Answer

    It roughly doubles, to 3,929 bp, while the median rises only from 56.3 to 62.1 bp and the share of days at least 100 bp below the peg from 26.0% to 30.3%. The 145 added days are the collapse of sUSD to 0.09 USD: they are 6% of the sample, but they fill the upper tail. The other nine coins change much less, and the two tiers keep the same members.

    The chart shows the distribution of the absolute deviation \(|d_{i,t}|\) of each coin as a box plot, from January 2020 to May 2026, with the coins ordered by their median and the vertical axis on a log scale. The red line is the median, the box covers the middle half of the days (the interquartile range), and the whiskers reach the 5% and 95% quantiles.

    On a log scale each step up multiplies the deviation by ten. The upper whisker, the 95% quantile, spans three orders of magnitude: 25 bp for USDC and almost 2,000 bp for sUSD. The two tiers of the table are visible as two groups of boxes, the dollar tokens on the left and GUSD, PAXG, EURS and sUSD on the right.

    In brief

    • Distribution of \(|d_{i,t}|\), January 2020 – May 2026, coins ordered by median; log scale
    • The 95% whisker spans three orders of magnitude: 25 bp for USDC, almost 2 000 bp for sUSD
    • Source: own computation on EODHD daily prices of 10 stablecoins, January 2020 – May 2026
    BA_L6_pegquality

    A Persistent Depeg: sUSD

    Slide 11 in the PDF
      Question for the room After September 2025 sUSD first closed below 0.95 USD on 24 November 2025, and it has stayed below 0.95 USD since 14 December 2025. The ▶ button lowers the threshold price. From which date does sUSD stay below 0.50 USD without interruption?
      Answer

      From 20 August 2026; it first closed below 0.50 on 12 June 2026, and on 23 October 2026 it closed at 0.09 USD. The fall from 0.95 to 0.50 took about nine months, with returns above each level before it was left for good. No arbitrage at par closes this gap, because redemption at 1 USD is not available: the depeg is a slow default, not a weekend dislocation.

      The chart follows the daily closing price of sUSD in USD from July 2024 to October 2026, against its peg of one dollar. sUSD is crypto-collateralised, issued by the Synthetix protocol. Since January 2025 it closed below 0.99 USD on 607 of 661 days, below 0.95 USD on every day from 24 November 2025, and at 0.09 USD on 23 October 2026.

      The contrast with SVB is instructive. During the SVB weekend USDC was below its peg for about two days, and arbitrage closed the gap once redemption was possible again. For sUSD redemption at par is not available, so no arbitrage closes the gap: the depeg is a slow default, in which the token loses its value step by step.

      In brief

      • Crypto-collateralised (Synthetix); below 0.99 USD on 607 of 661 days since January 2025; below 0.95 from 24 November 2025; 0.09 USD on 23 October 2026
      • Unlike Silicon Valley Bank (SVB; two days), no arbitrage closes this gap: redemption at par is not available; a depeg can be a slow default
      • Source: own computation on EODHD daily prices of sUSD, 2025 – 23 October 2026
      BA_L6_susd

      Section summary

      Reference adjustment measures each token against its own anchor, so dollar, euro and gold tokens can enter one model. Peg quality falls into two tiers, a few bp for the large dollar tokens and tens of bp with deep lower tails for the others, and a depeg without redemption at par can last for months.

      Quantile Regression and the Lasso

      The DPI is built from two statistical tools. Quantile regression estimates a quantile, here the 5% quantile, the lower tail, instead of a mean; the Lasso adds a penalty that sets the slopes of unimportant predictors to zero. The section introduces both, shows why a 5% quantile in a short window is noisy and why it counts ranks rather than depth, and explains how the penalty selected by the data becomes a risk measure.

      The Check Loss

      Slide 12 in the PDF
        Question for the room With n = 20 and τ = 0.05 the total check loss is flat between −120 and −35 bp. The ▶ button raises τ. At τ = 0.10, which values of q minimise the total loss?
        Answer

        Every q from −35 to −12 bp, the second and third smallest observations: n τ = 2 is again an integer, so the loss is flat between two observations (total loss 148.10). Whenever 20 τ is an integer the minimiser is an interval; for other τ it is one observation (τ = 0.06: q = −35 bp). For τ = 0.5 the interval is [2, 3] bp, around the median of 2 bp quoted on the slide. In a 90-day window n τ = 4.5, so the 5% quantile is unique, but still set by five days.

        The mean of a variable minimises the expected squared error, and the median minimises the expected absolute error. Any quantile can be found in the same way with an asymmetric absolute loss, the check loss of Koenker & Bassett (1978):

        \[\rho_\tau(e) = e\,\big(\tau - \mathbb{1}\{e < 0\}\big), \qquad q_\tau = \arg\min_q E\,\rho_\tau(Y - q)\]
        (6.2)

        Here \(\tau\) is the quantile level, \(e\) the error and \(\mathbb{1}\{e < 0\}\) equals 1 when the error is negative. Quantile regression replaces the constant \(q\) by a linear function of predictors \(\mathbf x_i\) and minimises \(\sum_i \rho_\tau(y_i - \alpha - \mathbf x_i'\beta)\), a problem that can be solved by linear programming.

        For \(\tau = 0.05\), an observation below the fitted line costs \(0.95|e|\) and one above it only \(0.05|e|\), so the fit is pushed down until about 5% of the observations lie below it: it describes the lower tail. The left panel of the chart draws the check loss for \(\tau = 0.05\), 0.5 and 0.95; at \(\tau = 0.5\) it is symmetric, half the absolute error.

        The right panel applies the loss to 20 daily deviations. For \(\tau = 0.5\) the total loss has a clear minimum at the median, 2 bp. For \(\tau = 0.05\) the loss is flat between −120 and −35 bp: every value in that range fits equally well.

        In brief

        • The \(\tau\)-quantile minimises an asymmetric absolute loss (Koenker & Bassett (1978)):
          \[ \rho_\tau(e) = e\,\big(\tau - \mathbb 1\{e < 0\}\big), \qquad q_\tau = \arg\min_q E\,\rho_\tau(Y - q) \]
        • Quantile regression: \(\;\min_{\alpha, \beta} \sum_i \rho_\tau(y_i - \alpha - \mathbf x_i'\beta)\), a linear programme
        • For \(\tau = 0.05\) a residual below the line costs \(0.95|e|\), above it \(0.05|e|\): the fit is pushed to the lower tail
        • Right: 20 daily deviations; for \(\tau = 0.5\) the minimum is the median (2 bp); for \(\tau = 0.05\) the loss is flat between \(-120\) and \(-35\) bp
        BA_L6_checkloss

        Worked Example: Observations Defining a 5% Quantile

        Slide 13 in the PDF

        With \(n = 20\) observations and \(\tau = 0.05\), \(n\tau = 1\): exactly one observation should lie below the 5% quantile. The total check loss is

        \[\sum_i \rho_{0.05}(y_i - q) = 116.55 \quad \text{for every } q \in [-120, -35]\]

        so any value between the smallest observation, −120 bp, and the second smallest, −35 bp, is a valid 5% quantile: the estimate is not unique.

        In a window of 90 days, \(n\tau = 4.5\) is not an integer, the minimiser is unique and equal to the 5th smallest observation. Even so, five days carry all the information about the lower tail. A rolling 5% quantile model is therefore noisy by construction: one extreme day entering or leaving the window can move the estimate a lot. This is why the DPI is analysed in regimes and in windows around events, not day by day.

        In brief

        • With \(n = 20\) and \(\tau = 0.05\): \(n\tau = 1\), so one observation lies below the quantile
          \[ \sum_i \rho_{0.05}(y_i - q) = 116.55 \;\text{ for every } q \in [-120, -35] \]
        • Any value between the smallest and second-smallest observation is a valid 5% quantile: the estimate is not unique
        • In a 90-day window \(n\tau = 4.5\) is not an integer: the minimiser is unique, the 5th smallest observation; still, five days carry the whole lower tail
        • Interpretation
          • rolling 5% quantile models are noisy by construction; one day entering or leaving the window can move the estimate
          • this is why the DPI is analysed in regimes and event windows, not day by day

        Sensitivity of the 5% Quantile to Extreme Days

        Slide 14 in the PDF
          Question for the room With W = 90 days the −285 bp close of USDC on 11 March 2023 moves the 5% quantile of the next day only from −1.90 to −2.15 bp. The ▶ button shortens the window. From which window length W does that close itself become the 5% quantile on 12 March 2023?
          Answer

          From W = 20 days down. The 5% quantile of W values is the value of rank ⌊0.05 (W − 1)⌋ + 1 from the bottom; it is the lowest value only while 0.05 (W − 1) < 1, that is W ≤ 20. With W = 21 to 30 it is the second lowest of the window to 11 March (−5.21 bp, the close of 10 March), and from 15 March the second SVB close, −79.3 bp. A short window gives a quantile that jumps with single days; a long window gives one that ignores how deep the worst day was.

          The worked example showed that in a 90-day window the 5% quantile is the 5th lowest observation. The chart shows what this means in real data, around the SVB weekend of March 2023. The thin lines are the daily deviations of USDC and DAI; the step lines are the 5% quantile of the 90 days before each day; the dotted lines mark the days when the SVB weekend enters and leaves the window.

          On 11 March 2023 USDC closed at −285 bp. Once that day is in the window, the 5% quantile moves only from −1.90 to −2.15 bp: the −285 bp day becomes the lowest of the 90, and the 5th lowest, which is all the estimator reports, barely changes. The next days bring −79, −11 and −8 bp. By 15 March, these four SVB days and the −5.2 bp of 10 March are the five lowest, and the quantile is −5.21 bp. With further low days later in March, it reaches its 2023 low of −10.5 bp on 22 March and stays there until 9 June.

          When the SVB days leave the window between 10 and 13 June, the quantile returns to −5.0 bp. The size of the worst day never entered the estimate; only its rank did. A rolling 5% quantile thus reacts to how many bad days there are in the window, not to how bad the worst one was: it is robust to the size of a single outlier, and for the same reason blind to it.

          In brief

          • Step lines: 5% quantile of the 90 days before \(t\), the 5th lowest deviation; thin lines: daily \(d_t\)
          • 11 March 2023: USDC closes at \(-285\) bp; once that day is in the window, the 5% quantile moves from \(-1.90\) to \(-2.15\) bp
          • 15 March: the four SVB days (\(-285\), \(-79\), \(-11\), \(-8\) bp) and 10 March (\(-5.2\) bp) are the five lowest, so the quantile is \(-5.21\) bp; it stays at its 2023 low, \(-10.5\) bp, from 22 March to 9 June
          • As the SVB days leave the window (10–13 June) it returns to \(-5.0\) bp: the size of the lowest day never enters the estimate, only its rank
          • Source: own computation on EODHD daily prices of the stablecoins, 2023
          BA_L6_rollq

          Quantile Lasso and the Selected Penalty

          Slide 15 in the PDF

          When a coin's lower tail may depend on many other coins, the regression has many candidate predictors and few tail observations. The Lasso (least absolute shrinkage and selection operator) of Tibshirani (1996) adds a penalty on the sum of the absolute slopes, the L1 norm \(\|\beta\|_1 = \sum_k |\beta_k|\); Belloni & Chernozhukov (2011) study it for quantile regression. For coin \(j\) and a window \(W_t\) of past days, the 5% Quantile Lasso is

          \[(\hat\alpha, \hat\beta)(\lambda) = \arg\min_{\alpha, \beta} \sum_{s \in W_t} \rho_{0.05}\big(d_{j,s} - \alpha - \mathbf x_{j,s}'\beta\big) + \lambda\,\|\beta\|_1\]
          (6.3)

          where \(d_{j,s}\) is the deviation of coin \(j\) on day \(s\) and \(\mathbf x_{j,s}\) the predictors. A larger penalty \(\lambda\) sets more slopes exactly to zero. All slopes are zero once \(\lambda\) reaches

          \[\lambda_{max} = \max_k \Big|\sum_{s} x_{k,s}\,\psi_\tau\big(d_{j,s} - \hat q_\tau\big)\Big|\]

          where \(\psi_\tau(u) = \tau - \mathbb{1}\{u < 0\}\) and \(\hat q_\tau\) is the plain 5% quantile of the target. This is a scaled covariance between each predictor and an indicator of the target being in its lower tail.

          The penalty is chosen by generalised approximate cross-validation (GACV) on a grid of values:

          \[\mathrm{GACV}(\lambda) = \frac{\sum_s \rho_\tau\big(d_{j,s} - \hat\alpha - \mathbf x_{j,s}'\hat\beta(\lambda)\big)}{n - \mathrm{df}(\lambda)}\]
          (6.4)

          The numerator is the check loss of the fit, and the denominator corrects it for the number of estimated parameters, the degrees of freedom \(\mathrm{df} = \#\{\hat\beta_k \neq 0\} + 1\), the number of non-zero slopes plus the intercept, so that a model with more predictors must fit clearly better to be chosen.

          The idea of the Financial Risk Meter (Mihoci, Althof, Chen & Härdle, 2020) is to use the selected penalty \(\hat\lambda_{j,t}\) as a proxy for risk: a higher penalty is chosen when the target's lower tail is tied more strongly to the predictors. Its level also depends on the scaling of the variables, the noise, the grid of \(\lambda\) and the selection criterion. It is a proxy that depends on the specification and must be checked against other measures of stress.

          In brief

          • With many candidate predictors, add an L1 penalty (Tibshirani (1996), Belloni & Chernozhukov (2011)):
            \[ (\hat\alpha, \hat\beta)(\lambda) = \arg\min_{\alpha, \beta} \sum_{s \in W_t} \rho_{0.05}\big(d_{j,s} - \alpha - \mathbf x_{j,s}'\beta\big) + \lambda\,\|\beta\|_1 \]
          • All slopes are zero for \(\lambda \geq \lambda_{max}\), with (\(\psi_\tau(u) = \tau - \mathbb 1\{u < 0\}\))
            \[ \lambda_{max} = \max_k \Big|\sum_{s} x_{k,s}\,\psi_\tau\big(d_{j,s} - \hat q_\tau\big)\Big| \]
            • a scaled covariance between each predictor and the target's lower-tail indicator
          • \(\lambda\) is chosen by generalised approximate cross-validation on a grid:
            \[ \mathrm{GACV}(\lambda) = \frac{\sum_s \rho_\tau\big(d_{j,s} - \hat\alpha - \mathbf x_{j,s}'\hat\beta(\lambda)\big)}{n - \mathrm{df}(\lambda)}, \qquad \mathrm{df} = \#\{\hat\beta_k \neq 0\} + 1 \]
          • FRM idea (Mihoci, Althof, Chen & Härdle (2020)): the selected \(\hat\lambda_{j,t}\) is used as a risk proxy
            • its level also depends on scaling, noise, the \(\lambda\) grid and the selection criterion: a specification-dependent proxy, to be checked against other stress measures

          Worked Example: the Quantile Lasso Equation for USDC after SVB

          Slide 16 in the PDF
            Question for the room The ▶ button lowers λ from λmax = 164.1 along the grid of 25 values. Does the number of active predictors grow at every step?
            Answer

            No. It goes 0, 1, 2, then back to 1 (at λ = 51.9 USDT leaves and only DAI is active), 2 up to λ = 5.2, then 3, 4, 5, 10, 12, 11, 12, 12 and 13. The Lasso path of a quantile regression is not monotone: a predictor can leave when others enter. At the selected λ̂ = 0.241 twelve of the 13 are active; the one left out is USDT, which is active at the neighbouring λ values.

            The example estimates one equation (6.3): the target is USDC, the window the 90 days up to 19 March 2023, and the 13 candidate predictors are the deviations of 9 other coins and the lagged changes of 4 macro variables, all scaled by their median absolute deviation (MAD), the median distance of the observations from their median. The left panel shows the Lasso path, the estimated slopes \(\hat\beta_k(\lambda)\) as the penalty grows from left to right on a log scale; the right panel shows the GACV (6.4) and the number of active predictors.

            The largest useful penalty is \(\lambda_{max} = 164.1\). GACV is lowest at \(\hat\lambda = 0.241\) (dotted red line), where 12 of the 13 predictors are active. The largest slopes belong to DAI (3.70), TUSD (−1.93) and PAXG (1.78). DAI enters the path first and stays until the largest values of \(\lambda\): in the SVB window, the lower tail of USDC was essentially the lower tail of DAI, the link already seen in Lecture 5.

            In brief

            • Target USDC, 90 days to 19 March 2023, 13 candidates (9 coins, 4 lagged macro changes), all scaled by the median absolute deviation (MAD)
            • \(\lambda_{max} = 164.1\); GACV minimum at \(\hat\lambda = 0.241\) with 12 active predictors; largest slopes: DAI \(3.70\), TUSD \(-1.93\), PAXG \(1.78\)
            • DAI keeps the largest slope along the path and is the only active predictor at \(\lambda \approx 52\) (USDT enters first, alone, at \(\lambda \approx 112\); DAI at \(\lambda \approx 76\)): in the SVB window USDC's lower tail was DAI's lower tail
            • Source: own computation on EODHD daily prices of 10 stablecoins and four market series, 90 days to 19 March 2023
            BA_L6_lasso_path

            Section summary

            The check loss turns a quantile into a minimisation problem, and the Lasso adds a penalty that selects predictors. A 5% quantile in a short window rests on very few days, so rolling tail models are noisy, and the selected penalty is a risk proxy whose level depends on the specification.

            The De-pegging Pressure Index

            The section builds the DPI from many Quantile Lasso equations, describes the index from 2020 to 2026 and its trend, computes it by hand for one day, splits it into the contributions of the coins, and measures how it responded to six market events, day by day and in one number, with a check of how much those responses depend on technical choices.

            For each coin \(j\) and each day \(t\), a 5% Quantile Lasso (6.3) is estimated on the preceding 90 days. The target is the coin's deviation \(d_{j,s}\). The predictors are the other coins' deviations on the same day and the one-day-lagged changes of four macro variables: the US dollar index (DXY), the CBOE volatility index (VIX), the Standard & Poor's 500 stock index (S&P 500) and the 1-year Treasury yield; the paper also uses a crypto volatility index (CVIX). Each variable is centred by its median in the window and divided by \(1.4826 \times \mathrm{MAD}\), which equals the standard deviation for Normal data. The penalty \(\hat\lambda_{j,t}\) is the one of 25 grid values that minimises GACV (6.4). The window uses only prices known on day \(t\), as first recorded, so the index contains no look-ahead: it could have been computed on that day.

            The index is the average of the selected penalties over the \(J\) coins:

            \[\mathrm{DPI}_t = \frac{1}{J}\sum_{j=1}^{J} \hat\lambda_{j,t}\]
            (6.5)

            Two estimates of the index appear in this lecture. The paper uses 11 coins with CoinGecko prices and 2,252 days up to 31 May 2026. The estimate on EODHD prices uses 10 coins, because IDRT has no EODHD prices after 2023, and 2,397 days up to 23 October 2026, so 23,970 equations.

            The DPI is not a probability of a depeg and not a percentage. It is a model-implied measure of how strongly the coins depend on each other in their lower tails.

            In brief

            • For each coin \(j\) and day \(t\), a 5% Quantile Lasso on the preceding 90 days
              • target: \(d_{j,s}\); predictors: the other coins' \(d_{k,s}\) (same day) and one-day-lagged changes of the US dollar index (DXY), the CBOE volatility index (VIX), S&P 500, 1-year Treasury yield (the working paper also uses the CVIX)
              • every variable centred by its window median and divided by \(1.4826 \times \mathrm{MAD}\)
              • 25 values of \(\lambda\); \(\hat\lambda_{j,t}\) minimises GACV
              • real time: the window uses only prices known on day \(t\), as first recorded, so the index contains no look-ahead
            • Index:
              \[ \mathrm{DPI}_t = \frac{1}{J}\sum_{j=1}^{J} \hat\lambda_{j,t} \]
            • Two estimates of the index appear in this lecture
              • working paper: 11 coins, CoinGecko prices, 2 252 days to 31 May 2026
              • EODHD prices: 10 coins (IDRT has no EODHD prices after 2023), 2 397 days to 23 October 2026, 23 970 equations
            • The DPI is not a depeg probability or a percentage; it is a model-implied measure of lower-tail dependence

            The DPI, April 2020 – October 2026

            Slide 18 in the PDF
              Question for the room With the trend sample to 31 May 2026 the slope of ln DPI is +0.046 a year (p = 0.21). The ▶ button moves the end of the trend sample. With the sample ending on 31 December 2025, is a zero slope rejected at 5%?
              Answer

              Yes: +0.088 a year, HAC(90) standard error 0.032, p = 0.01, 95% interval [0.026, 0.151]. Adding the first five months of 2026, when the DPI was lower (annual median 0.32 against 0.52 in 2025), brings the slope to +0.046 and the p-value to 0.21; to 23 October 2026 it is +0.035 (p = 0.25). The test result depends on where the sample ends, which is one more reason to read the DPI in regimes, not as a trend.

              The chart shows the daily DPI (6.5) on a log scale. Each point is coloured by its quintile over the whole sample, from low (quintile 1) through general, elevated and high to severe (quintile 5). The dotted lines mark six events, the dashed red line the end of the paper's sample.

              The index moves in regimes rather than as day-by-day noise around a constant level. Its annual medians were 0.27 in 2020, between 0.39 and 0.40 in 2021–2023, 0.44 in 2024, 0.52 in 2025 and 0.32 in 2026. The highest regime came in late 2025. Most of it comes from one thinly traded token, BUSD, which carries 63% of the index in October–December 2025, and not from the sUSD depeg that began in the same months (see Contributions of Individual Coins to the DPI below).

              A log-linear trend fitted up to May 2026 gives \(+0.046\) per year, with a standard error (SE) of 0.036 that is consistent under heteroskedasticity and autocorrelation over 90 days (HAC(90)), a p-value of 0.21 and a 95% confidence interval (CI) of −0.025 to 0.117. A zero slope is not rejected. The interval is wide, from a small fall to a rise of 0.117 per year, so this is not evidence that there is no trend either. On CoinGecko prices to 31 May 2026, the paper estimates −0.004 per year with p = 0.78: neither price source rejects a zero trend, and the two estimates have opposite signs.

              In brief

              • Annual medians: 0.27 (2020), 0.39–0.40 (2021–2023), 0.44 (2024), 0.52 (2025), 0.32 (2026); highest regime in late 2025, carried mostly by BUSD (63% of the index in October–December 2025), not by the sUSD depeg of the same months
              • Log-linear trend to May 2026: \(+0.046\) per year, heteroskedasticity and autocorrelation consistent (HAC(90)) standard error (SE) 0.036, p \(= 0.21\), 95% CI \([-0.025, 0.117]\)
              • A zero slope is not rejected; the wide interval is not evidence of no trend either
              • Working paper (CoinGecko, to 31 May 2026): \(-0.004\) per year, p \(= 0.78\): also no rejection, with the opposite sign
              • Source: own computation on EODHD daily prices of 10 stablecoins and four market series, to 23 October 2026
              BA_L6_dpi

              Worked Example: The DPI on 12 March 2023

              Slide 19 in the PDF

              The index (6.5) is a simple average, and one day shows what goes into it. On 12 March 2023, the day after the USDC low, each of the ten Quantile Lasso equations, estimated on the 90 days to 11 March, selected a penalty \(\hat\lambda_{j,t}\): 1.096 for USDT, 0.902 for USDC, 0.861 for DAI, 0.110 for TUSD, 0.586 for USDP, 1.369 for GUSD, 0.045 for BUSD, 0.072 for sUSD, 0.050 for PAXG and 0.144 for EURS. Their sum is 5.235, so the DPI is \(5.235/10 = 0.52\).

              To place this value, compare it with the quintile bounds of the daily index from April 2020 to October 2026: 0.275, 0.350, 0.427 and 0.550. A DPI of 0.52 lies in the fourth quintile, labelled "high", as stated in the comparison of the three objects at the start of the lecture.

              The ten numbers are very unequal. The three largest, GUSD, USDT and USDC, carry 64% of the sum, and the median penalty is only 0.36. Leaving out GUSD alone gives \((5.235 - 1.369)/9 = 0.430\), just above the lower bound of the fourth quintile (0.427). The DPI is a mean of skewed numbers, and one coin can move it from the top of a quintile to its bottom. The next frame asks which coins drive the index over the whole sample.

              In brief

              • Selected penalties \(\hat\lambda_{j,t}\), each from the 90 days to 11 March 2023:
              USDTUSDCDAITUSDUSDPGUSDBUSDsUSDPAXGEURS
              1.0960.9020.8610.1100.5861.3690.0450.0720.0500.144
              • Sum 5.235, so \(\mathrm{DPI} = 5.235/10 = 0.52\)
              • Quintile bounds of the DPI, April 2020 – October 2026: 0.275, 0.350, 0.427, 0.550, so 0.52 lies in the fourth quintile (“high”)
              • Three coins (GUSD, USDT, USDC) carry 64% of the sum; the median penalty is only 0.36
              • Without GUSD: \((5.235 - 1.369)/9 = 0.430\), just above the lower bound of the fourth quintile (0.427)
              • The DPI is a mean of skewed numbers: one coin moves it from the top of a quintile to its bottom
              • Source: own computation on EODHD daily prices, 90 days to 11 March 2023

              Contributions of Individual Coins to the DPI

              Slide 20 in the PDF
                Question for the room Leaving out BUSD lowers the 2025 median of the DPI from 0.52 to 0.34. The ▶ button leaves out each coin in turn. Leaving out which other coin lowers the 2025 median the most?
                Answer

                None by more than 0.02: without EURS the 2025 median is 0.50; without PAXG 0.52, without USDC 0.52; leaving out any of the other six raises it, to between 0.53 (USDP) and 0.56 (TUSD), because the weight of BUSD then rises from 1/10 to 1/9. Leaving out sUSD gives 0.55: the late-2025 peak does not come from the sUSD depeg. Leaving out BUSD changes the whole index most (correlation of the logs 0.82; 0.96 to 0.98 for the other coins).

                Because the DPI is an average, it can be split into the contributions \(\hat\lambda_{j,t}/J\) of the coins. The left panel shows, for each year, the share of each coin in the yearly mean of the index; the right panel compares the annual median of the DPI with all ten coins and without BUSD.

                In most years the coins contribute about equally, around 10% each, as an average of comparable equations should. The exception is 2025, when BUSD alone carries 41.4% of the index. Paxos stopped minting BUSD in February 2023, and the token has since traded thinly: the standard deviation of its daily deviation was 5 bp in 2023 and 84 bp in 2025. Noisy prices give its equation a very large selected penalty, and through the average that penalty lifts the whole index.

                Without BUSD the picture of 2025 changes. Its annual median falls from 0.52 to 0.34, and in October–December 2025 from 0.76 to 0.31, because BUSD accounts for 63% of the index in that quarter. The high regime of late 2025 seen in the DPI chart is therefore mostly the signal of one thinly traded coin, not a system-wide rise in tail co-movement.

                The lesson is about how to read any average index. Before interpreting a regime, recompute the index leaving out one coin at a time; a regime that survives is a property of the system, one that disappears belongs to a single component.

                In brief

                • Left: share of each coin in the yearly mean of the DPI, from the terms \(\hat\lambda_{j,t}/J\); most coins carry about 10% in most years
                • 2025 is the exception: BUSD carries 41.4%; Paxos stopped minting BUSD in February 2023, and its daily deviation has a standard deviation of 5 bp in 2023 and 84 bp in 2025
                • Right: without BUSD the 2025 median DPI is 0.34 instead of 0.52; October–December 2025: 0.31 instead of 0.76 (BUSD 63% of the index)
                • The peak of late 2025 comes from one thinly traded coin: check leave-one-coin-out indices before reading a regime
                • Source: own computation on EODHD daily prices of 10 stablecoins, April 2020 – 23 October 2026
                BA_L6_dpi_coins

                Event Responses

                Slide 21 in the PDF

                The response of the DPI to an event is measured by comparing its average after the event with its average before:

                \[\Delta = \frac{\overline{\mathrm{DPI}}_{[0, 30]}}{\overline{\mathrm{DPI}}_{[-60, -8]}} - 1\]
                (6.6)

                where \(\overline{\mathrm{DPI}}_{[a, b]}\) is the mean of the index from day \(a\) to day \(b\) relative to the event. The test is run on \(\ln \mathrm{DPI}\) with HAC standard errors, and the p-values are adjusted with the Holm correction for testing six events at once.

                The table gives the response for six events: the KuCoin exchange hack (26 September 2020), the China sell-off (19 May 2021), the Terra collapse (9 May 2022), the FTX crisis (8 November 2022), the SVB / USDC depeg (11 March 2023) and the Iran–Hormuz shock (28 February 2026). In the paper, the only significant rises are +19.5% after FTX and +11.0% after SVB, the two events about intermediaries and confidence in reserves; the Terra response, +2.2%, is not significant (n.s.).

                On EODHD prices the picture changes. The responses are +141.1% after the KuCoin hack (Holm p below 0.001), −20.2% after the China sell-off (p = 0.033) and +17.9% after Terra (p = 0.021), while FTX (+13.4%, p = 0.961) and SVB (+21.0%, p = 0.092) are not significant after the correction. The sign agrees with the CoinGecko estimates for 4 of the 6 events, but the sizes and the significance change with the data source, and even with the grid of \(\lambda\): the last column, with a narrower grid, gives +158.6% for SVB.

                In brief

                • Response: \(\Delta = \overline{\mathrm{DPI}}_{[0, 30]} / \overline{\mathrm{DPI}}_{[-60, -8]} - 1\); test on \(\ln \mathrm{DPI}\) with HAC errors; Holm correction for six events
                EventdateCoinGecko (paper)EODHDHolm p, EODHDEODHD, narrow \(\lambda\) grid
                KuCoin hack26 Sep 2020\(-14.4\%\)\(+141.1\%\)\(<0.001\)\(+95.3\%\)
                China sell-off19 May 2021\(-8.6\%\)\(-20.2\%\)\(0.033\)\(-11.9\%\)
                Terra collapse9 May 2022\(+2.2\%\) (n.s.)\(+17.9\%\)\(0.021\)\(+15.4\%\)
                FTX crisis8 Nov 2022\(+19.5\%\)\(+13.4\%\)\(0.961\)\(+36.7\%\)
                SVB / USDC depeg11 Mar 2023\(+11.0\%\)\(+21.0\%\)\(0.092\)\(+158.6\%\)
                Iran–Hormuz shock28 Feb 2026\(-22.9\%\)\(+2.4\%\)\(0.961\)\(+1.9\%\)
                • Working paper: significant rises only after FTX and SVB, the two events about intermediation and reserve confidence
                • EODHD prices: same sign as the paper for 4 of 6 events; sizes and significance change with the data source and even with the \(\lambda\) grid
                • Source: CoinGecko column: Fan et al. (2026), working paper; EODHD columns: own computation on EODHD daily prices

                Event Windows, Day by Day

                Slide 22 in the PDF

                The response (6.6) compresses 84 days of the index into one number. The chart shows the days behind each number. In each panel the daily DPI is divided by its mean over days −60 to −8, so the pre-event mean is 1 (green line); the thin blue line is the daily value, the thick one its 7-day centred mean, and the red line the mean over days 0 to 30, whose height is \(1 + \Delta\).

                The six paths look quite different. After the KuCoin hack (+141%) the DPI was already rising in the days before the event and kept rising for three weeks, so the response mixes the event with a trend that started before it. After SVB (+21%) the index stayed above its pre-event mean on 24 of the 31 post-event days, the most consistent rise among the six. Around the China sell-off (−20%) and the Iran–Hormuz shock (+2%) there is no step at day 0; before Iran–Hormuz the DPI dipped about 20 days early and then recovered.

                A ratio of two window means cannot distinguish a jump at the event from a drift that started earlier, or a late spike from a lasting shift. Plotting the paths is a cheap check that should accompany every event response.

                In brief

                Event Windows, Day by Day
                • DPI divided by its mean over days \(-60\) to \(-8\) (green line at 1); red line: mean of days 0 to 30
                • KuCoin (\(+141\%\)): the DPI was rising before day 0 and kept rising for three weeks; SVB (\(+21\%\)): above the pre-event mean on 24 of 31 post-event days
                • China sell-off (\(-20\%\)) and Iran–Hormuz (\(+2\%\)): no step at day 0; before Iran–Hormuz the DPI dips about 20 days early and then recovers
                • A ratio of two window means compresses paths that differ in timing and shape
                • Source: own computation on EODHD daily prices of 10 stablecoins and four market series
                BA_L6_eventpaths

                Worked Example: The SVB Response by Hand

                Slide 23 in the PDF

                The SVB response can be computed directly from the daily index. The event day is \(t_0\) = 11 March 2023. The pre-event window, days −60 to −8, runs from 10 January to 3 March 2023, 53 days; the DPI values in it sum to 22.58, so their mean is \(22.58/53 = 0.426\). The post-event window, days 0 to 30, runs from 11 March to 10 April 2023, 31 days; the values sum to 15.98, so their mean is \(15.98/31 = 0.515\).

                The response (6.6) is \(\Delta = 0.515/0.426 - 1 = +21.0\%\). The test works with logarithms: the difference between the window means of \(\ln \mathrm{DPI}\) is 0.212, which corresponds to a ratio of geometric means of \(e^{0.212} = 1.24\), a rise of 24%, close to the 21% of the arithmetic means.

                The size depends on the summary statistic. With medians instead of means, the index goes from 0.386 to 0.514, a rise of 33%. Together with the narrow grid of penalties, which gives +159%, this shows how many choices sit behind a single event response.

                In brief

                • Event day \(t_0\) = 11 March 2023; daily DPI from EODHD prices
                • Pre-event window, days \(-60\) to \(-8\): 10 January – 3 March 2023, 53 days
                  \[ \overline{\mathrm{DPI}}_{[-60,-8]} = 22.58/53 = 0.426 \]
                • Post-event window, days 0 to 30: 11 March – 10 April 2023, 31 days
                  \[ \overline{\mathrm{DPI}}_{[0,30]} = 15.98/31 = 0.515 \]
                • Response: \(\Delta = 0.515/0.426 - 1 = +21.0\%\)
                • The test works on \(\ln \mathrm{DPI}\): difference of the window means of the logs \(0.212\), about the same here
                • With medians instead of means: \(0.386 \to 0.514\), \(+33\%\): the size depends on the summary statistic, as well as on the \(\lambda\) grid

                Robustness of Event Responses

                Slide 24 in the PDF
                  Question for the room With the post-event window of days 0..30, three own responses are significant after the Holm correction: KuCoin, the China sell-off and Terra. With a window of days 0..10, which events have a Holm p-value below 0.05?
                  Answer

                  KuCoin (+49.6%), Terra (+23.5%) and the Iran–Hormuz shock (−34.8%, Holm p = 0.049); the China sell-off is no longer significant (−12.5%, Holm p = 0.123). The window length changes which events respond, and even the sign: the Iran–Hormuz response is −35% over 10 days and +2.4% over 30. Neither window gives FTX and SVB, the two events the paper finds significant. The narrow-λ-grid index on this page is re-estimated with the notebook (+94.9% for KuCoin and +158.6% for SVB at 30 days; the lecture reports +95.3% and +158.6%).

                  The chart compares the event responses (6.6) in three versions: the paper (CoinGecko, 11 coins), the index on EODHD prices with a wide grid of penalties, from \(10^{-4}\lambda_{max}\) to \(\lambda_{max}\), and the same index with a narrow grid, from \(10^{-2}\lambda_{max}\) to \(\lambda_{max}\). The vertical axis is linear near zero and logarithmic beyond ±10%, so large responses look shorter than they are.

                  The grid matters because of a boundary problem. With the narrow grid, 19.6% of equations select the smallest \(\lambda\) on the grid, a boundary solution that would probably have gone lower on a wider grid; with the wide grid only 4.0% do. The two indices correlate 0.84 in logs, so they tell the same story over time, yet the SVB response is +21% in one and +159% in the other. A 31-day mean of a noisy index is sensitive to tuning choices. Event responses should be reported together with their specification, and only the results that survive several specifications should be trusted.

                  In brief

                  • Narrow grid \([10^{-2}, 1]\lambda_{max}\): 19.6% of equations select the smallest \(\lambda\) (a boundary solution); wide grid \([10^{-4}, 1]\lambda_{max}\): 4.0%
                  • The two indices correlate 0.84 in logs, yet the SVB response is \(+21\%\) or \(+159\%\): a 31-day mean of a noisy index is sensitive to tuning choices
                  • Lesson: report event responses with their specification; trust the results that survive several specifications
                  BA_L6_events

                  Section summary

                  The DPI averages the penalties selected by GACV in daily Quantile Lasso equations. It moves in regimes, a zero trend is not rejected, and its responses to events change with the data source and the penalty grid, so a single event response is fragile evidence.

                  Tail Dependence between Stablecoins

                  The DPI is an average. The section opens it up: which coins enter each other's equations, how the models change under pressure, whether the DPI is just another stress gauge, whether a good fit means a good forecast, and whether the concentration of the market explains the pressure.

                  Tail Links between Stablecoins: Signed Selection Frequencies

                  Slide 25 in the PDF

                  The heat map shows, for each pair of a target coin (rows) and a predictor coin (columns), how often and with which sign the predictor enters the target's daily equation. A positive slope means the coins are aligned: a lower deviation of the predictor goes with a lower 5% quantile of the target. Each cell is a signed selection frequency: the percentage of daily equations in which the slope is positive minus the percentage in which it is negative, in percentage points (pp). It is a frequency, not a size in bp. Red cells are mostly aligned links, blue cells mostly inverse ones.

                  Across all equations, 47.4% of the coin coefficients are aligned and 33.1% inverse; the paper finds 39.6% and 23.9%. The most frequently aligned links are USDC predicted by DAI (75 pp), TUSD by BUSD (64 pp) and DAI by USDC (60 pp); the most frequently inverse link is BUSD predicted by GUSD (−32 pp). The macro variables account for 18.4% of the total absolute slope weight, against 20.8% in the paper.

                  The paper adds a caution: no difference between pairs survives a correction for multiple testing, and the rankings of links change between the two halves of the sample. The coins share a common structure in their lower tails, but there is no fixed sender of pressure.

                  In brief

                  • Positive slope = aligned: a lower predictor deviation goes with a lower target quantile
                  • Cell = signed selection frequency: % of daily equations with a positive slope minus % with a negative slope, in percentage points (pp); not a size in bp
                  • Aligned 47.4% vs inverse 33.1% of coin coefficients (working paper: 39.6% vs 23.9%)
                  • Most often aligned: USDC \(\leftarrow\) DAI (75 pp), TUSD \(\leftarrow\) BUSD (64), DAI \(\leftarrow\) USDC (60); most often inverse: BUSD \(\leftarrow\) GUSD (\(-32\))
                  • Macro variables: 18.4% of absolute slope weight (working paper: 20.8%)
                  • Working paper: no pair difference survives multiple testing; rankings change between halves: common structure, no fixed sender
                  • Source: own computation on EODHD daily prices; working-paper values from Fan et al. (2026)
                  BA_L6_links

                  Model Sparsity across DPI Quintiles

                  Slide 26 in the PDF
                    Question for the room With quintiles of the DPI the mean active set falls from 10.96 in the lowest group to 9.68 in the highest. The ▶ button raises the number of groups from 2 to 10. With deciles, what happens to the gap between the lowest and the highest group?
                    Answer

                    It widens: 11.06 active predictors in the lowest decile against 9.51 in the highest, and the mean falls from each decile to the next. The Spearman correlation of the daily DPI with the mean active set is −0.73. Part of this is mechanical: the DPI is the mean selected λ, and a larger λ removes predictors by construction.

                    The chart gives the mean number of active predictors, out of 13, in each quintile of the DPI. On EODHD prices the models become sparser as pressure rises, from 10.96 active predictors in the lowest quintile to 9.68 in the highest; the paper finds 10.11 to 9.00, out of 15. The direction is the same with both price sources, so the result does not depend on the data.

                    The result is partly mechanical. A larger selected penalty removes predictors by construction, and the DPI is the average penalty, so higher DPI and fewer predictors go together whatever the economics. An economic reading, that a few strong tail links dominate under pressure, is a hypothesis that this chart does not establish.

                    In brief

                    • EODHD prices: from 10.96 active predictors (of 13) in the lowest DPI quintile to 9.68 in the highest; working paper: 10.11 to 9.00 (of 15)
                    • Partly mechanical: a larger selected penalty removes predictors by construction
                    • Economic reading (a few strong tail links dominate under pressure): a hypothesis, not shown by this chart
                    • Same direction with both price sources: a result robust to the data
                    • Source: own computation on EODHD daily prices; working-paper values from Fan et al. (2026)
                    BA_L6_active

                    The DPI and Simple Stress Gauges

                    Slide 27 in the PDF

                    If the DPI only repeated what simple gauges already show, the Quantile Lasso machinery would not be needed. The chart compares it with two such gauges, each averaged over the same window of 90 days before the day as the DPI: the cross-coin median absolute deviation \(|d_{i,t}|\), which measures how far the stablecoins are from their pegs, and the VIX, which measures fear in the US equity market. Each point is one of the 2,397 days from April 2020 to October 2026, coloured by year.

                    Neither gauge explains the DPI. The Spearman rank correlation of the DPI with the median deviation is −0.10, and with the VIX −0.16; both are close to zero and slightly negative. The year 2020 shows why: it has the largest deviations and the highest VIX of the sample, but a low DPI, with an annual median of 0.27.

                    The DPI and the gauges measure different things. The gauges measure levels: how far prices are from their pegs, how volatile equities are. The DPI measures dependence: how strongly the lower tails of the coins move together. Coins can be far from their pegs, each for its own reasons, while their tails are unrelated. The DPI therefore complements simple gauges; it does not replace them.

                    In brief

                    The DPI and Simple Stress Gauges
                    • Each point is one day, April 2020 – October 2026 (2 397 days); both gauges are means over the 90 days before the day, as the DPI window
                    • Left: cross-coin median \(|d_{i,t}|\); Spearman rank correlation with the DPI \(-0.10\)
                    • Right: VIX; Spearman rank correlation \(-0.16\)
                    • 2020 has the largest deviations and the highest VIX, but a low DPI (annual median 0.27)
                    • The DPI measures how coins move together in their lower tails, not how far they are from the peg: it complements simple gauges, it does not replace them
                    • Source: own computation on EODHD daily prices and the VIX, April 2020 – 23 October 2026
                    BA_L6_stress

                    In-Sample Fit and Out-of-Sample Forecasts of the DPI

                    Slide 28 in the PDF

                    The table, from the paper's sample to 31 May 2026, compares three models: macro predictors only, coin predictors only, and both (joint). Skill is defined as \(1 - \) (loss of the model / loss of an intercept-only model), so a positive skill means the model beats a constant quantile and a negative skill that it does worse. The first column is the in-window skill, computed from GACV inside each 90-day estimation window, which is not a loss on held-out data. The second is the out-of-sample (OOS) skill of a nowcast, with the check loss on days outside the estimation window and coin predictors taken from the target's own day. The third is the OOS skill of a forecast, with every predictor dated \(t-1\); the benchmark is a rolling intercept-only model.

                    The ranking reverses between the columns. The joint model gives the best in-window description, a skill of 37.0%, and the worst forecast, −38.6%, once all predictors are dated the day before. Even the nowcasts have negative skill, between −12.5% and −19.8%.

                    Three further results of the paper point the same way. The DPI alone has a negative out-of-sample \(R^2\) at horizons of 10 to 150 days. The crypto FRM forecasts crypto volatility with \(R^2\) between 0.17 and 0.29, but adding the DPI does not improve the forecast. On-chain variables from Curve and Uniswap enter 91% of equations, yet they do not reduce the held-out loss significantly: being selected often is not the same as having forecast value. Finally, the DPI and the crypto FRM correlate only 0.11: stablecoin pressure is not the same as crypto stress.

                    In brief

                    Modelin-window GACV skillsame-day OOS nowcast skillfully lagged OOS forecast skill
                    macro only\(8.3\%\)\(-12.5\%\)\(-12.5\%\)
                    coins only\(33.7\%\)\(-15.9\%\)\(-34.4\%\)
                    joint\(37.0\%\)\(-19.8\%\)\(-38.6\%\)
                    • Skill \(= 1 - \) loss of the model / loss of the intercept-only model, on the paper's sample (to 31 May 2026)
                      • in-window: GACV inside each 90-day estimation window, not a held-out loss
                      • OOS: check loss on days outside the estimation window; nowcast: coin predictors from the target's own day; forecast: every predictor dated \(t-1\); benchmark: rolling intercept-only model
                    • The best in-window description (joint, 37%) is the worst forecast once predictors are dated \(t-1\) (\(-38.6\%\))
                    • The DPI alone has negative out-of-sample (OOS) \(R^2\) at horizons of 10 to 150 days; the crypto FRM forecasts crypto volatility (\(R^2\) 0.17–0.29) but adding the DPI does not help
                    • On-chain variables (Curve, Uniswap) enter 91% of equations, yet do not reduce held-out loss significantly: frequent selection is not forecast value
                    • The DPI and crypto FRM correlate only 0.11: stablecoin pressure is not crypto stress
                    • Source: Fan et al. (2026), working paper, sample to 31 May 2026

                    Market Concentration and the DPI

                    Slide 29 in the PDF

                    Concentration of the stablecoin market is measured with the Herfindahl–Hirschman index (HHI), \(\mathrm{HHI} = \sum_j s_j^2\), where \(s_j\) is coin \(j\)'s share of the market value of the paper's \(J = 11\) coins. The paper normalises it to run from 0, equal shares, to 1, one coin:

                    \[\mathrm{HHI}^{\ast} = \frac{\mathrm{HHI} - 1/J}{1 - 1/J}\]

                    It was 0.640 in 2020, 0.291 in 2022 and 0.513 in 2026; on 31 May 2026 USDT held 69.4% and USDC 28.0% of the 11 coins. These values are not comparable with the raw HHI of Lecture 1 or with the shares of all USD stablecoins in Lecture 5, which use another set of coins and another date.

                    The paper regresses \(\ln \mathrm{DPI}_t\) on lagged concentration, control variables and year effects, with HAC(90) standard errors. An increase of one standard deviation (SD) in the HHI goes with a change of −1.52% in the DPI, with a 95% interval of −4.24% to 1.27%; for the share of the two largest coins (Top-2 share) the change is −2.01%, interval −4.93% to 1.01%, and the Holm-adjusted p-value is 0.38. Without year effects, higher concentration goes with a higher DPI; with them the link disappears, because both variables trend with the market regime.

                    Concentration and pressure answer different questions. Concentration tells how much of the system depends on two issuers, which is useful for scenarios and exposure limits (Lecture 1 introduced the HHI). It does not tell when pressure builds.

                    In brief

                    • Normalised Herfindahl–Hirschman index (HHI) of the paper's \(J = 11\) coins: \((\mathrm{HHI} - 1/J)/(1 - 1/J)\), \(\mathrm{HHI} = \sum_j s_j^2\), \(s_j\) = market-value share within the 11 coins
                      • 0.640 (2020), 0.291 (2022), 0.513 (2026); on 31 May 2026 USDT 69.4%, USDC 28.0% of the 11 coins
                      • not comparable with the raw HHI of Lecture 1 or with shares of all USD stablecoins (Lecture 5): other universe and date
                    • Regression of \(\ln \mathrm{DPI}_t\) on lagged concentration, controls and year effects (HAC(90)):
                      • \(+1\) standard deviation (SD) of HHI: \(-1.52\%\) DPI [\(-4.24\%\), \(1.27\%\)]; Top-2 share: \(-2.01\%\) [\(-4.93\%\), \(1.01\%\)]; Holm p \(= 0.38\)
                    • Without year effects, higher concentration goes with higher DPI; with them the link disappears: both trend with the market regime
                    • Interpretation
                      • concentration tells you how much of the system depends on two issuers: use it for scenarios and limits (Lecture 1: HHI)
                      • it does not tell you when pressure builds
                    • Source: Fan et al. (2026), working paper, sample to 31 May 2026

                    Section summary

                    Inside the DPI, aligned tail links outnumber inverse ones and the models become sparser under pressure, both results that hold with either price source. But a good in-window fit is not a forecast: held-out skill is negative and gets worse as more information is added, and concentration measures exposure rather than timing.

                    Summary

                    Uses of the DPI: Measurement, Forecasting and Event Attribution

                    Slide 30 in the PDF

                    The table collects the verdicts of the lecture by task. To measure the current tail structure, the evidence is positive: aligned links outnumber inverse ones, the index moves in regimes, and a zero trend is not rejected; the DPI is useful as one layer of monitoring. To forecast, the evidence is negative: out-of-sample skill is below zero, lagged models are worse, and the DPI adds nothing to the FRM. For event attribution, the paper finds rises after FTX and SVB, but the sizes change with the data source and the \(\lambda\) grid, so the results are fragile and must be reported with their specification. A measure of the present state is not a leading indicator, and each of the two claims needs its own test.

                    In brief

                    TaskEvidenceVerdict
                    Measure the current tail structurealigned \(>\) inverse links; regimes, zero trend not rejectedDPI useful, as one layer of monitoring
                    Forecastnegative OOS skill; lagged models worse; DPI adds nothing to FRMno
                    Event attributionFTX and SVB in the paper; sizes change with source and \(\lambda\) gridfragile: report the specification
                    • A state measure is not a leading indicator: test each claim separately

                    Reference adjustment (6.1) puts dollar, euro and gold tokens on one scale, and peg quality splits into two tiers: a median absolute deviation of 1.3 bp for USDC against 51 bp for EURS and 56 bp for sUSD. A 5% quantile in a 90-day window is the 5th lowest observation, so rolling tail models are noisy by construction.

                    The DPI (6.5) averages the Quantile Lasso penalties selected by GACV. It moves in regimes, and a zero trend is not rejected: \(+0.046\) a year, with a 95% CI of −0.025 to 0.117. A good fit is not a forecast: held-out skill is negative, and worse when more information is used.

                    In brief

                    • Reference adjustment puts dollar, euro and gold tokens on one scale; peg quality splits into two tiers (USDC 1.3 bp median vs EURS 51 bp, sUSD 56 bp)
                    • A 5% quantile in a 90-day window is the 5th lowest observation: rolling tail models are noisy by construction
                    • The DPI averages GACV-selected Quantile Lasso penalties; it moves in regimes, and a zero trend is not rejected (\(+0.046\) a year, 95% CI \([-0.025, 0.117]\))
                    • Good fit is not forecast: held-out skill is negative, and worse with more information

                    Interpretation Exercises

                    Slide 32 in the PDF

                    The four exercises use the numbers of the lecture: the reference adjustment of a gold token, the flat check loss in a short window, the sensitivity of an event response, and a regression coefficient whose interval contains zero. Each asks for an interpretation in a few sentences. Lecture 7 turns to networks, sentiment and asset classes, where a "link" becomes a correlation or a share of variance rather than a selection frequency.

                    In brief

                    1. PAXG trades at 3 998 USD while gold is 3 994 USD. Compute nothing: explain why its dollar return is not a depeg measure and what the reference adjustment changes.
                    2. In a 20-day window the 5% check loss is flat between the two lowest observations. What does this imply for a rolling 5% quantile, and what changes in a 90-day window?
                    3. The DPI rose by 159% after SVB in one specification and by 21% in another. Which results of the lecture would you still report with confidence, and why?
                    4. The HHI coefficient is \(-1.52\%\) with interval \([-4.24\%, 1.27\%]\) after year effects. What is the conclusion, and what is concentration still useful for?
                    • Next: Lecture 7, networks, sentiment and asset classes; there a “link” becomes a correlation or a variance share, not a selection frequency

                    Appendix: Further Material

                    The Appendix holds optional material for self-study. It shows how bad prints in daily prices are detected and how much they matter, decomposes the in-window fit into the contributions of coins and macro variables, and asks whether the DPI is useful for building a stablecoin portfolio.

                    Appendix: Further Material

                    Slide 33 in the PDF
                    Optional material for self-study.
                    • Bad Prints in Daily Prices
                    • Shapley Decomposition: Coin and Macro Variables
                    • Four Long-Only Rules
                    • Results: Covariance-Based and Penalty-Based Portfolio Rules
                    • Value and Drawdown, 2020–2026
                    • What the Minimum-Variance Portfolio Holds
                    • Worked Example: Penalty Coefficients as Portfolio Weights

                    Bad Prints in Daily Prices

                    Slide 34 in the PDF

                    A bad print is a recorded closing price far from where the coin actually traded, which reverses the next day. Daily price files of small tokens contain such prints, and a single one can dominate any statistic built on returns. The rule of Lecture 1 flags a daily move beyond 1,000 bp that reverses by at least 90% the next day, and replaces the close by the previous one. The rule needs the next day, so it can clean a descriptive table after the fact, but an index or a portfolio weight computed on day \(t\) sees day \(t-1\) only as first recorded.

                    In the ten stablecoins from 2020 to 2026 the rule flags nine prints; the table lists the five largest. The largest is a log return of −12,446 bp for sUSD on 6 August 2025. A reversal alone does not prove an error: that day sUSD traded as low as 0.21 USD, closed at 0.27 and closed at 0.98 the next day, on low volume, which is also what a brief real collapse would look like. Only a second, independent price source can decide.

                    How much the prints matter depends on the statistic. The DPI, built on 5% quantiles, is little affected: without any cleaning it changes on 437 days, but its trend moves only from +0.046 to +0.041 a year (p = 0.24), and the KuCoin response from +141% to +153%. Measures built on returns are not: with uncleaned prices the maximum drawdown of the linear-lambda portfolio rises from 19.0% to 29.2%, and the volatility of the equal-weight portfolio from 3.9% to 6.2%.

                    In brief

                    • A bad print is a recorded close far from where the coin traded, reversed the next day
                    • Detection rule (Lecture 1): a daily move beyond 1 000 bp that reverses by at least 90% the next day; the close is replaced by the previous close
                    • The rule needs the next day: it can clean a descriptive table, but an index or a portfolio weight on day \(t\) sees day \(t-1\) only as first recorded
                    • Ten stablecoins, 2020 – 2026: nine flagged prints, the five largest below
                    coindateclose (USD)log return (bp)
                    sUSD6 Aug 20250.2743\(-12\,446\)
                    EURS15 Jun 20250.8788\(-2\,720\)
                    sUSD6 Oct 20260.0724\(+2\,633\)
                    EURS23 Jun 20250.9362\(-1\,953\)
                    sUSD14 Feb 20200.8372\(-1\,566\)
                    • A reversal alone does not prove an error: sUSD on 6 August 2025 traded as low as 0.21 USD, closed at 0.27 and closed at 0.98 the next day, on low volume; only a second price source can decide
                    • The DPI, built on 5% quantiles, is little affected: without any cleaning it changes on 437 days, the trend moves from \(+0.046\) to \(+0.041\) a year (p \(= 0.24\)), the KuCoin response from \(+141\%\) to \(+153\%\)
                    • Return-based measures are not: uncleaned prices raise the linear-lambda maximum drawdown from 19.0% to 29.2% and equal-weight volatility from 3.9% to 6.2%
                    • Source: own computation on EODHD daily prices of 10 stablecoins, 2020 – 23 October 2026

                    Shapley Decomposition: Coin and Macro Variables

                    Slide 35 in the PDF

                    The conditional skill of a model \(m\) is \(S_m = 1 - \overline{\mathrm{GACV}}_m / \overline{\mathrm{GACV}}_0\), where model 0 has only an intercept. With two blocks of predictors, coins and macro variables, the Shapley value of the coin block averages its contribution over the two orders in which the blocks can be added, \(\phi_{coin} = \tfrac12[(S_{coin} - S_0) + (S_{joint} - S_{macro})]\), and the two Shapley values add up to the joint skill. On EODHD prices, with the equations estimated on every 7th day, the skills are 6.1% for macro only, 38.5% for coins only and 41.6% for the joint model, so \(\phi_{coin} = \tfrac12[(38.5 - 0) + (41.6 - 6.1)] = 37.0\), and coins account for \(37.0/41.6 \approx 89\%\) of the joint skill; the paper finds 84.3%. This in-window result is robust to the price source: the other stablecoins explain most of a coin's lower tail, and macro variables add little.

                    In brief

                    • Conditional skill of a model \(m\): \(S_m = 1 - \overline{\mathrm{GACV}}_m / \overline{\mathrm{GACV}}_0\) (0 = intercept only)
                    • Two-block Shapley value of the coin block:
                      \[ \phi_{coin} = \tfrac12\big[(S_{coin} - S_0) + (S_{joint} - S_{macro})\big], \qquad \phi_{coin} + \phi_{macro} = S_{joint} \]
                    working paper (CoinGecko)EODHD prices, every 7th day
                    ModelGACVskillGACVskill
                    intercept only\(0.1808\)–\(0.1679\)–
                    macro only\(0.1658\)\(8.3\%\)\(0.1577\)\(6.1\%\)
                    coins only\(0.1199\)\(33.7\%\)\(0.1033\)\(38.5\%\)
                    joint\(0.1139\)\(37.0\%\)\(0.0980\)\(41.6\%\)
                    coin share of joint skill \(\phi_{coin}/S_{joint}\)\(84.3\%\)\(88.9\%\)
                    • Worked example, EODHD columns: \(\phi_{coin} = \frac12[(38.5 - 0) + (41.6 - 6.1)] = 37.0\); \(\;37.0/41.6 \approx 0.89\)
                    • In-window (GACV) result, robust to the price source: the other stablecoins explain most of a coin's lower tail; macro variables add little

                    Four Long-Only Rules

                    Slide 36 in the PDF

                    The portfolio question concerns a holder of several stablecoins, such as a treasury, a market maker or a lending desk, measured by reference-adjusted returns \(r_{i,t}\) in bp. Four rules are rebalanced every 7 days with information up to \(t-1\), with no short positions and every weight capped at 30%. Equal weight gives each coin \(1/J\). Global minimum variance (GMV) minimises the portfolio variance \(w'\hat\Sigma w\), where \(\hat\Sigma = 0.9\,S + 0.1\,\mathrm{diag}(S)\) shrinks the 90-day covariance matrix \(S\) by 10% towards its diagonal. Linear lambda minimises \(\sum_i w_i \hat\lambda_{i,t-1}\), filling the coins with the lowest penalties first; quadratic lambda, used only in the paper, combines the penalty level with the covariance. Between rebalances the weights drift with the daily simple returns; turnover is half the sum of the absolute weight changes at a rebalance; there are no transaction costs. The outcomes are the annualised mean return and volatility, the 5% expected shortfall (ES 5%) of daily returns, the mean loss on the worst 5% of days, the maximum drawdown and the turnover.

                    In brief

                    • Holding reference-adjusted stablecoin positions (a treasury, a market maker, a lending desk); returns \(r_{i,t}\) in bp
                    • Rules, rebalanced every 7 days with information to \(t-1\), weights capped at 30%:
                      • equal weight: \(w_i = 1/J\)
                      • global minimum variance (GMV): \(\min_w w'\hat\Sigma w\), \(\;\sum w_i = 1\), \(0 \leq w_i \leq 0.3\); \(\hat\Sigma = 0.9\,S + 0.1\,\mathrm{diag}(S)\), \(S\) the covariance of the last 90 days (10% shrinkage)
                      • linear lambda: \(\min_w \sum_i w_i \hat\lambda_{i,t-1}\) under the same constraints (fill the lowest-penalty coins first)
                      • quadratic lambda (only in the working paper): penalty level plus covariance
                    • Holdings simulation: weights drift with daily simple returns \(e^{r_{i,t}/10^4} - 1\) between rebalances; turnover \(= \frac12\sum_i |w_i^{target} - w_i^{drifted}|\); no transaction costs
                    • Outcomes: annualised mean return and volatility, 5% ES of daily returns, maximum drawdown, turnover per rebalance

                    Results: Covariance-Based and Penalty-Based Portfolio Rules

                    Slide 37 in the PDF

                    The table and chart compare the three rules from 30 June 2020 to 31 May 2026, on EODHD prices and in the paper. The minimum-variance portfolio has the lowest risk on every measure: an annualised volatility of 1.43% against 3.92% for equal weight, an ES 5% of 15.9 bp against 49.0 bp, and a maximum drawdown of 1.7% against 6.5%. It cuts volatility by 63% and ES by 67% relative to equal weight; the paper finds 71% and 62%. The linear-lambda rule has the highest volatility, 6.72%, the deepest drawdown, 19.0%, and a turnover of 57.0% per rebalance: weights based on the penalty add risk and trading without improving returns. The paper's levels are higher, but the ranking is the same.

                    In brief

                    EODHD prices, 30 Jun 2020 – 31 May 2026working paper
                    RulereturnvolES 5% (bp)max DDturnovervolES 5%max DD
                    equal weight\(-0.51\%\)\(3.92\%\)\(49.0\)\(6.5\%\)\(0.3\%\)\(10.93\%\)\(101.4\)\(15.6\%\)
                    min. variance\(0.05\%\)\(1.43\%\)\(15.9\)\(1.7\%\)\(3.7\%\)\(3.14\%\)\(39.0\)\(2.6\%\)
                    linear lambda\(-2.46\%\)\(6.72\%\)\(84.2\)\(19.0\%\)\(57.0\%\)\(10.76\%\)\(125.7\)\(16.5\%\)
                    Results: Covariance-Based and Penalty-Based Portfolio Rules
                    • Robust result: GMV cuts volatility by 63% and ES by 67% against equal weight (working paper: 71%, 62%); penalty-based weights add risk and turnover
                    BA_L6_portfolios

                    Value and Drawdown, 2020–2026

                    Slide 38 in the PDF

                    The left panel shows the value of 1 USD invested in each rule, in reference-adjusted terms, and the right panel its drawdown, the percentage fall from the previous peak, from mid-2020 to October 2026; the dotted red line marks the end of the paper's sample. The GMV portfolio stays within 1.7% of its peak over six years. The equal-weight portfolio falls 19.9% below its peak once sUSD collapses, after the paper's sample ends. The linear-lambda portfolio loses about 6% in its first months in 2020 and more in each stress episode: risk without a reward.

                    In brief

                    Value and Drawdown, 2020–2026
                    • GMV stays within 1.7% of its peak over six years; equal weight falls 19.9% below its peak once sUSD collapses (after the paper's sample)
                    • Linear lambda loses about 6% in its first months (2020) and more in each stress episode: risk without a reward
                    • Source: own computation on EODHD daily prices of 10 stablecoins, 30 June 2020 – 23 October 2026
                    BA_L6_nav

                    What the Minimum-Variance Portfolio Holds

                    Slide 39 in the PDF

                    The chart stacks the weights of the GMV portfolio over time. The average weights are 28.8% for USDC, 21.1% for USDT, 16.9% for DAI, 14.9% for BUSD and 8.2% for TUSD; sUSD, PAXG and EURS hold less than 1% each. The cap of 30% binds often, visible as flat bands at 30% and 60%: without it the portfolio would hold even more USDC and USDT, that is more concentration on two issuers.

                    In brief

                    What the Minimum-Variance Portfolio Holds
                    • Average weights: USDC 28.8%, USDT 21.1%, DAI 16.9%, BUSD 14.9%, TUSD 8.2%; sUSD, PAXG, EURS below 1% each
                    • The cap binds often: without it GMV would hold even more USDC and USDT, i.e. more issuer concentration
                    • Source: own computation on EODHD daily prices of 10 stablecoins, 30 June 2020 – 23 October 2026
                    BA_L6_gmv_weights

                    Worked Example: Penalty Coefficients as Portfolio Weights

                    Slide 40 in the PDF

                    The selected penalty \(\hat\lambda_{j,t}\) is a regularisation choice for one equation. It depends on how the other coins move in coin \(j\)'s lower tail, not on \(j\)'s own risk. A small, erratic token, with high variance and few co-movements, can therefore receive a low \(\hat\lambda\) and then a 30% weight under the linear rule. On EODHD prices to May 2026, the linear-lambda volatility is 6.7% against 1.4% for GMV, and its turnover of 57.0% per week shows that the ranking of the penalties changes at almost every rebalance. Extending the sample to 23 October 2026, when sUSD fell from 0.39 to 0.09 USD, makes the difference larger: equal weight reaches a volatility of 5.74%, an ES 5% of 70.5 bp and a maximum drawdown of 19.9%, three times its 6.5% to May 2026; linear lambda reaches 10.62%, 111.4 bp and 30.6%; GMV stays at 1.39%, 15.3 bp and 1.7%, because it had already moved out of sUSD when its 90-day volatility exploded to 364 bp a day before 12 June 2026.

                    In brief

                    • \(\hat\lambda_{j,t}\) is the regularisation chosen for one equation; it depends on how the other coins move in \(j\)'s lower tail, not on \(j\)'s own risk
                    • A small, erratic token (high variance, few co-movements) can get a low \(\hat\lambda\) and then a 30% weight under the linear rule
                    • In numbers (EODHD prices, to May 2026): linear lambda volatility \(6.7\%\) vs \(1.4\%\) for GMV; turnover \(57.0\%\) per week: the ranking of \(\hat\lambda\) changes almost every rebalance
                    • Over the longer sample to 23 October 2026, sUSD falls from 0.39 to 0.09 USD:
                    30 Jun 2020 – 23 Oct 2026volES 5% (bp)max DD
                    equal weight\(5.74\%\)\(70.5\)\(19.9\%\)
                    min. variance\(1.39\%\)\(15.3\)\(1.7\%\)
                    linear lambda\(10.62\%\)\(111.4\)\(30.6\%\)
                    • Equal weight triples its drawdown (6.5% to 19.9%) with sUSD's collapse; GMV had already moved out because sUSD's 90-day volatility exploded (364 bp a day before 12 June 2026)

                    References

                    • Adrian, T. & Brunnermeier, M. K. (2016). CoVaR. American Economic Review, 106(7), 1705–1741. https://doi.org/10.1257/aer.20120555
                    • Belloni, A. & Chernozhukov, V. (2011). ℓ1-Penalized Quantile Regression in High-Dimensional Sparse Models. The Annals of Statistics, 39(1), 82–130. https://doi.org/10.1214/10-AOS827
                    • Fan (2026). Large Quantile Lasso FRM for Stable Coin Depegging Risks. MSc thesis, Humboldt-Universität zu Berlin.
                    • Fan, Sha, Härdle & Pele (2026). Reference-adjusted Quantile Lasso reveals regime-dependent stablecoin depegging pressure. Working paper.
                    • Härdle, W. K., Wang, W. & Yu, L. (2016). TENET: Tail-Event Driven NETwork Risk. Journal of Econometrics, 192(2), 499–513. https://doi.org/10.1016/j.jeconom.2016.02.013
                    • Koenker, R. & Bassett, G. (1978). Regression Quantiles. Econometrica, 46(1), 33–50. https://doi.org/10.2307/1913643
                    • Lyons, R. K. & Viswanath-Natraj, G. (2023). What Keeps Stablecoins Stable? Journal of International Money and Finance, 131, 102777. https://doi.org/10.1016/j.jimonfin.2022.102777
                    • Mihoci, A., Althof, M., Chen, C. Y.-H. & Härdle, W. K. (2020). FRM Financial Risk Meter. In The Econometrics of Networks (Advances in Econometrics, Vol. 42), 335–368. Emerald Publishing. https://doi.org/10.1108/S0731-905320200000042016
                    • Tibshirani, R. (1996). Regression Shrinkage and Selection via the Lasso. Journal of the Royal Statistical Society: Series B (Methodological), 58(1), 267–288. https://doi.org/10.1111/j.2517-6161.1996.tb02080.x
                    • Yu, L., Härdle, W. K., Borke, L. & Benschop, T. (2023). An AI Approach to Measuring Financial Risk. The Singapore Economic Review, 68(5), 1529–1549. https://doi.org/10.1142/S0217590819500668