Blockchain Applications

Lecture 3: Crypto Indices: CRIX, VCRIX and Investible Index Funds

Lecture 3: Crypto Indices: CRIX, VCRIX and Investible Index Funds

Daniel Traian PELE

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

Lecture Question and Outline

Slide 2 in the PDF

The question of the lecture is practical: how can a market of thousands of coins be summarised in one number that an investor can actually buy? A fund needs a benchmark against which its managers are judged; an exchange-traded product needs an index that it can replicate at low cost, every day, with coins that can be bought and sold in size.

The route has six steps, followed by a wider view. The first explains why an index is needed and what makes crypto difficult: concentration and survivorship bias. The second builds CRIX step by step: the Laspeyres formula, the divisor, the liquidity rule and the information criterion that fixes the number of constituents. The third replicates the index on real data and compares the replica with the published CRIX. The fourth describes the real CRIX from 2017 to 2026, and the fifth its volatility index VCRIX. The sixth moves from the index to a fund and measures how closely a spot Bitcoin fund follows Bitcoin. The last section places CRIX among nine types of crypto indices, with their formulas. Lecture 4 then develops the generalised autoregressive conditional heteroskedasticity (GARCH) model used here for CRIX into a full toolkit for volatility and tail risk.

In brief

  • Question: how do we summarise a market of thousands of coins in one number that an investor can actually buy?
    • a fund needs a benchmark; an exchange-traded product needs an index it can replicate at low cost
  • Outline
    • why an index: concentration and survivorship bias
    • how CRIX is built: Laspeyres formula, divisor, liquidity rule, number of constituents by an information criterion
    • our own replication on real data, compared with the published CRIX
    • the real CRIX 2017–2026 and VCRIX, its volatility index
    • from index to fund: tracking difference and tracking error of a spot Bitcoin fund
    • beyond CRIX: nine types of crypto indices, their formulas, two of them rebuilt on our data
  • Next lecture (Lecture 4): volatility models and tail risk; the generalised autoregressive conditional heteroskedasticity (GARCH) model of this lecture is its starting point

Learning Outcomes

Slide 3 in the PDF
  • Explain what a market index must do: be representative, investible and stable
  • Compute a Laspeyres index with a divisor and explain why the divisor changes at a rebalancing
  • Read an information-criterion curve and explain how CRIX chooses its number of constituents
  • Classify a crypto index into one of nine types and read its weighting formula
  • Interpret correlation, tracking error and tracking difference between an index, a replica and a fund
  • Read compound annual growth rate (CAGR), volatility and drawdown statistics of CRIX, and the VCRIX volatility index

Abbreviations (1)

Slide 4 in the PDF
ADTV average daily traded volume
AIC Akaike information criterion
BGCI Bloomberg Galaxy Crypto Index
BFIC CoinDesk Bitcoin Futures Excess Return Index
BITX Bitwise 10 Large Cap Crypto Index
BTC, ETH, … coin tickers: Bitcoin, Ethereum, Dogecoin (DOGE), Cardano (ADA), Litecoin (LTC), Bitcoin Cash (BCH), Ethereum Classic (ETC), Monero (XMR), Dash (DASH), Zcash (ZEC)
CAGR compound annual growth rate
CD20 CoinDesk 20 index
CME Chicago Mercantile Exchange
CMC CoinMarketCap
COINTILT Coinbase Size Tilt Index
COVID-19 coronavirus disease 2019
CRIX CRypto IndeX
DACS Digital Asset Classification Standard (CoinDesk)
ECRIX, EFCRIX CRIX variants: first rise / global minimum of the AIC
EODHD EOD (end-of-day) Historical Data, data vendor
ETF exchange-traded fund
ETP exchange-traded product
EVIV Ethereum Volmex Implied Volatility 30 Day Index
EWMA exponentially weighted moving average
FABIZ Faculty of Business Administration in Foreign Languages, ASE
FRED Federal Reserve Economic Data (St. Louis Fed)
FTX crypto exchange (company name), bankrupt in November 2022

Abbreviations (2)

Slide 5 in the PDF
GARCH generalised autoregressive conditional heteroskedasticity
HAC heteroskedasticity- and autocorrelation-consistent
HAR heterogeneous autoregressive (model)
HHI Herfindahl–Hirschman index
IBIT iShares Bitcoin Trust ETF (ticker)
IDA Institute Digital Assets
IQR interquartile range
MBB Master Blockchain for Business
MDVT median daily value traded
MVTT10F MarketVector Token Terminal Fundamental Index
NCI Nasdaq Crypto Index
OLS ordinary least squares
RV realised volatility
S&P Standard & Poor's; S&P 500: US stock index; DAX: German stock index
SE standard error
SEC US Securities and Exchange Commission
TD, TE tracking difference, tracking error
TMI total market index
USD US dollar; US United States
UTC Coordinated Universal Time
VCRIX volatility index of CRIX
VIX Cboe Volatility Index (S&P 500 options)
VWAP volume-weighted average price

Reading and Tools

Slide 6 in the PDF

Three texts are the core reading. Trimborn & Härdle (2018) define CRIX and its rule for the number of constituents; Kim, Trimborn & Härdle (2021) build VCRIX, the volatility index of CRIX; Găman, Będowska-Sójka, Pele & Härdle (2026) is a white paper of the IDA Institute Digital Assets that classifies crypto indices into types and describes their methodologies. For further reading, Trimborn, Li & Härdle (2020) study how liquidity constraints change portfolios of cryptocurrencies, and Petukhina et al. (2021) evaluate portfolio allocation strategies that include cryptocurrencies.

The data come from five sources. CRIX and VCRIX were read from the chart data of thecrix.de. Daily prices and traded volumes in US dollars (USD) of ten coins — Bitcoin (BTC), Ethereum (ETH), Dogecoin (DOGE), Cardano (ADA), Litecoin (LTC), Bitcoin Cash (BCH), Ethereum Classic (ETC), Monero (XMR), Dash (DASH) and Zcash (ZEC) — and of IBIT come from EODHD, to 30 October 2026. The circulating supply of each coin, needed for market values, comes from Coin Metrics, and the effective federal funds rate from the Federal Reserve Economic Data (FRED) of the St. Louis Fed. All files are in the `data/` folder of the course repository, and the lecture notebook reproduces every number.

In brief

  • Core reading
    • Trimborn & Härdle (2018), CRIX an index for cryptocurrencies, Journal of Empirical Finance 49, 107–122
    • Kim, Trimborn & Härdle (2021), VCRIX: a volatility index for crypto-currencies, International Review of Financial Analysis 78, 101915
    • Găman, Będowska-Sójka, Pele & Härdle (2026), White paper on cryptocurrency indices, IDA Institute Digital Assets, working paper: types of crypto indices and their methodologies
  • Data
    • Royalton CRIX and VCRIX, daily, from the chart data of thecrix.de (CRIX from 6 September 2017, VCRIX from 4 January 2018, to 6 November 2026)
    • prices and volumes from EOD Historical Data (EODHD), to 30 October 2026: ten coins, Bitcoin (BTC), Ethereum (ETH), Dogecoin (DOGE), Cardano (ADA), Litecoin (LTC), Bitcoin Cash (BCH), Ethereum Classic (ETC), Monero (XMR), Dash (DASH), Zcash (ZEC), and the iShares Bitcoin Trust exchange-traded fund (ETF), ticker IBIT
    • coin supplies from the Coin Metrics Community API; effective federal funds rate from FRED: DFF
  • Video units on Quantinar (course Blockchain Analytics)
  • Further: Trimborn, Li & Härdle (2020); Petukhina et al. (2021)

While most CC indices look similar at the structural level, their design choices create fundamentally different exposures.

— Găman, Będowska-Sójka, Pele & Härdle (2026), White paper on cryptocurrency indices

Lectures 1 and 2 studied individual coins: their prices, returns, volatility and crashes. An investor, a fund manager or a regulator who wants to speak about "the crypto market" needs something else: one number that summarises thousands of coins, that moves like the market as a whole, and that a fund can actually buy. This number is a market index. Building it for crypto is harder than for equities, because the market is concentrated in one or two coins, because coins appear and disappear continuously, and because many small coins are hardly traded.

The lecture follows the CRIX (CRypto IndeX) of Trimborn & Härdle (2018), today published as the Royalton CRIX Crypto Index. CRIX is a market-value index with a divisor, like the major stock indices, but it chooses its number of constituents with a statistical rule: an information criterion that compares candidate indices with the total market. The lecture explains the construction with full formulas, replicates the index on real data, describes the real CRIX and its volatility index VCRIX from 2017 to 2026, and asks the question every index product faces: how closely does a fund follow its index? It ends with a wider view: following Găman, Będowska-Sójka, Pele & Härdle (2026), CRIX is one of nine types of crypto indices, from market and sector indices to volatility, yield and sentiment indices, and two of them are rebuilt on our data.

The data are the daily values of CRIX and VCRIX from thecrix.de (CRIX from 6 September 2017, VCRIX from 4 January 2018, both to 6 November 2026), daily prices and volumes of ten coins and of the iShares Bitcoin Trust exchange-traded fund (IBIT) from the data vendor EOD Historical Data (EODHD), hourly Bitcoin prices from the same vendor, and the circulating supply of the ten coins from the Coin Metrics Community API. The effective federal funds rate, used for the cash part of a yield index, comes from FRED: DFF.

Left: the published Royalton CRIX (red) and our replica built on ten coins with the CRIX rules (blue), both 1,000 on 31 March 2018, log scale; dashed green: the replica with weights capped below 50%. Right: the correlation of the daily log returns of CRIX and the replica on rolling 180-day windows. Data to 30 October 2026.
Left: the published Royalton CRIX (red) and our replica built on ten coins with the CRIX rules (blue), both 1,000 on 31 March 2018, log scale; dashed green: the replica with weights capped below 50%. Right: the correlation of the daily log returns of CRIX and the replica on rolling 180-day windows. Data to 30 October 2026.

What you will be able to do

  • Explain what a market index must do: be representative, investible and stable.
  • Compute a Laspeyres index with a divisor and explain why the divisor changes at a rebalancing.
  • Read an information-criterion curve and explain how CRIX chooses its number of constituents.
  • Classify a crypto index into one of nine types and read its weighting formula.
  • Interpret correlation, tracking error and tracking difference between an index, a replica and a fund.
  • Read compound annual growth rate (CAGR), volatility and drawdown statistics of CRIX, and the VCRIX volatility index.

The Appendix holds data notes and the details of the kernel density estimator.

Lecture plan

  1. Motivation for a Crypto IndexWhat an index must do, the concentration of the crypto market and survivorship bias.
  2. Constructing CRIXThe Laspeyres formula, the divisor, the liquidity rule, the information criterion for the number of constituents, its smoothing over six months and capped weights.
  3. Our Replication on Real DataA CRIX-style index on ten coins compared with the published CRIX; tracking error and tracking difference.
  4. The Real CRIX, 2017–2026Growth, volatility, drawdowns, crisis episodes and calendar years, against Bitcoin.
  5. VCRIX: A Volatility IndexConstruction, the 30-day window, volatility after falls and rises, a GARCH comparison and a HAR forecast.
  6. Investible Index ProductsFrom an academic index to an investible one, from an index to a fund, and the tracking of a spot Bitcoin fund.
  7. Nine Types of Crypto IndicesMarket, sector, price, futures, volatility, yield, smart-beta, fundamentals and sentiment indices; three weightings and a protection floor on real data.
  8. SummaryThe main numbers, two short quizzes and five interpretation exercises.
  9. Appendix: Further MaterialData notes and kernel density estimation with leave-one-out.

Motivation for a Crypto Index

An index is useful only if it does three things at once: represent the market, be investible, and be stable. The section explains these requirements, shows how concentrated the crypto market is, and measures on real data how much a backtest built from today's winners overstates past returns.

Requirements for a Market Index

Slide 7 in the PDF

A benchmark is one number for the performance of a market. Fund managers are compared with it, traders hedge with products built on it, and researchers use it as "the market" in their models. For equities the S&P 500 or the DAX play this role. For crypto there is no single obvious choice, and the choice matters.

Three requirements define a good index. It must be representative: its returns should move like those of the whole market, which we call the total market index (TMI), the market-value index of all eligible coins. It must be investible: it should contain few and liquid constituents, so that a fund can buy and hold them at low cost. And it must be stable: its rules are fixed in advance, and its constituents do not change every day.

Two features of crypto make these requirements hard to meet together. The market is concentrated: one or two coins hold most of the value, so that a market-value index is close to a Bitcoin index. And coins enter and exit continuously: new coins are launched every day and others are abandoned, so a list of today's coins hides the ones that died. The trade-off is clear: more constituents represent the market better but cost more to trade. CRIX resolves it with a statistical rule that chooses the number of constituents from the data (Trimborn & Härdle, 2018).

In brief

  • Benchmark: one number for the performance of a market; funds, traders and researchers compare against it
  • Three requirements
    • representative: it moves like the whole market, the total market index (TMI) of all eligible coins
    • investible: few, liquid constituents that a fund can buy and hold at low cost
    • stable: rules fixed in advance; constituents do not change every day
  • Two features of crypto make this hard
    • concentration: one or two coins hold most of the value
    • entry and exit: coins appear and vanish; a list of today's coins hides the dead ones
  • Trade-off: more constituents represent the market better but cost more to trade; CRIX chooses the number with a statistical rule Trimborn & Härdle (2018)

Concentration of the Crypto Market

Slide 8 in the PDF

The chart shows the market values of the ten coins with public supply data, computed as price times current supply, from 1 January 2018 to 30 October 2026. Bitcoin's share of the ten rose from 57.3% on 1 January 2018 to 82.2% on 30 October 2026, with a maximum of 87.1% on 21 April 2025. Bitcoin held more than half of the value of the ten coins on 99.8% of the days.

The right panel turns the shares into one number, the effective number of coins \(1/\mathrm{HHI}\), where HHI is the Herfindahl–Hirschman index defined on the next slide. It was 2.97 in January 2018, when Ethereum and several other coins still held large shares, and is 1.44 today. In this universe, ten coins behave like fewer than one and a half equal coins.

In brief

Concentration of the Crypto Market
  • Ten coins with public supply data, market value = price \(\times\) current supply, 1 January 2018 – 30 October 2026
  • Bitcoin: 57.3% of the ten on 1 January 2018, 82.2% on 30 October 2026 (maximum 87.1%, 21 April 2025); above 50% on 99.8% of the days
  • Effective number of coins \(1/\mathrm{HHI}\), with HHI the Herfindahl–Hirschman index: 2.97 in January 2018, 1.44 today
BA_L3_concentration

Worked Example: the Herfindahl–Hirschman Index

Slide 9 in the PDF

The Herfindahl–Hirschman index (HHI) is the measure that competition authorities use for market concentration. With value shares \(w_i = M_i/\sum_j M_j\) of \(N\) coins,

\[\mathrm{HHI} = \sum_{i=1}^{N} w_i^2, \qquad N_{\text{eff}} = \frac{1}{\mathrm{HHI}}\]
(3.1)

The HHI ranges from \(1/N\), when all coins have equal shares, to 1, when one coin holds everything; its inverse is the number of equal coins that would give the same concentration.

On 30 October 2026 Bitcoin held 82.2% of the ten coins, Ethereum 14.5% and the other eight 3.4% together. The squares are \(0.822^2 = 0.675\) and \(0.145^2 = 0.021\); the eight small coins add almost nothing, so \(\mathrm{HHI} = 0.696\) and \(N_{\text{eff}} = 1/0.696 = 1.44\). A market-value index of these coins is therefore mostly a Bitcoin index. A capped index limits the weight of the largest constituent, for example below 50%, a rule discussed later in the lecture.

In brief

  • HHI of the value shares \(w_i = M_i/\sum_j M_j\):
    \[ \mathrm{HHI} = \sum_{i=1}^{N} w_i^2, \qquad N_{\text{eff}} = 1/\mathrm{HHI} \]
  • Ten coins, 30 October 2026: Bitcoin 82.2%, Ethereum 14.5%, the other eight 3.4% together
    • \(0.822^2 + 0.145^2 = 0.675 + 0.021 = 0.696\); the eight small coins add almost nothing
    • \(N_{\text{eff}} = 1/0.696 = 1.44\): the ten coins behave like 1.4 equal coins
  • Consequence for an index
    • a market-value index of these coins is mostly a Bitcoin index
    • a capped index limits the largest constituent, for example below 50% (later in this lecture)

Survivorship Bias

Slide 10 in the PDF

Survivorship bias arises when a sample is built from the assets that are alive today: the assets that failed or shrank are missing, and the past returns of the survivors look better than what an investor could actually have earned. Brown et al. (1992) showed how strong this effect is in studies of fund performance. In crypto it is severe, because coins are launched, abandoned or delisted every day.

A closely related error is look-ahead selection: choosing the constituents of a historical index with information that was not available at the time. A backtest that starts in 2018 with the largest coins of 2026 knows, in 2018, which coins will grow. CRIX avoids both errors by construction: at each reselection date its constituents are chosen from the data known on that date (Trimborn & Härdle, 2018).

The test on our ten coins compares two baskets, both bought on 1 January 2018. The ex-ante basket holds the five largest coins of that day, the choice an investor could really have made. The ex-post basket holds the five largest coins of 30 October 2026, the choice a backtest built today would make.

In brief

  • Survivorship bias: a sample built from the assets alive today ignores those that failed or shrank Brown et al. (1992)
    • past returns of the survivors look better than what an investor could have earned
  • In crypto it is severe: coins are launched, abandoned or delisted every day
  • A related error: look-ahead selection, choosing constituents with information that was not available at the time
  • Remedy built into CRIX: constituents are chosen at each reselection date from the data known on that date Trimborn & Härdle (2018)
  • Test on our ten coins
    • ex ante: buy on 1 January 2018 the five largest coins of that day
    • ex post: buy on 1 January 2018 the five largest coins of 30 October 2026

Survivorship Bias: Ex-Ante and Ex-Post Baskets since 2018

Slide 11 in the PDF
    Question for the room Switch the purchase date to 1 January 2024, with five coins and equal weights. How large is the gap between the ex-post and the ex-ante basket?
    Answer

    Very large: the ex-ante top 5 of 1 January 2024 (BTC, ETH, ADA, DOGE, LTC) returned −3.0% a year to 30 October 2026, the ex-post top 5 (BTC, ETH, ZEC, DOGE, XMR) 129.8% a year. The shorter the period, the more an ex-post list is made of the recent winners: Zcash enters the top 3 only after its rise in 2025–2026. From 1 January 2018 the gap is 17.6% against 7.1% a year; with N = 2 it disappears, because Bitcoin and Ethereum are the top two at both dates.

    With equal weights, 100 USD invested in the ex-ante basket (BTC, ETH, BCH, ADA, LTC) became 184 USD by 30 October 2026, a compound annual growth rate of 7.1%. The same 100 USD in the ex-post basket (BTC, ETH, ZEC, DOGE, XMR) became 418 USD, 17.6% a year. The difference comes almost entirely from the selection: Zcash ranked 9th of the ten coins by market value in 2018 and 3rd in 2026, Dogecoin 10th and 4th, while Bitcoin Cash, Cardano and Litecoin, large in 2018, lost rank.

    With market-value weights the gap is smaller, 401 USD against 491 USD (CAGR 17.1% against 19.8%), because Bitcoin and Ethereum dominate both baskets. The lesson is general: any historical performance of a list of "today's top coins" overstates what an investor could have earned, and the overstatement is largest for equally weighted baskets, which give the most weight to the small coins whose fortunes changed most.

    In brief

    • Ex ante (BTC, ETH, BCH, ADA, LTC), equal weights: 100 USD became 184 USD, CAGR 7.1%
    • Ex post (BTC, ETH, ZEC, DOGE, XMR): 100 USD became 418 USD, CAGR 17.6%; Zcash ranked 9th of the ten in 2018 and 3rd in 2026
    • With market-value weights the gap is smaller (401 against 491 USD), because Bitcoin and Ethereum dominate both baskets
    BA_L3_survivorship

    Section summary

    The crypto market is highly concentrated, with an effective number of about 1.4 coins among the ten studied, and a basket chosen with today's knowledge earns far more in a backtest than a basket an investor could have chosen at the time. A good crypto index must therefore select its constituents with information available at each date, and must decide how much weight the dominant coin may carry.

    Constructing CRIX

    The section builds CRIX from its formulas: the Laspeyres index with a divisor, the rules for rebalancing and reselection, the liquidity rule, the information criterion that fixes the number of constituents, its smoothed version over six months, and capped weights. Each step is illustrated with numbers from our own replication.

    The Laspeyres Index and the Divisor

    Slide 12 in the PDF

    A Laspeyres index values today, at today's prices, a basket fixed at a base time 0:

    \[\mathrm{CRIX}_t = 1000 \times \frac{\sum_i P_{it}\,Q_{i0}}{\sum_i P_{i0}\,Q_{i0}}\]
    (3.2)

    where \(P_{it}\) is the price of coin \(i\) at time \(t\) and \(Q_{i0}\) the number of coins of \(i\) in circulation at time 0. The numerator is the market value of the basket today, the denominator its value at the base time; the index starts at 1,000.

    CRIX uses the same idea in market-value form (Trimborn & Härdle, 2018):

    \[\mathrm{CRIX}_t = \frac{\sum_{i=1}^{k} \mathrm{MV}_{it}}{\mathrm{Divisor}_t}, \qquad \mathrm{MV}_{it} = P_{it}\,Q_i, \qquad \mathrm{Divisor}_0 = \frac{\sum_i \mathrm{MV}_{i0}}{1000}\]
    (3.3)

    with \(k\) constituents. Each coin enters with a weight proportional to its market value, \(w_{it} = \mathrm{MV}_{it}/\sum_j \mathrm{MV}_{jt}\). Between two rebalancing days the quantities \(Q_i\) are fixed, so only prices move the index: a coin whose price doubles doubles its contribution to the numerator.

    In brief

    • Laspeyres index: today's prices of a basket fixed at base time 0
      \[ \mathrm{CRIX}_t = 1000 \times \frac{\sum_{i} P_{it}\,Q_{i0}}{\sum_{i} P_{i0}\,Q_{i0}} \]
      • \(P_{it}\): price of coin \(i\) at time \(t\); \(Q_{i0}\): number of coins of \(i\) in circulation at time 0
    • Market-value form used by CRIX Trimborn & Härdle (2018):
      \[ \mathrm{CRIX}_t = \frac{\sum_{i=1}^{k} \mathrm{MV}_{it}}{\mathrm{Divisor}_t}, \qquad \mathrm{MV}_{it} = P_{it}\,Q_{i}, \qquad \mathrm{Divisor}_0 = \frac{\sum_i \mathrm{MV}_{i0}}{1000} \]
      • \(k\): number of constituents; the index starts at 1 000
      • weights are proportional to market value: \(w_{it} = \mathrm{MV}_{it}/\sum_j \mathrm{MV}_{jt}\)
    • Between two rebalancing days the quantities \(Q_i\) are fixed; only prices move the index

    Reasons for Changes in the Divisor

    Slide 13 in the PDF

    On a rebalancing day \(\tau\) the basket changes: a coin enters or leaves, or the quantities \(Q_i\) are updated to the new circulating supply. Without an adjustment the index would jump on that day even if no price moved. The divisor removes this jump: it is chosen so that the index computed with the new basket equals the index computed with the old basket,

    \[\frac{\sum_i \mathrm{MV}^{\text{old}}_{i\tau}}{\mathrm{Divisor}_{\text{old}}} = \mathrm{CRIX}_\tau = \frac{\sum_i \mathrm{MV}^{\text{new}}_{i\tau}}{\mathrm{Divisor}_{\text{new}}} \quad\Longrightarrow\quad \mathrm{Divisor}_{\text{new}} = \frac{\sum_i \mathrm{MV}^{\text{new}}_{i\tau}}{\mathrm{CRIX}_\tau}\]
    (3.4)

    The same device keeps stock indices such as the S&P 500 or the DAX continuous when a company enters or leaves the index. In CRIX two calendars apply (thecrix.de). Monthly, the weights are recalculated from current market values (rebalancing). Quarterly, the constituents themselves are reselected (reconstitution). The difference matters for an investor: rebalancing changes the amounts held, reconstitution may replace coins.

    In brief

    • At a rebalancing day \(\tau\) the basket changes: a coin enters or leaves, or the quantities \(Q_i\) are updated
    • The index must not jump because of the change; only prices may move it:
      \[ \frac{\sum_i \mathrm{MV}^{\text{old}}_{i\tau}}{\mathrm{Divisor}_{\text{old}}} = \mathrm{CRIX}_\tau = \frac{\sum_i \mathrm{MV}^{\text{new}}_{i\tau}}{\mathrm{Divisor}_{\text{new}}} \;\Longrightarrow\; \mathrm{Divisor}_{\text{new}} = \frac{\sum_i \mathrm{MV}^{\text{new}}_{i\tau}}{\mathrm{CRIX}_\tau} \]
    • The same device keeps stock indices such as the S&P 500 (US) or the DAX (Germany) continuous when a company enters or leaves
    • CRIX rules thecrix.de
      • monthly: the weights are recalculated from current market values (rebalancing)
      • quarterly: the constituents are reselected (reconstitution)

    Worked Example: a Rebalancing Day of the Replica

    Slide 14 in the PDF

    Our replica, described later in the lecture, had a quarterly reselection day on 30 September 2026, with four constituents: Bitcoin, Ethereum, Zcash and Dogecoin. The index stood at \(\mathrm{CRIX}_\tau = 9{,}521.15\). The table gives prices, supplies, market values and weights on that day, and the prices one month later.

    The calculation has three steps. First, the market value of the new basket is \(\sum_i \mathrm{MV}_{i\tau} = 1{,}678.74 + 327.65 + 24.27 + 14.75 = 2{,}045.41\) billion USD. Second, the new divisor follows from (3.4): \(2{,}045.41 \times 10^9 / 9{,}521.15 = 214.83\) million, against 214.67 million before the change. The index level is unchanged on that day; only the divisor moves. Third, on 30 October 2026 the quantities are the same and the prices have changed: the basket is worth \(\sum_i P_{it}Q_i = 1{,}857.94\) billion USD, and the index is \(1{,}857.94 \times 10^9/(214.83 \times 10^6) = 8{,}648.5\), a fall of 9.2% in one month, mostly because Bitcoin fell from 83,553.85 to 77,198.21 USD.

    In brief

    Coinprice, 30 Sep 2026 (USD)supply (million)market value (billion USD)weight (%)price, 30 Oct 2026 (USD)
    Bitcoin83 553.8520.091 678.7482.0777 198.21
    Ethereum2 683.68122.09327.6516.022 241.05
    Zcash1 436.4016.9024.271.191 268.64
    Dogecoin0.0945156 13614.750.720.0759
    • Our replica (later in this lecture) on its quarterly reselection day, 30 September 2026: index level \(\mathrm{CRIX}_\tau = 9\,521.15\)
    • Step 1: new basket value \(\sum_i \mathrm{MV}_{i\tau} = 2\,045.41\) billion USD
    • Step 2: \(\mathrm{Divisor}_{\text{new}} = 2\,045.41 \times 10^9 / 9\,521.15 = 214.83\) million (before: 214.67 million)
    • Step 3, 30 October 2026: same quantities, new prices, \(\sum_i P_{it}Q_i = 1\,857.94\) billion USD
      • \(\mathrm{CRIX}_t = 1\,857.94 \times 10^9 / (214.83 \times 10^6) = 8\,648.5\), a fall of 9.2% in one month

    The Divisor through Time

    Slide 15 in the PDF

    The left panel shows the replica index, 1,000 on 31 March 2018; the right panel its divisor, recomputed at each of the 103 monthly rebalancing days. The divisor moves in steps: when a coin enters or leaves, and when supplies change. The index level is unchanged on each of these days, by construction.

    The general upward drift of the divisor has a simple reason. New Bitcoin, Ethereum and Dogecoin are issued continuously, so the circulating supply grows; at each rebalancing the larger quantities raise \(\sum_i \mathrm{MV}_{i\tau}\), and the divisor rises in proportion so that the index does not. Growth in supply is not a return for the holder of one coin, and the divisor keeps it out of the index.

    In brief

    The Divisor through Time
    • Left: the replica index, 1 000 on 31 March 2018; right: its divisor, recomputed at each of the 103 monthly rebalancing days
    • The divisor moves in steps (new coins, new supplies); the index level is unchanged on each of these days
    • Rising supply (new Bitcoin, Ethereum and Dogecoin issued) raises \(\sum_i \mathrm{MV}_{i\tau}\) and with it the divisor
    BA_L3_divisor

    Eligibility of Coins

    Slide 16 in the PDF

    Before ranking, CRIX filters the universe. The liquidity rule (thecrix.de) keeps a coin only if its average daily traded volume (ADTV) in the last period is at least the 25th percentile of the ADTVs of all coins:

    \[\mathrm{ADTV}_i \ge \mathrm{ADTV}_{0.25}\]
    (3.5)

    The least traded quarter of the universe is excluded: a coin may have a large market value but trade so little that a fund could not buy it without moving its price.

    Stablecoins are excluded: their price is pegged, not formed by the market (Royalton Partners: CRIX). The eligible coins are then ranked by market value, and the candidate indices take the first \(k\) coins of the ranking.

    In our replica the liquidity rule is applied on the last 90 days before each quarterly reselection. Monero failed it in 33 of 35 quarters, Dash in 24 and Zcash in 22; Bitcoin, Ethereum, Litecoin and Bitcoin Cash always passed.

    In brief

    • Liquidity rule thecrix.de: coin \(i\) is eligible if
      \[ \mathrm{ADTV}_i \ge \mathrm{ADTV}_{0.25} \]
      • \(\mathrm{ADTV}_i\): average daily traded volume (USD) of coin \(i\) in the last period
      • \(\mathrm{ADTV}_{0.25}\): the 25th percentile of the ADTVs of all coins: the least traded quarter is excluded
    • Stablecoins are excluded: their price is pegged, not formed by the market Royalton Partners: CRIX
    • Eligible coins are ranked by market value; the candidate indices take the first \(k\) of the ranking
    • In our replica (90-day window): Monero failed the liquidity rule in 33 of 35 quarters, Dash in 24, Zcash in 22

    Number of Constituents and the Information Criterion

    Slide 17 in the PDF

    The central idea of CRIX is to choose the number of constituents \(k\) by comparing candidate indices with the total market (Trimborn & Härdle, 2018). The candidate \(\mathrm{CRIX}(k, \beta)\) holds the largest coin with weight 1 and the next \(k - 1\) coins with free weights \(\beta\), estimated so that the candidate tracks the TMI as closely as possible. Its tracking residual on day \(t\) is the difference of daily log returns:

    \[\hat\varepsilon_t(k, \beta) = r^{\mathrm{TMI}}_t - r^{\mathrm{CRIX}(k,\beta)}_t, \qquad r_t = \ln I_t - \ln I_{t-1}\]
    (3.6)

    The distribution of these residuals is estimated without assuming a Normal shape. The residuals of the one-coin index, \(\hat\varepsilon_s(1,1)\), define a kernel density with a Gaussian kernel \(\varphi\) and a bandwidth \(h\); to avoid evaluating each point on its own bump, the leave-one-out version is used (details in the Appendix):

    \[\hat f_{-t}(x) = \frac{1}{(n-1)h}\sum_{s \ne t}\varphi\Big(\frac{x - \hat\varepsilon_s(1,1)}{h}\Big), \qquad h = 0.9\,\min\big(\hat\sigma, \widehat{\mathrm{IQR}}/1.349\big)\,n^{-1/5}\]
    (3.7)

    where IQR is the interquartile range and \(\hat\sigma\) the standard deviation of the residuals. The Akaike information criterion (AIC) of Akaike (1974) then scores each candidate:

    \[\mathrm{AIC}(k) = -2\sum_{t=1}^{n}\ln \hat f\big(\hat\varepsilon_t(k, \hat\beta)\big) + 2\,(k - 1)\]
    (3.8)

    The first term measures fit: each added coin pushes the residuals towards zero, where the density is highest, so the log-likelihood rises. The second term is the penalty, 2 for each added coin. Two stopping rules are used: the ECRIX rule stops at the first \(k\) at which the AIC rises; the EFCRIX rule takes the global minimum.

    In brief

    • Candidate index \(\mathrm{CRIX}(k, \beta)\): the largest coin with weight 1, the next \(k - 1\) with free weights \(\beta\) Trimborn & Härdle (2018); its tracking residual:
      \[ \hat\varepsilon_t(k, \beta) = r^{\mathrm{TMI}}_t - r^{\mathrm{CRIX}(k,\beta)}_t, \qquad r_t = \ln I_t - \ln I_{t-1} \]
    • Density of the residuals of the one-coin index, by a Gaussian kernel with leave-one-out:
      \[ \hat f_{-t}(x) = \frac{1}{(n-1)h}\sum_{s \ne t}\varphi\Big(\frac{x - \hat\varepsilon_s(1,1)}{h}\Big), \qquad h = 0.9\,\min\big(\hat\sigma, \widehat{\mathrm{IQR}}/1.349\big)\,n^{-1/5} \]
      • \(\varphi\): standard Normal density; IQR: interquartile range; \(\hat\sigma\): standard deviation
    • Akaike information criterion (AIC) Akaike (1974) of the candidate with \(k\) constituents:
      \[ \mathrm{AIC}(k) = -2\sum_{t=1}^{n}\ln \hat f\big(\hat\varepsilon_t(k, \hat\beta)\big) + 2\,(k - 1) \]
      • fit: more coins push the residuals towards 0, where the density is highest; penalty: \(2\) per added coin
      • ECRIX rule: stop at the first \(k\) where the AIC rises; EFCRIX: the global minimum

    Smoothing the Decision over Six Months

    Slide 18 in the PDF

    An AIC curve computed on one day is noisy: a small change of prices can move its minimum from one \(k\) to another, and an index whose constituents change for such reasons is costly for a fund that follows it. Our replica therefore smooths the decision over six months. At each of the last six month-ends the AIC curve is computed on its own 90-day window; the six curves are pooled, polynomials in \(k\) of degree 1 to \(k_{\max} - 2\) are fitted through them, the degree with the lowest AIC of the fit is kept, and the minimum of the fitted polynomial gives the number of constituents.

    The smoothing has a price: when the market really changes, the rule reacts later than a rule based on one day.

    In brief

    • A single-day AIC curve is noisy: a small change of prices can move its minimum from one \(k\) to another
      • an index whose constituents change for such reasons is costly for a fund that follows it
    • Smoothed rule, used in our replica
      • compute the AIC curve at each of the last six month-ends, each on its own 90-day window
      • pool the six curves and fit polynomials in \(k\) of degree 1 to \(k_{\max} - 2\) through them
      • keep the degree with the lowest AIC of the fit; the minimum of the fitted polynomial gives \(k\)
    • The price: the rule reacts later when the market really changes

    The Decision Curve on Our Data

    Slide 19 in the PDF
      Question for the room The ▶ button runs through the 35 quarterly reselections, March 2018 to September 2026. How often does the first rise of the AIC choose only two constituents?
      Answer

      In 9 of 35 quarters; 3 constituents in 13, 4 in 11, 5 and 6 once each (mean 3.2). The curve always falls steeply from k = 1 to k = 2 (adding Ethereum) and is almost flat after that, so the choice between 2, 3 and 4 moves from quarter to quarter. The smoothed AIC rule (six months, polynomial) chooses 3 in 30 of the 35 quarters: smoothing over six months makes the number of constituents more stable.

      The left panel shows \(\mathrm{AIC}(k)\) for our replica on the 90 days to 30 September 2026. Seven coins passed the liquidity rule on that date; Ethereum Classic, Monero and Dash were excluded. From \(k = 1\) to \(k = 2\) the AIC falls by 284 points: adding Ethereum to Bitcoin is decisive for tracking the market. After \(k = 3\) the curve is almost flat.

      The right panel applies the smoothed rule: the AIC curves at the six month-ends up to 30 September 2026, each on its own 90-day window, pooled and smoothed by a polynomial in \(k\) whose degree, 5, is chosen by the AIC of the fit. The minimum of the fitted polynomial lies at \(k = 3\). On a small universe the rule illustrates the principle, not the commercial calculation.

      In brief

      • Left: \(\mathrm{AIC}(k)\) on the 90 days to 30 September 2026; seven coins pass the liquidity rule (ETC, XMR, DASH excluded)
      • From \(k = 1\) to \(k = 2\) the AIC falls by 284: adding Ethereum is decisive; after \(k = 3\) the curve is almost flat
      • Right: our smoothed rule, six month-end curves and a polynomial of degree 5; minimum at \(k = 3\)
      BA_L3_aic_curve

      Worked Example: Reading the AIC Values

      Slide 20 in the PDF

      The table lists the AIC values of the last quarter. The eligible coins, ordered by market value on 30 September 2026, are BTC, ETH, ZEC, DOGE, ADA, BCH and LTC.

      Reading the table takes three steps. First, the AIC falls until \(k = 4\) (\(-953.6\)) and rises at \(k = 5\) (\(-952.3\)), so the ECRIX rule chooses four constituents. Second, the global minimum is also at \(k = 4\), so the EFCRIX rule agrees. Third, the size of the gains tells how firm the choice is: the fourth coin improves the AIC by only \(953.6 - 953.3 = 0.3\) points, against 284 points for the second coin. On such a flat curve a small change in the data could move the choice between three and four constituents; this fragility is exactly what the smoothing over six months is designed to remove.

      In brief

      \(k\)1234567
      \(\mathrm{AIC}(k)\)\(-662.2\)\(-946.2\)\(-953.3\)\(-953.6\)\(-952.3\)\(-950.7\)\(-948.8\)
      • 90 days to 30 September 2026; eligible coins by market value: BTC, ETH, ZEC, DOGE, ADA, BCH, LTC
      • Step 1: the AIC falls until \(k = 4\) (\(-953.6\)) and rises at \(k = 5\) (\(-952.3\)): ECRIX chooses \(k = 4\)
      • Step 2: the global minimum is also at \(k = 4\): EFCRIX agrees
      • Step 3: the gain from the 4th coin is \(953.6 - 953.3 = 0.3\) AIC points, against a gain of 284 from the 2nd
        • a flat curve: the choice between 3 and 4 is fragile; this is the reason for the smoothing

      The Number of Constituents, Quarter by Quarter

      Slide 21 in the PDF

      Repeating the rule at each of the 35 quarterly reselections from 31 March 2018 to 30 September 2026 gives the chart. On average 7 of the 10 coins passed the liquidity rule. The first-rise (ECRIX) rule chose two constituents in 9 quarters, three in 13, four in 11 and five or six once each, with a mean of 3.2. The global-minimum (EFCRIX) rule usually agrees.

      The smoothed rule, which pools six months of curves, chose three constituents in 30 of the 35 quarters. The smoothing makes the index more stable: fewer changes of constituents mean fewer trades for a fund that follows it. The small numbers reflect our small universe, dominated by Bitcoin and Ethereum.

      In brief

      The Number of Constituents, Quarter by Quarter
      • 35 quarterly reselections, 31 March 2018 – 30 September 2026; on average 7 of the 10 coins pass the liquidity rule
      • First rise of the AIC: \(k = 2\) in 9 quarters, 3 in 13, 4 in 11, 5 or 6 once each; mean 3.2
      • The smoothed rule chooses \(k = 3\) in 30 of 35 quarters: the smoothing makes the index more stable
      BA_L3_k_choice

      Capped Weights

      Slide 22 in the PDF

      A market-value index of a concentrated market is dominated by its largest coin. A capped market-value index limits the weight of each constituent:

      \[\mathrm{CRIX}(k)_t = \frac{\sum_{i=1}^{k}\mathrm{MV}_{it}\,\mathrm{AW}_{it}}{\mathrm{Divisor}}, \qquad \mathrm{AW}_{it} = \frac{\mathrm{CW}_{it}}{W_{it}}, \qquad W_{it} = \frac{\mathrm{MV}_{it}}{\sum_j \mathrm{MV}_{jt}}\]
      (3.9)

      Here \(W_{it}\) is the market-value weight, \(\mathrm{CW}_{it}\) the capped weight, chosen so that no coin has an influence of 50% or more on the index, and \(\mathrm{AW}_{it}\) the adjustment factor applied to the market value. The capping algorithm is the one used by other capped indices: set any weight above the cap to the cap, share the excess among the other constituents in proportion to their weights, and repeat until no weight exceeds the cap.

      On 30 September 2026 the market-value weights of the replica were Bitcoin 82.1%, Ethereum 16.0%, Zcash 1.2% and Dogecoin 0.7%. Capped just below 50%, they become Bitcoin 49.99%, Ethereum 44.7%, Zcash 3.3% and Dogecoin 2.0%: the excess of 32.1 percentage points goes mostly to Ethereum, the second largest coin.

      In brief

      • A capped market-value index:
        \[ \mathrm{CRIX}(k)_t = \frac{\sum_{i=1}^{k}\mathrm{MV}_{it}\,\mathrm{AW}_{it}}{\mathrm{Divisor}}, \qquad \mathrm{AW}_{it} = \frac{\mathrm{CW}_{it}}{W_{it}}, \qquad W_{it} = \frac{\mathrm{MV}_{it}}{\sum_j \mathrm{MV}_{jt}} \]
        • \(W_{it}\): market-value weight; \(\mathrm{CW}_{it}\): capped weight, so that no coin has an influence of 50% or more
        • \(\mathrm{AW}_{it}\): the adjustment factor applied to the market value
      • Capping algorithm (as in other capped indices): set the weight above the cap to the cap; share the excess among the other constituents in proportion to their weights; repeat
      • Worked example, replica on 30 September 2026
        • market-value weights: BTC 82.1%, ETH 16.0%, ZEC 1.2%, DOGE 0.7%
        • capped below 50%: BTC 49.99%, ETH 44.7%, ZEC 3.3%, DOGE 2.0%; the excess of 32.1 points goes mostly to Ethereum

      Market-Value Weights against Capped Weights

      Slide 23 in the PDF

      The chart shows the weights of the replica at each monthly rebalancing from April 2018 to September 2026. With market-value weights (left), Bitcoin's weight ranged from 59.1% to 89.3%, with a mean of 76.9%. In this universe the 50% cap would therefore bind at every rebalancing.

      The capped index (right) is a different product: about half Bitcoin and a much larger share of Ethereum. Whether a cap is appropriate depends on the purpose of the index: a cap improves diversification and limits the influence of one coin, while the uncapped index stays closer to the market as it is. The next section shows which of the two is closer to the published CRIX on our data.

      In brief

      Market-Value Weights against Capped Weights
      • Weights of the replica at each monthly rebalancing, April 2018 – September 2026
      • Market-value weight of Bitcoin: from 59.1% to 89.3%, mean 76.9%; the 50% cap would bind at every rebalancing in this universe
      • The capped index is a different product: much more Ethereum, less Bitcoin
      BA_L3_replica_weights

      Section summary

      CRIX is a Laspeyres market-value index with a divisor that keeps the level continuous when the basket changes. Its constituents pass a liquidity rule and are ranked by market value; their number is chosen by an information criterion that compares candidate indices with the total market, smoothed over six months. On our ten coins the rule chooses between two and four constituents in most quarters, and the curve is steep for the second coin and nearly flat afterwards.

      Our Replication on Real Data

      The section rebuilds a CRIX-style index from public data with the published rules and compares it with the real CRIX. It then defines tracking error and tracking difference, the two numbers that measure how closely one series follows another.

      Replication Design

      Slide 24 in the PDF

      The universe is the ten coins whose circulating supply is public in the Coin Metrics community data: BTC, ETH, DOGE, ADA, LTC, BCH, ETC, XMR, DASH and ZEC. Prices and volumes in USD are EODHD daily closes, and market values are price times current supply. Coins without a public supply history in this source, such as BNB or Solana (SOL), and coins whose reported supply includes tokens held by the issuer, such as XRP, are not included. The replica therefore covers a smaller universe than the published CRIX.

      The rules follow Trimborn & Härdle (2018) and thecrix.de. At each quarter-end from 31 March 2018 to 30 September 2026 the liquidity rule is applied on the last 90 days, the eligible coins are ranked by market value, and the number of constituents is set by the first rise of the AIC on the last 90 days (the ECRIX rule). Within each quarter the weights are rebalanced monthly with the new supplies, and the divisor keeps the index continuous; the index starts at 1,000 on 31 March 2018.

      Three benchmarks help to read the result: the TMI of the eligible coins, Bitcoin alone, and a version of the replica with weights capped below 50%. All are compared with the published CRIX on common days, 1 April 2018 to 30 October 2026, with \(n = 3{,}105\) daily log returns.

      In brief

      • Universe: ten coins whose circulating supply is public (Coin Metrics Community API): BTC, ETH, DOGE, ADA, LTC, BCH, ETC, XMR, DASH, ZEC
        • prices and volumes in USD: EODHD daily closes; market value = price \(\times\) current supply
        • not included: coins without a public supply history in this source (e.g. BNB, SOL) or whose reported supply includes issuer holdings (XRP)
      • Rules, as in Trimborn & Härdle (2018) and thecrix.de
        • quarterly reselection (31 March 2018 – 30 September 2026): liquidity rule on the last 90 days, ranking by market value
        • number of constituents: first rise of the AIC on the last 90 days (ECRIX)
        • monthly rebalancing with new supplies; divisor; start at 1 000 on 31 March 2018
      • Benchmarks: the TMI of the eligible coins, Bitcoin alone, a version with weights capped below 50%
      • Comparison with the published CRIX on common days, 1 April 2018 – 30 October 2026, \(n = 3\,105\) daily log returns

      The Replica against the Real CRIX

      Slide 25 in the PDF
        Question for the room Change the number of constituents: the AIC rules, the TMI, or a fixed top 1 to 7. Which choice tracks the published CRIX best, and by how much?
        Answer

        All choices with 3 to 7 coins are close: tracking error 9.0% (top 5) to 9.4% (top 3) a year, correlation 0.989–0.990; the AIC rules give 9.3%. Bitcoin alone (top 1) has 15.6% and correlation 0.971; the top 2 has 10.0%. Capping the weights below 50% raises the tracking error of the ECRIX replica from 9.3% to 14.5%. Beyond the first two or three coins the remaining gap comes from coins outside our ten-coin universe, not from the rule.

        Both series start at 1,000 on 31 March 2018. On 30 October 2026 the published CRIX stands at 8,380 on this scale, the replica at 8,648 and the capped replica at 8,153. The three lines are hardly distinguishable on a log scale over eight and a half years that include two crypto winters.

        The right panel shows the correlation of daily log returns on rolling 180-day windows. It stays between 0.976, its lowest value in December 2021, and 0.998. A handful of constituents from ten coins, chosen with the published rules, reproduces most of the daily movement of the published index.

        In brief

        • Both start at 1 000 on 31 March 2018; on 30 October 2026: CRIX 8 380, replica 8 648, capped replica 8 153
        • 180-day correlation of daily log returns between 0.976 (December 2021) and 0.998
        • Ten coins and a handful of constituents reproduce most of the published index
        BA_L3_replica_vs_crix

        Accuracy of the Replica

        Slide 26 in the PDF

        The table compares four series with the published CRIX. The replica has a correlation of 0.990 with CRIX, a tracking error of 9.3% a year and a beta of 1.00. The TMI of the eligible coins is almost identical (0.990, 9.1%). The capped replica is further away (0.979, 14.5%, beta 0.91), and Bitcoin alone further still (0.971, 15.6%).

        Three readings follow. First, the replica tracks its own TMI very closely, with a correlation of 0.9995 and a tracking error of 2.2% a year: the rule does what it is designed to do, representing the market with few coins. Second, the remaining gap between the replica and CRIX comes from the universe, the coins we do not have, not from the rule. Third, on these data the uncapped replica is closer to the published CRIX than the capped one. And a basket of a few coins tracks the market better than Bitcoin alone, even though Bitcoin holds most of its weight.

        In brief

        Series compared with the published CRIXcorrelationtracking error (% a year)beta
        Replica (ECRIX rule, market-value weights)\(0.990\)\(9.3\)\(1.00\)
        TMI of the eligible coins\(0.990\)\(9.1\)\(0.99\)
        Replica, weights capped below 50%\(0.979\)\(14.5\)\(0.91\)
        Bitcoin alone\(0.971\)\(15.6\)\(1.02\)
        • Daily log returns, 1 April 2018 – 30 October 2026, \(n = 3\,105\); beta: slope of CRIX returns on the series' returns
        • Interpretation
          • the replica tracks its own TMI closely: correlation 0.9995, tracking error 2.2% a year
          • the remaining gap to CRIX comes from the universe (coins we do not have), not from the rule
          • on these data the uncapped replica is closer to the published CRIX than the capped one
          • a basket of a few coins tracks the market better than Bitcoin alone

        Tracking Error and Tracking Difference

        Slide 27 in the PDF

        Two numbers describe how a portfolio, a replica or a fund, follows its benchmark. With daily log returns \(r^{P}_t\) of the portfolio and \(r^{B}_t\) of the benchmark, and \(A\) returns per year,

        \[\mathrm{TE} = \sqrt{A}\;\mathrm{sd}\big(r^{P}_t - r^{B}_t\big), \qquad \mathrm{TD} = \frac{A}{n}\sum_{t=1}^{n}\big(r^{P}_t - r^{B}_t\big)\]
        (3.10)

        The tracking error (TE) measures how much the portfolio wanders around the benchmark; the tracking difference (TD) measures by how much it lags or leads on average. \(A\) is 365 for coins, which trade every day, and 252 for exchange-traded funds.

        For the replica against CRIX, \(\mathrm{TE} = 9.3\%\) a year. On a daily basis this is \(9.3/\sqrt{365} = 0.49\) percentage points: on a typical day on which CRIX moves by 3%, the replica moves by about 2.5% to 3.5%. A good index fund has a small TE, because it follows the index closely, and a TD close to minus its fees, because fees are the one steady drag it cannot avoid.

        In brief

        • Tracking error (TE) and tracking difference (TD) of a portfolio (replica or fund) with daily log returns \(r^{P}_t\) against its benchmark \(r^{B}_t\):
          \[ \mathrm{TE} = \sqrt{A}\;\mathrm{sd}\big(r^{P}_t - r^{B}_t\big), \qquad \mathrm{TD} = \frac{A}{n}\sum_{t=1}^{n}\big(r^{P}_t - r^{B}_t\big) \]
          • TE: how much the portfolio wanders around the benchmark; TD: by how much it lags or leads on average
          • \(A\): returns per year (365 for coins, 252 for exchange-traded funds)
        • Worked example: replica against CRIX, TE \(= 9.3\%\) a year
          • daily: \(9.3/\sqrt{365} = 0.49\) percentage points
          • a typical day: if CRIX moves by 3%, the replica moves by about 2.5% to 3.5%
        • A good index fund has a small TE (it follows the index) and a TD close to minus its fees

        Section summary

        A CRIX-style index built on ten coins with the published rules tracks the published CRIX with a correlation of 0.990 and a tracking error of 9.3% a year, much closer than Bitcoin alone. The replica follows its own total market almost perfectly, so the remaining gap reflects the universe of coins rather than the method.

        The Real CRIX, 2017–2026

        The section turns to the published index itself: its path from 2017 to 2026, its growth, volatility and drawdowns, the crisis episodes, and a comparison with Bitcoin year by year.

        CRIX, 2017–2026

        Slide 28 in the PDF
          Question for the room Choose the period 1 January 2023 – 31 October 2026. Did the broad index CRIX beat Bitcoin?
          Answer

          No: CAGR 40.4% for CRIX against 49.3% for Bitcoin, with volatility 47.4% against 46.1% and maximum drawdowns of −56.1% against −53.1%. Over the official history (16 March 2018 – 31 October 2026) the picture is the same: 24.0% against 29.4% a year. In this period Bitcoin's share of the market rose, so a market-value index of many coins grew more slowly and was not less volatile.

          The chart shows the daily values of CRIX from thecrix.de, 6 September 2017 to 6 November 2026, with Bitcoin scaled to CRIX on 16 March 2018. No values are published on 22 March 2026 and from 14 May to 7 June 2026.

          The index fell to its low of 582.01 on 15 December 2018, at the end of the first crypto winter, reached its record of 21,523.34 on 6 October 2025, and stood at 12,250.34 on 6 November 2026, 43.1% below the record. Before 16 March 2018 the series comes from a different data source and moves one day ahead of the Bitcoin close of EODHD; the statistics that follow therefore start on 16 March 2018, the start of the official history.

          In brief

          • Daily values from thecrix.de, 6 September 2017 – 6 November 2026; no values on 22 March 2026 and from 14 May to 7 June 2026
          • Low 582.01 on 15 December 2018; record 21 523.34 on 6 October 2025; last value 12 250.34 (6 November 2026), 43.1% below the record
          • Before 16 March 2018 the series comes from a different data source and is shifted by one day against Bitcoin; statistics below start on 16 March 2018
          BA_L3_crix_path

          CRIX and Bitcoin in Numbers

          Slide 29 in the PDF

          From 16 March 2018 to 31 October 2026 CRIX grew at a compound annual rate of 24.0%, with an annualised volatility of daily log returns of 65.0% and a maximum drawdown of \(-77.4\%\). Bitcoin grew at 29.4% a year with a volatility of 61.7% and a maximum drawdown of \(-76.6\%\). The daily returns of CRIX are skewed to the left (\(-1.09\)) and heavy-tailed (excess kurtosis 14.4), as found for single coins in Lecture 2; its worst day was 12 March 2020 (\(-46.5\%\)) and its best 2 April 2019 (\(+16.8\%\)).

          A regression of CRIX returns on Bitcoin returns over 3,122 common days gives a beta of 1.020 (standard error 0.005), an \(R^2\) of 0.94 and a correlation of 0.971. Two conclusions follow. CRIX is slightly more volatile than Bitcoin: the other coins add volatility rather than diversification, because they move with Bitcoin and move more. And over this period CRIX grew less than Bitcoin, 24.0% against 29.4% a year, because Bitcoin's share of the market rose.

          In brief

          16 Mar 2018 – 31 Oct 2026CAGR (%)ann. vol. (%)max. drawdown (%)skewnessexc. kurtosisworst day (%)
          CRIX\(24.0\)\(65.0\)\(-77.4\)\(-1.09\)\(14.4\)\(-46.5\)
          Bitcoin\(29.4\)\(61.7\)\(-76.6\)–––
          • Daily log returns, \(A = 365\); CRIX: \(n = 3\,121\) returns between consecutive days; worst day 12 March 2020, best day 2 April 2019 (\(+16.8\%\))
          • Regression of CRIX returns on Bitcoin returns (\(n = 3\,122\) common days): beta \(1.020\) (standard error, SE, 0.005), \(R^2 = 0.94\), correlation 0.971
          • Interpretation
            • CRIX is slightly more volatile than Bitcoin: the other coins add volatility, not diversification
            • over this period CRIX grew less than Bitcoin (24.0% against 29.4% a year): Bitcoin's share of the market rose

          Worked Example: CAGR and the Last Drawdown

          Slide 30 in the PDF

          The compound annual growth rate (CAGR) is the constant yearly growth that turns the first value into the last: \((I_T/I_0)^{1/T} - 1\), with \(T\) in years. For CRIX, \(I_0 = 1{,}960.12\) on 16 March 2018 and \(I_T = 12{,}512.61\) on 31 October 2026, \(T = 8.63\) years; the ratio is 6.384 and \(6.384^{1/8.63} - 1 = 24.0\%\) a year. Bitcoin, on the same dates, went from 8,338 to 77,214 USD, \((77{,}214/8{,}338)^{1/8.63} - 1 = 29.4\%\) a year.

          The drawdown \(D_t = I_t/\max_{s \le t} I_s - 1\) measures the loss from the last record. CRIX set its record of 21,523.34 on 6 October 2025 and fell to 9,439.22 on 30 June 2026: \(D = 9{,}439.22/21{,}523.34 - 1 = -56.1\%\). To recover, the index must gain \(21{,}523.34/9{,}439.22 - 1 = 128\%\). On 6 November 2026, at 12,250.34, it was still 43.1% below the record.

          In brief

          • CAGR: \((I_T/I_0)^{1/T} - 1\) with \(T\) in years
            • CRIX: \(I_0 = 1\,960.12\) (16 March 2018), \(I_T = 12\,512.61\) (31 October 2026), \(T = 8.63\)
            • \((12\,512.61/1\,960.12)^{1/8.63} - 1 = 6.384^{0.116} - 1 = 24.0\%\) a year
            • Bitcoin, same dates: \((77\,214/8\,338)^{1/8.63} - 1 = 29.4\%\) a year
          • Drawdown \(D_t = I_t/\max_{s \le t} I_s - 1\)
            • record 21 523.34 on 6 October 2025; trough 9 439.22 on 30 June 2026: \(D = 9\,439.22/21\,523.34 - 1 = -56.1\%\)
            • gain needed to recover: \(21\,523.34/9\,439.22 - 1 = 128\%\)
            • 6 November 2026: 12 250.34, still 43.1% below the record

          Drawdowns of CRIX and Bitcoin

          Slide 31 in the PDF
            Question for the room At 50%, CRIX spent 48.8% of the days more than 50% below its record, Bitcoin 39.9%. What are the shares at 70%?
            Answer

            17.5% of the days for CRIX and 9.5% for Bitcoin. CRIX has two falls deeper than 70%: 2018 (−87.1%, record again on 31 December 2020) and 2021–2022 (−77.4%, record again on 11 November 2024). The broad index stays deeper under water for longer than Bitcoin, because the smaller coins fall more in crypto winters.

            The drawdown curves of CRIX and Bitcoin almost coincide: the two fall and recover together. CRIX falls a little deeper in each crash, consistent with its beta above one. Since September 2017 CRIX was more than 50% below its record on 48.8% of the days, Bitcoin on 39.9%. For an investor in a broad crypto index, being far below the last peak is the normal state, not the exception.

            In brief

            • The two drawdown curves almost coincide; CRIX falls a little deeper in each crash
            • Since September 2017 CRIX was more than 50% below its record on 48.8% of the days, Bitcoin on 39.9%
            BA_L3_crix_drawdown

            Episodes: Crypto Winters and Shocks

            Slide 32 in the PDF

            The table lists the falls of CRIX deeper than 50%. The first, from 4,510.28 on 5 January 2018 to 582.01 on 15 December 2018, cost 87.1% and took 1,091 days from peak to the next record; its peak lies before the official history. In 2021 the index lost 52.6% between May and July and recovered within 163 days. The second long winter, from 14,484.51 on 8 November 2021 to 3,268.34 on 9 November 2022, cost 77.4% and lasted 1,099 days from peak to record. The latest fall, from the record of 6 October 2025 to 30 June 2026, cost 56.1% and has not yet been recovered.

            Three shocks, measured as cumulative log returns over the same windows as in Lecture 2, show how the broad index reacts to different events. In the COVID-19 crash (19 February to 23 March 2020) CRIX lost 45.0 log points and Bitcoin 40.6. In the collapse of the stablecoin TerraUSD and its sister token Luna (5 to 12 May 2022) CRIX lost 31.0 and Bitcoin 23.0; in the bankruptcy of the exchange FTX (5 to 21 November 2022), 33.8 against 29.9. In crypto-specific shocks the broad index loses more than Bitcoin: smaller coins fall harder when confidence in the crypto ecosystem itself is shaken.

            In brief

            Peakleveltroughdepth (%)new recorddays peak to record
            5 Jan 20184 510.2815 Dec 2018 (582.01)\(-87.1\)31 Dec 20201 091
            9 May 202112 644.7020 Jul 2021 (5 994.23)\(-52.6\)19 Oct 2021163
            8 Nov 202114 484.519 Nov 2022 (3 268.34)\(-77.4\)11 Nov 20241 099
            6 Oct 202521 523.3430 Jun 2026 (9 439.22)\(-56.1\)not yet–
            • Falls of CRIX deeper than 50%; the 2018 peak lies before the official history
            • Three shocks, cumulative log returns (same windows as Lecture 2)
              • COVID-19, 19 February – 23 March 2020: CRIX \(-45.0\), Bitcoin \(-40.6\)
              • Terra/Luna, 5 – 12 May 2022: CRIX \(-31.0\), Bitcoin \(-23.0\)
              • FTX, 5 – 21 November 2022: CRIX \(-33.8\), Bitcoin \(-29.9\)
            • In crypto-specific shocks the broad index loses more than Bitcoin: smaller coins fall harder

            Year by Year: CRIX against Bitcoin

            Slide 33 in the PDF

            Calendar-year returns make the comparison concrete (2018 from 16 March, 2026 to 31 October). CRIX beat Bitcoin only in 2021, with \(+125.7\%\) against \(+59.7\%\), the year of the large rally of alternative coins. In the falling years 2018, 2022, 2025 and 2026 CRIX lost more than Bitcoin, and in the rising years 2019, 2023 and 2024 it gained less.

            For a Bitcoin investor, then, a broad crypto index has not offered protection: it is more volatile, falls more in bad years and has lagged in most good years. Its value lies elsewhere: it measures the market as a whole, and in a year like 2021 the market as a whole was very different from Bitcoin.

            In brief

            Year by Year: CRIX against Bitcoin
            • Simple returns per calendar year; *2018 from 16 March, 2026 to 31 October
            • CRIX beat Bitcoin only in 2021 (\(+125.7\%\) against \(+59.7\%\)), the year of the large altcoin rally
            • In falling years (2018, 2022, 2025, 2026) CRIX lost more than Bitcoin
            BA_L3_crix_years

            Section summary

            From 16 March 2018 to 31 October 2026 CRIX grew by 24.0% a year with a volatility of 65.0% and a maximum drawdown of 77.4%; its beta on Bitcoin is 1.02 and its correlation 0.971. It reached its record of 21,523.34 on 6 October 2025 and lost 56.1% by 30 June 2026. CRIX falls more than Bitcoin in crypto-specific shocks and outperformed it in only one calendar year, 2021.

            VCRIX: A Volatility Index

            Volatility is the central risk number of the crypto market. The section explains why a volatility index is useful, how VCRIX is built, what its 30-day window implies, how it reacts to falls and rises, and how it compares with a GARCH model and with our own forecast of next month's volatility.

            Motivation for a Volatility Index

            Slide 34 in the PDF

            Equity markets have a "fear gauge": the Cboe Volatility Index (VIX), the volatility over the next 30 days implied by the prices of S&P 500 options (Whaley, 2000). When investors fear large moves, they pay more for options, implied volatility rises and the VIX rises with it. For the crypto market as a whole there was no liquid option market on an index, and therefore no implied volatility to read.

            Kim, Trimborn & Härdle (2021) built VCRIX to fill this gap: a forecast of the volatility of CRIX that plays the role of implied volatility. As a proxy for implied volatility they use the realised volatility of CRIX over a 30-day window, and they choose the forecasting model by its out-of-sample accuracy among GARCH models, the heterogeneous autoregressive (HAR) model and a neural network; the HAR model performed best. VCRIX serves as a daily risk thermometer for the whole market and as an input for risk limits, option pricing and volatility-timing strategies.

            In brief

            • Equity markets have a “fear gauge”: the Cboe Volatility Index (VIX), the 30-day volatility implied by S&P 500 option prices Whaley (2000)
            • For the crypto market as a whole there was no liquid option market on an index, hence no implied volatility
            • Kim, Trimborn & Härdle (2021): VCRIX, a forecast of CRIX volatility that plays the role of implied volatility
              • proxy for implied volatility: the realised volatility of CRIX over a 30-day window
              • forecasting model chosen by out-of-sample accuracy among GARCH models, HAR and a neural network; HAR performed best
            • Uses
              • a daily risk thermometer for the whole market
              • an input for risk limits, option pricing and volatility-timing strategies

            VCRIX: Construction

            Slide 35 in the PDF

            The building block is the realised volatility (RV) of CRIX over the last \(w\) days, annualised, with \(r_t\) the daily log return in percent:

            \[\mathrm{RV}^{(w)}_t = \sqrt{\frac{365}{w}\sum_{j=0}^{w-1} r_{t-j}^2}\]
            (3.11)

            The HAR model of Corsi (2009) forecasts volatility from its own recent values at three horizons, daily, weekly and monthly:

            \[\mathrm{RV}_{t+1} = c + \beta_d\,\mathrm{RV}^{(1)}_t + \beta_w\,\mathrm{RV}^{(7)}_t + \beta_m\,\mathrm{RV}^{(30)}_t + u_{t+1}\]
            (3.12)

            The three components stand for three kinds of traders: day traders who react to yesterday's move, and weekly and monthly investors who react to longer averages. The model is estimated by ordinary least squares (OLS), and the forecast \(\widehat{\mathrm{RV}}_{t+1}\) is the volatility expected for tomorrow. The index divides the forecast by a divisor, \(\mathrm{VCRIX}_t = \widehat{\mathrm{RV}}_t/\mathrm{Divisor}\), set so that VCRIX equals 1,000 on its start date; the published series starts at 1,000 on 4 January 2018.

            In brief

            • Realised volatility (RV) of CRIX over the last \(w = 30\) days, annualised, with \(r_t\) the daily log return in %:
              \[ \mathrm{RV}^{(w)}_t = \sqrt{\frac{365}{w}\sum_{j=0}^{w-1} r_{t-j}^2} \]
            • Heterogeneous autoregressive (HAR) model Corsi (2009): daily, weekly and monthly components forecast the next value
              \[ \mathrm{RV}_{t+1} = c + \beta_d\,\mathrm{RV}^{(1)}_t + \beta_w\,\mathrm{RV}^{(7)}_t + \beta_m\,\mathrm{RV}^{(30)}_t + u_{t+1} \]
              • three horizons of traders: day traders, weekly and monthly investors
              • estimated by ordinary least squares (OLS); the forecast \(\widehat{\mathrm{RV}}_{t+1}\) is the volatility expected for tomorrow
            • The index: \(\mathrm{VCRIX}_t = \widehat{\mathrm{RV}}_t/\mathrm{Divisor}\), with the divisor set so that VCRIX \(= 1\,000\) on the start date (4 January 2018 in the published series)

            VCRIX, 2018–2026

            Slide 36 in the PDF
              Question for the room The ▶ button changes the window of the realised volatility from 7 to 90 days. For which window does VCRIX correlate most with the realised volatility of the previous day?
              Answer

              At exactly 30 days: correlation 0.943 (21 days 0.873, 45 days 0.839, 90 days 0.657); a scan of every window from 5 to 120 days peaks at 30. This is consistent with the construction of VCRIX on a 30-day realised volatility (Kim, Trimborn and Härdle, 2021). With the same-day volatility the correlation is lower, 0.930: VCRIX on day t reflects the returns up to day t − 1.

              The chart shows VCRIX from thecrix.de together with the 30-day realised volatility of CRIX computed from the published index. The two move together: the correlation of VCRIX with the realised volatility of the previous day is 0.943 over \(n = 3{,}020\) days.

              VCRIX averaged 458 (median 427). Its maximum, 1,215.57, came on 13 March 2020, the day after the COVID-19 crash; its minimum, 105.09, on 14 August 2023, in the calm market of that summer; on 6 November 2026 it stood at 342.13. A regression of VCRIX on the previous day's realised volatility gives VCRIX \(\approx 74.4 + 6.11\,\mathrm{RV}^{(30)}_{t-1}\) with \(R^2 = 0.89\): a realised volatility of 50% a year corresponds to a VCRIX of about 380. The shaded period in 2026 marks the gap in the published CRIX and the 30 days after it, discussed next.

              In brief

              • VCRIX (thecrix.de) and the 30-day realised volatility of CRIX: correlation 0.943 with the realised volatility of the previous day (\(n = 3\,020\))
              • Mean 458, median 427; maximum 1 215.57 on 13 March 2020, minimum 105.09 on 14 August 2023; 342.13 on 6 November 2026
              • VCRIX \(\approx 74.4 + 6.11\,\mathrm{RV}^{(30)}_{t-1}\) (\(R^2 = 0.89\)): 50% realised volatility corresponds to a VCRIX of about 380
              BA_L3_vcrix_path

              Worked Example: the 30-Day Window

              Slide 37 in the PDF

              On 12 March 2020 CRIX fell by 46.5% in log terms. VCRIX was 405.66 on 12 March and 1,215.57 on 13 March: the crash entered the window one day later. One day can move the index this much because squared returns enter the sum: the crash adds \(46.5^2/30 = 72\) to the mean square of the 30-day window, which alone corresponds to \(\sqrt{365 \times 72} = 162\%\) of annual volatility.

              Thirty days later the opposite happened. VCRIX was 1,159.78 on 11 April and 542.82 on 12 April 2020, although nothing special happened in the market that day: the crash day left the window. A rolling window produces such a "ghost" effect, a jump in and a jump out exactly one window apart.

              The published series shows the same mechanism at a data gap. CRIX has no values from 14 May to 7 June 2026; a return over this 25-day gap enters a 30-day window as if it were one day. VCRIX jumped from 213.04 to 673.24 on 10 June and fell from 793.12 to 283.93 on 9 July 2026. This is a data artefact, and the window is excluded from our statistics. The general lesson is that a rolling window treats a shock as fully relevant for \(w\) days and then irrelevant, while a GARCH model lets old shocks fade gradually.

              In brief

              • 12 March 2020: CRIX log return \(-46.5\%\)
                • VCRIX 405.66 on 12 March, 1 215.57 on 13 March: the crash enters the window one day later
                • one day enters the 30-day sum: \(46.5^2/30 = 72\) added to the mean square, i.e. \(\sqrt{365 \times 72} = 162\%\) of annual volatility from one day
              • 30 days later: VCRIX 1 159.78 on 11 April, 542.82 on 12 April 2020
                • the crash leaves the window; nothing happened in the market that day
              • Data gap of 14 May – 7 June 2026: VCRIX jumps from 213.04 to 673.24 on 10 June and falls from 793.12 to 283.93 on 9 July
                • a return over a 25-day gap inside a 30-day window: a data artefact; this window is excluded from our statistics
              • Lesson: a rolling window gives a “ghost” effect; a GARCH model lets old shocks fade gradually

              Volatility after Falls and Rises

              Slide 38 in the PDF
                Question for the room Raise the threshold of a large move to 10%. What happens to VCRIX the day after a fall of more than 10%, and after a rise of more than 10%?
                Answer

                After the 35 falls of more than 10% the log of VCRIX rises by 21.6% on average, after the 19 rises of more than 10% by 18.3%; on the other days it falls slightly (−0.4%). Large moves in either direction raise the volatility index; the effect after falls is somewhat larger, as the leverage effect suggests, but rises also raise it, because VCRIX is built from squared returns.

                In equity markets volatility rises more after falls than after rises of the same size, the so-called leverage effect. The chart checks this for crypto: the daily change of \(\ln\) VCRIX on day \(t\) against the CRIX return on day \(t-1\), from 16 March 2018 to 13 May 2026, \(n = 2{,}978\) days. The correlation is \(-0.132\); with the same day's return it is 0.007, because VCRIX reacts one day later.

                A regression with separate slopes for falls and rises, with heteroskedasticity- and autocorrelation-consistent (HAC) standard errors, gives a slope of \(-1.61\) (SE 0.14) after falls and \(+1.38\) (SE 0.12) after rises. Both are significant, and their sizes are close: a test of equal size is not rejected (p-value 0.10). The right panel shows the same pattern in groups: after a fall of more than 5% VCRIX rises by 8.6% on average, after a rise of more than 5% by 6.9%, and after quiet days it drifts down. In crypto, large moves in either direction raise expected volatility; the asymmetry towards falls is small on these data.

                In brief

                • Change of \(\ln\) VCRIX on day \(t\) against the CRIX return on day \(t-1\), 16 March 2018 – 13 May 2026, \(n = 2\,978\); correlation \(-0.132\) (same day: 0.007)
                • Regression with heteroskedasticity- and autocorrelation-consistent (HAC) SE: slope \(-1.61\) (SE 0.14) after falls, \(+1.38\) (SE 0.12) after rises; equal size not rejected (p 0.10)
                • Large moves in either direction raise VCRIX: after a fall beyond 5% by 8.6% on average, after a rise beyond 5% by 6.9%
                BA_L3_vcrix_leverage

                A GARCH(1,1) Volatility for CRIX

                Slide 39 in the PDF

                A model-based alternative to the rolling window is the GARCH(1,1) model of Bollerslev (1986). With Student-\(t\) errors for the daily log returns \(r_t\) (in %) of CRIX,

                \[r_t = \mu + \sigma_t z_t, \quad z_t \sim t_\nu, \qquad \sigma_t^2 = \omega + \alpha\,(r_{t-1} - \mu)^2 + \beta\,\sigma_{t-1}^2\]
                (3.13)

                today's variance is a weighted combination of a constant, yesterday's squared shock and yesterday's variance; \(z_t\) has unit variance and \(\nu\) degrees of freedom.

                Estimated on 16 March 2018 to 13 May 2026 (\(n = 2{,}979\) returns; the two returns around the missing 22 March 2026 are skipped), the model gives \(\hat\mu = 0.128\), \(\hat\omega = 0.254\), \(\hat\alpha = 0.084\), \(\hat\beta = 0.914\) and \(\hat\nu = 3.14\), with robust standard errors in the table. The persistence \(\alpha + \beta = 0.998\) is very high: a shock to variance halves only after \(\ln 0.5/\ln 0.998 \approx 370\) days. The estimate \(\hat\nu = 3.1\) signals very heavy tails, as found in Lecture 2. The annualised GARCH volatility \(\sqrt{365}\,\hat\sigma_{t+1}\) averaged 69.8% and peaked at 269% on 13 March 2020. Lecture 4 develops GARCH models, value at risk and expected shortfall in detail.

                In brief

                • GARCH(1,1) Bollerslev (1986) with Student-\(t\) errors for the daily log returns \(r_t\) (%) of CRIX:
                  \[ r_t = \mu + \sigma_t z_t, \quad z_t \sim t_\nu \text{ (unit variance)}, \qquad \sigma_t^2 = \omega + \alpha\,(r_{t-1} - \mu)^2 + \beta\,\sigma_{t-1}^2 \]
                \(\mu\)\(\omega\)\(\alpha\)\(\beta\)\(\nu\)
                estimate\(0.128\)\(0.254\)\(0.084\)\(0.914\)\(3.14\)
                SE (robust)\(0.042\)\(0.107\)\(0.020\)\(0.019\)\(0.20\)
                • 16 March 2018 – 13 May 2026, \(n = 2\,979\) returns (the two returns around the missing 22 March 2026 skipped)
                • Persistence \(\alpha + \beta = 0.998\): a shock to variance halves in \(\ln 0.5/\ln 0.998 \approx 370\) days
                • \(\hat\nu = 3.1\): very heavy tails, as in Lecture 2
                • Annualised GARCH volatility \(\sqrt{365}\,\hat\sigma_{t+1}\): mean 69.8%, maximum 269% on 13 March 2020
                • Lecture 4 develops GARCH, value at risk and expected shortfall

                VCRIX against GARCH and Realised Volatility

                Slide 40 in the PDF
                  Question for the room The ▶ button shifts the GARCH and realised volatility series against VCRIX by k = −10 to +10 days. At which shift is the correlation of VCRIX with the realised volatility highest?
                  Answer

                  At k = −1: 0.931, against 0.916 at k = 0, i.e. VCRIX on day t matches the 30-day realised volatility of day t − 1. For the GARCH one-day-ahead volatility the maximum is at k = 0 (0.844), since it is already a forecast made on day t − 1. For positive k the correlations fall (GARCH 0.53 at k = +10): VCRIX does not anticipate future volatility better than the two models.

                  The chart puts the three volatility measures side by side over \(n = 2{,}949\) days. VCRIX correlates 0.916 with the 30-day realised volatility and 0.844 with the one-day-ahead GARCH forecast; GARCH and realised volatility correlate 0.908.

                  All three agree on the volatility regimes: the turbulent 2018, March 2020, May 2021 and November 2022. They differ in their dynamics. VCRIX follows the 30-day window, with its jumps in and out; the GARCH volatility reacts at once to a large shock and then fades smoothly, at the speed set by its persistence. For a risk manager the GARCH reading is the more timely one after a shock, the VCRIX reading the more directly comparable with the realised risk of the last month.

                  In brief

                  • Correlations, \(n = 2\,949\) days: VCRIX with the 30-day realised volatility 0.916, with the one-day-ahead GARCH forecast 0.844; GARCH with realised volatility 0.908
                  • The three measures agree on the regimes (2018, March 2020, May 2021, November 2022); VCRIX follows the 30-day window, GARCH reacts faster and fades smoothly
                  BA_L3_vcrix_garch

                  Our HAR Forecast of the Next Month's Volatility

                  Slide 41 in the PDF

                  As an exercise we estimate a HAR model of our own, with a harder target: the realised volatility of CRIX over the next 30 days, from the daily, weekly and monthly components known today. The model is estimated on 16 March 2018 to 31 December 2023 (\(n = 2{,}117\) days) and tested, without re-estimation, from 2024 to February 2026 (\(n = 781\) days); training and test periods are strictly chronological.

                  The estimates, with HAC standard errors, are \(c = 48.0\) (6.5), \(\beta_d = 0.021\) (0.007), \(\beta_w = 0.112\) (0.042) and \(\beta_m = 0.153\) (0.089); the in-sample \(R^2\) is only 0.09. In the test period the root mean squared error of the forecast is 18.4 percentage points of annual volatility, against 18.6 for the naive forecast "next month equals last month"; the forecast correlates 0.73 with VCRIX. A month ahead, crypto volatility is hard to forecast: the HAR forecast stays close to the long-run mean and improves only slightly on the naive rule.

                  In brief

                  Our HAR Forecast of the Next Month's Volatility
                  • Target: realised volatility of CRIX over the next 30 days; training 16 March 2018 – 31 December 2023 (\(n = 2\,117\)), test 2024 – February 2026 (\(n = 781\))
                  • Estimates (HAC SE): \(c = 48.0\) (6.5), \(\beta_d = 0.021\) (0.007), \(\beta_w = 0.112\) (0.042), \(\beta_m = 0.153\) (0.089); in-sample \(R^2 = 0.09\)
                  • Test period: root mean squared error 18.4 points, against 18.6 for “next month = last month”; the forecast correlates 0.73 with VCRIX
                  • A month ahead, crypto volatility is hard to forecast: the forecast stays close to the long-run mean
                  BA_L3_vcrix_har

                  Section summary

                  VCRIX is a HAR-based forecast of the 30-day realised volatility of CRIX, normalised to 1,000 on its first day. It correlates 0.94 with the realised volatility, peaked at 1,215.57 after the crash of March 2020, and shows the jumps in and out that a 30-day window produces. Large moves in either direction raise it, and a GARCH(1,1) model with persistence 0.998 tells the same story about regimes with smoother dynamics.

                  Investible Index Products

                  An index is a number; investors buy funds. The section traces how CRIX went from research to a commercial index, explains how a fund follows an index and why it deviates, and measures the tracking of a spot Bitcoin fund on real data.

                  Academic and Investible Indices

                  Slide 42 in the PDF

                  CRIX began as academic research. In 2015 and 2016 its first versions appeared as discussion papers at Humboldt-Universität zu Berlin, and in 2018 Trimborn & Härdle (2018) published the method in the Journal of Empirical Finance. The official history of the index starts on 16 March 2018 (Royalton Partners: CRIX).

                  In 2021 Royalton Partners AG acquired the rights to the index. From November 2021 the index has been run commercially as the Royalton CRIX Crypto Index. The two public sources describe the calculation agent differently: Royalton Partners: CRIX states that the index has been calculated by Royalton Partners since November 2021, while thecrix.de states that it has been calculated by S&P Global since November 2021. The stated target is to track the TMI "with as few as possible but more than 5 liquid cryptocurrencies", with stablecoins excluded (Royalton Partners: CRIX). In 2021 Kim, Trimborn & Härdle (2021) added VCRIX, the volatility index of CRIX.

                  In brief

                  • 2015–2016: CRIX discussion papers, Humboldt-Universität zu Berlin
                  • 2018: Trimborn & Härdle (2018) publish the method; official history of the index starts on 16 March 2018 Royalton Partners: CRIX
                  • 2021: Royalton Partners AG acquires the rights to the index
                  • From November 2021: the index is run commercially as the Royalton CRIX Crypto Index
                  • 2021: Kim, Trimborn & Härdle (2021) add VCRIX, the volatility index

                  Index Funds and Exchange-Traded Products

                  Slide 43 in the PDF

                  An index provider publishes the constituents and their weights. The fund issues shares, takes investors' money and buys the constituents in the index weights. When investors buy or redeem shares, the fund buys or sells the coins.

                  An exchange-traded product (ETP) is a fund or note whose shares trade on a stock exchange. On 10 January 2024 the US Securities and Exchange Commission (SEC) approved spot Bitcoin exchange-traded funds in the US (SEC, 2024); a spot fund holds the coins themselves, not futures contracts on them.

                  A fund deviates from its index for several reasons. Fees are a steady drag and produce the tracking difference. Cash held for redemptions, trading costs, the timing of rebalancing and different price sources produce noise around the index, the tracking error. And small constituents may be illiquid, so that buying them moves their prices: this is why investible indices use a liquidity rule and a small number of constituents.

                  In brief

                  • An index is a number; a fund holds the coins
                    • the index provider publishes constituents and weights
                    • the fund issues shares, takes investors' money and buys the constituents in the index weights
                  • Exchange-traded product (ETP): a fund or note whose shares trade on a stock exchange
                    • 10 January 2024: the US Securities and Exchange Commission (SEC) approved spot Bitcoin ETFs in the US SEC (2024)
                    • a spot ETF holds the coins themselves, not futures
                  • Why a fund deviates from its index
                    • fees (a steady drag: tracking difference)
                    • cash, trading costs, timing of rebalancing, different price sources (noise: tracking error)
                    • small constituents may be illiquid: hence the liquidity rule and a small number of constituents

                  A Spot Bitcoin Fund against Bitcoin

                  Slide 44 in the PDF
                    Question for the room Move the start of the period to July 2025 and switch the benchmark clock. What happens to the tracking error and to the tracking difference?
                    Answer

                    With Bitcoin at 16:00 New York the tracking error falls to 1.7% a year (from 3.5% since January 2024), with Bitcoin at 24:00 UTC it is still 16.1%: most of the latter is the time gap between the two prices, not the fund. The tracking difference with synchronous prices stays near −0.26% a year, close to the sponsor fee of 0.25%; with midnight prices it is a noisy −0.06%.

                    The iShares Bitcoin Trust (IBIT) is a spot Bitcoin fund that has traded since January 2024. The chart compares its adjusted closing price (EODHD) with the price of Bitcoin at 16:00 New York time, the time at which the fund is priced, taken from EODHD hourly data, from 12 January 2024 to 30 October 2026, \(n = 697\) trading days.

                    The left panel shows the ratio of IBIT to Bitcoin, in effect the amount of Bitcoin behind one share. It fell by 1.1% over the period, along a trend of \(-0.28\%\) a year: this is the tracking difference, close to the sponsor fee of 0.25% a year (iShares: IBIT). The right panel shows the daily differences of returns. With prices taken at the same time, the tracking error is 3.5% a year; with Bitcoin's close at midnight, Coordinated Universal Time (UTC), it is 20.6%.

                    In brief

                    • IBIT (EODHD, adjusted close) against Bitcoin at 16:00 New York time (EODHD hourly), 12 January 2024 – 30 October 2026, \(n = 697\) trading days
                    • Left: IBIT per Bitcoin fell by 1.1%; trend \(-0.28\%\) a year, close to the sponsor fee of 0.25% iShares: IBIT
                    • Right: daily return differences; TE 3.5% a year with synchronous prices, 20.6% with Bitcoin's close at midnight, Coordinated Universal Time (UTC)
                    BA_L3_etf_tracking

                    Worked Example: Tracking Error and the Benchmark Closing Time

                    Slide 45 in the PDF

                    IBIT trades from 9:30 to 16:00 New York time on weekdays; Bitcoin trades 24 hours a day, 7 days a week. On the same days we can therefore compare the fund with two different Bitcoin prices.

                    With Bitcoin at 16:00 New York time, the standard deviation of the daily return differences is 0.223 percentage points, and \(\mathrm{TE} = 0.223 \times \sqrt{252} = 3.5\%\) a year; the correlation of returns is 0.998. With Bitcoin's daily close at 24:00 UTC, as in the EODHD daily file, the standard deviation is 1.299 points and \(\mathrm{TE} = 1.299 \times \sqrt{252} = 20.6\%\); the correlation drops to 0.907. The fund did not change; only the clock of the benchmark did. Over the whole period IBIT gained 73.6% and Bitcoin 75.6%, and the tracking difference was \(-0.28\%\) a year.

                    The lesson applies to every index product: compare a fund with its index at the time the fund is priced, and read the tracking error and the tracking difference separately. The first tells how closely the fund follows the index from day to day; the second, how much it costs the investor over time.

                    In brief

                    • IBIT trades 9:30–16:00 New York time on weekdays; Bitcoin trades 24 hours a day, 7 days a week
                    • Same days, two Bitcoin prices
                      • at 16:00 New York: sd of daily differences 0.223 percentage points, \(\mathrm{TE} = 0.223 \times \sqrt{252} = 3.5\%\); correlation 0.998
                      • at 24:00 UTC (EODHD daily close): sd 1.299 points, \(\mathrm{TE} = 1.299 \times \sqrt{252} = 20.6\%\); correlation 0.907
                    • The fund did not change; only the clock of the benchmark did
                    • Tracking difference: \(-0.28\%\) a year, IBIT \(+73.6\%\) against Bitcoin \(+75.6\%\) over the period
                    • Lesson for index products: compare a fund with its index at the time the fund is priced, then read TE and TD separately

                    Section summary

                    A fund turns an index into a product by buying its constituents in the index weights; fees create a tracking difference and frictions a tracking error. CRIX became a commercial index in 2021, after its academic start. A spot Bitcoin fund, measured with synchronous prices, follows Bitcoin with a tracking error of 3.5% a year and a tracking difference of \(-0.28\%\), close to its fee.

                    Nine Types of Crypto Indices

                    CRIX is a broad market index, but crypto indices serve many other purposes. Following Găman, Będowska-Sójka, Pele & Härdle (2026), the section sorts them into nine types, reads the formula of one representative index of each type from its provider's methodology, and rebuilds two of the ideas on our own data: square-root weights and a protection floor.

                    Nine Types of Crypto Indices

                    Slide 46 in the PDF

                    Găman, Będowska-Sójka, Pele & Härdle (2026) reviewed the crypto indices whose methodology is public and grouped them by scope into nine types. A multi-asset index summarises several coins in one number, usually by market value, as the Bitwise 10 (BITX) and CRIX do. A sector index restricts the universe to one segment, such as decentralised finance (DeFi), smart-contract platforms or the Metaverse. A price index gives a reference price of one coin from several exchanges, and a futures index follows futures contracts that are rolled from one maturity to the next. A volatility index measures expected or realised variability, a "fear gauge" such as EVIV or VCRIX. A yield index tracks income from cash, lending or staking together with crypto exposure. A smart-beta index uses rule-based weights other than market value, a fundamentals index weights coins by on-chain activity such as fees and active users, and a sentiment index combines market and behavioural signals into one score.

                    Most of these indices share one backbone, the Laspeyres index with a divisor of the previous sections. They differ in four design layers: the universe, the weights and constraints, the data, and the operating rules for rebalancing and reconstitution. These choices, not the name of the index, determine the exposure an investor buys.

                    In brief

                    Typewhat it tracksexample
                    Multi-assetseveral coins in one number, usually by market valueBITX; CRIX
                    Sectorone segment: DeFi, smart-contract platforms, MetaverseCoinDesk Metaverse Select
                    Pricethe spot price of one coin, from several exchangesMarketVector Polygon VWAP Close
                    Futuresfutures contracts, rolled from one maturity to the nextCoinDesk BFIC
                    Volatilityexpected or realised variability: a “fear gauge”EVIV; VCRIX
                    Yieldincome from cash, lending or staking, with crypto exposureLIXX Risk-Adjusted Bitcoin
                    Smart betarule-based weights other than market valueCOINTILT
                    Fundamentalson-chain activity: fees, active usersMVTT10F
                    Sentimentmarket and behavioural signals in one scoreCMC Fear and Greed
                    • Taxonomy and examples from Găman, Będowska-Sójka, Pele & Härdle (2026); DeFi: decentralised finance
                    • Most of them share one backbone, the Laspeyres index with a divisor; they differ in universe, weights, data and rules

                    Multi-Asset: the Bitwise 10 (BITX)

                    Slide 47 in the PDF

                    The Bitwise 10 Large Cap Crypto Index (BITX) holds the ten largest eligible coins by free-float market value (Bitwise, 2025; Găman, Będowska-Sójka, Pele & Härdle, 2026):

                    \[I_t = \frac{1}{D_t}\sum_{i \in \mathcal{I}_t} P_{it}\,Q^{\text{ff}}_{it}\]
                    (3.14)

                    where \(Q^{\text{ff}}_{it}\) is the free-float supply, the coins that can actually trade, and \(D_t\) the divisor. The index was set to 964 on 1 January 2017, the price of Bitcoin that day. The weights are then rescaled so that at least 87.5% go to coins that have a single-asset exchange-traded product (ETP) approved by the US Securities and Exchange Commission (SEC).

                    Both the trading venues and the coins must pass filters. A venue must be regulated, free of capital controls, have secure interfaces and carry at least 0.1% of global crypto volume. A coin must not be pegged, must be listed on several approved venues and have no critical security flaws; its volume over 30 days must exceed 1% of its free-float market value, and its risk of being classified as a security under US law must be low. The index is reconstituted monthly, and a committee and an advisory board review the rules. CRIX differs on one point: the number of constituents is chosen by the data, not fixed at ten.

                    In brief

                    • The ten largest eligible coins by free-float market value Bitwise (2025), Găman, Będowska-Sójka, Pele & Härdle (2026)
                      \[ I_t = \frac{1}{D_t}\sum_{i \in \mathcal{I}_t} P_{it}\,Q^{\text{ff}}_{it} \]
                      • \(Q^{\text{ff}}_{it}\): free-float supply (coins that can trade); \(D_t\): divisor; \(I = 964\) on 1 January 2017, the price of Bitcoin that day
                      • weights rescaled so that at least 87.5% go to coins with an SEC-approved single-asset ETP
                    • Eligible trading venues
                      • regulated, no capital controls, secure interfaces, at least 0.1% of global crypto volume
                    • Eligible coins
                      • not pegged, listed on several approved venues, no critical security flaws
                      • 30-day volume above 1% of the free-float market value; low risk of being a security under US law
                    • Monthly reconstitution; a committee and an advisory board review the rules
                    • CRIX differs on one point: the number of constituents is chosen by the data, not fixed at ten

                    Sector and Price Indices

                    Slide 48 in the PDF

                    The CoinDesk Metaverse Select Index is a sector index (CoinDesk, 2024). Its universe is the Metaverse group of the Digital Asset Classification Standard (DACS) of CoinDesk; meme coins are excluded, and the coins must be held in secure custody and be accessible to US investors. The coins are ranked by size and checked for liquidity with the median daily value traded (MDVT) over 30 days; between 5 and 10 coins are kept, each with a weight of at least 1%. The index is a Laspeyres index,

                    \[\mathrm{Index}_t = \frac{\sum_{i=1}^{N} P_{it}\,S_i\,\mathrm{WAF}_i}{\mathrm{Divisor}}\]
                    (3.15)

                    with \(S_i\) the circulating supply at the weighting date and \(\mathrm{WAF}_i\) a weighting adjustment factor; it is reconstituted quarterly.

                    The MarketVector Polygon VWAP Close Index is a price index (MarketVector, 2025a). It holds one coin, Polygon (MATIC), with a weight of 100%, and its daily close is the one-hour volume-weighted average price (VWAP) from 15:00 to 16:00 Central European Time, with a base value of 100 on 30 September 2020. After a hard fork the new coins may enter the index for a while, weighted by market value. The formula has the same structure,

                    \[\mathrm{Index}_t = \frac{\sum_{i=1}^{n} p_i\,q_i\,\mathit{cf}_i\,\mathit{fx}_i}{D}, \qquad \mathrm{VWAP} = \frac{\sum_j p_j v_j}{\sum_j v_j}\]
                    (3.16)

                    with \(\mathit{cf}_i\) a cap factor, usually 1, \(\mathit{fx}_i\) the exchange rate to USD, and \(p_j, v_j\) the price and volume of trade \(j\) in the hour. Averaging over an hour of trades makes the close harder to move with a single trade.

                    In brief

                    • Sector: CoinDesk Metaverse Select CoinDesk (2024)
                      • coins of the Metaverse group of the Digital Asset Classification Standard (DACS); no meme coins; custody and US access required
                      • ranked by size; liquidity by the median daily value traded (MDVT) over 30 days; 5 to 10 coins; minimum weight 1%
                      \[ \mathrm{Index}_t = \frac{\sum_{i=1}^{N} P_{it}\,S_i\,\mathrm{WAF}_i}{\mathrm{Divisor}} \]
                      • \(S_i\): circulating supply at the weighting date; \(\mathrm{WAF}_i\): weighting adjustment factor; quarterly reconstitution
                    • Price: MarketVector Polygon VWAP Close MarketVector (2025a)
                      • one coin, Polygon (MATIC), weight 100%; close = one-hour volume-weighted average price (VWAP), 15:00–16:00 Central European Time
                      • base value 100 on 30 September 2020; after a hard fork the new coins may enter for a while, weighted by market value
                      \[ \mathrm{Index}_t = \frac{\sum_{i=1}^{n} p_i\,q_i\,\mathit{cf}_i\,\mathit{fx}_i}{D}, \qquad \mathrm{VWAP} = \frac{\sum_j p_j v_j}{\sum_j v_j} \]
                      • \(\mathit{cf}_i\): cap factor (usually 1); \(\mathit{fx}_i\): exchange rate to USD; \(p_j, v_j\): price and volume of trade \(j\) in the hour

                    Futures: Rolling CME Contracts

                    Slide 49 in the PDF

                    The CoinDesk Bitcoin Futures Excess Return Index (BFIC) follows Bitcoin futures of the Chicago Mercantile Exchange (CME) (CoinDesk, 2025; Găman, Będowska-Sójka, Pele & Härdle, 2026). It was launched on 7 February 2024 with a base of 14,690.00 on 27 December 2017. The index holds the front contract; before it expires, the exposure is moved to the next contract, which is called the roll:

                    \[I_t = U_{L,t}\,P_{L,t} + U_{N,t}\,P_{N,t}, \qquad U_{L,t} = U_{L,\text{start}} - r, \quad U_{N,t} = r\]
                    (3.17)

                    where \(U\) are the units of the lead (L) and next (N) contract and \(P\) their CME settlement prices at 16:00 New York time. The roll starts on the Friday before the last trade date and ends on the Wednesday: a quarter of the position moves each day over four days.

                    There are no market-value weights; the exposure is set in contract units. The return of the index has two parts: the change of the futures price and the roll yield. If the next contract is dearer than the lead one, a situation called contango, each roll buys fewer units, and the index lags spot Bitcoin.

                    In brief

                    • CoinDesk Bitcoin Futures Excess Return Index (BFIC) CoinDesk (2025), Găman, Będowska-Sójka, Pele & Härdle (2026)
                      • Bitcoin futures of the Chicago Mercantile Exchange (CME); launched 7 February 2024, base 14 690.00 on 27 December 2017
                      • holds the front contract; before it expires, exposure moves to the next contract (the roll)
                      \[ I_t = U_{L,t}\,P_{L,t} + U_{N,t}\,P_{N,t}, \qquad U_{L,t} = U_{L,\text{start}} - r, \quad U_{N,t} = r \]
                      • \(U\): units of the lead (L) and next (N) contract; \(P\): CME settlement price at 16:00 New York time
                      • the roll starts on the Friday before the last trade date and ends on the Wednesday: a quarter of the position per day over four days
                    • No market-value weights: exposure is set in contract units
                    • The return has two parts: the change of the futures price and the roll yield
                      • if the next contract is dearer than the lead (contango), each roll buys fewer units: a drag against spot Bitcoin

                    Volatility: Implied Volatility from Options

                    Slide 50 in the PDF

                    The Ethereum Volmex Implied Volatility 30 Day Index (EVIV) measures the volatility over the next 30 days expected by the Ether options market (Volmex; Găman, Będowska-Sójka, Pele & Härdle, 2026). As for the VIX (Whaley, 2000), the variance of each maturity is computed from the prices of out-of-the-money calls and puts:

                    \[\sigma_i^2 = \frac{1}{T_i}\Big(2\sum_j \frac{\Delta K_j}{K_j^2}\,e^{r_iT_i}\,V(K_j) - \Big(\frac{F_i}{K_{i,\text{ATM}}} - 1\Big)^2\Big), \qquad i \in \{\text{near}, \text{next}\}\]
                    (3.18)

                    where \(V(K_j)\) is the price of the option with strike \(K_j\), \(F_i\) the forward price and \(K_{i,\text{ATM}}\) the at-the-money strike. The near and next variances are interpolated to 30 days, \(\sigma^2_{\text{raw}} = \omega_{\text{near}}\sigma^2_{\text{near}} + \omega_{\text{next}}\sigma^2_{\text{next}}\), and then smoothed by an exponentially weighted moving average (EWMA):

                    \[\sigma^2_{\text{smooth},t} = \lambda\,\sigma^2_{\text{smooth},t-1} + (1 - \lambda)\,\sigma^2_{\text{raw},t}, \qquad \lambda = e^{-\ln 2/\tau}, \qquad \mathrm{EVIV}_t = 100\,\sigma_{\text{smooth},t}\]
                    (3.19)

                    The half-life \(\tau\) is the number of periods after which an old value has lost half of its weight. VCRIX, studied earlier in this lecture, has no options to use: it forecasts the volatility of CRIX from its own past with a HAR model.

                    In brief

                    • Ethereum Volmex Implied Volatility 30 Day Index (EVIV) Volmex, Găman, Będowska-Sójka, Pele & Härdle (2026): the 30-day expected volatility from Ether options
                    • Variance of each maturity from out-of-the-money calls and puts, as for the VIX Whaley (2000):
                      \[ \sigma_i^2 = \frac{1}{T_i}\Big(2\sum_j \frac{\Delta K_j}{K_j^2}\,e^{r_iT_i}\,V(K_j) - \Big(\frac{F_i}{K_{i,\text{ATM}}} - 1\Big)^2\Big), \qquad i \in \{\text{near}, \text{next}\} \]
                      • \(V(K_j)\): price of the option with strike \(K_j\); \(F_i\): forward price; \(K_{i,\text{ATM}}\): at-the-money strike
                    • Interpolation to 30 days, \(\sigma^2_{\text{raw}} = \omega_{\text{near}}\sigma^2_{\text{near}} + \omega_{\text{next}}\sigma^2_{\text{next}}\), then smoothing by an exponentially weighted moving average (EWMA):
                      \[ \sigma^2_{\text{smooth},t} = \lambda\,\sigma^2_{\text{smooth},t-1} + (1 - \lambda)\,\sigma^2_{\text{raw},t}, \qquad \lambda = e^{-\ln 2/\tau}, \qquad \mathrm{EVIV}_t = 100\,\sigma_{\text{smooth},t} \]
                      • \(\tau\): half-life; an old value loses half of its weight after \(\tau\) periods
                    • VCRIX (this lecture) has no options to use: it forecasts the volatility of CRIX from its own past with a HAR model

                    Yield: Bitcoin and Cash with a Protection Floor

                    Slide 51 in the PDF

                    The LIXX Risk-Adjusted Bitcoin & USD Overnight Interest Index combines two parts: Bitcoin, and USD cash that earns the overnight interest rate (LIXX, 2023; Găman, Będowska-Sójka, Pele & Härdle, 2026). It starts with 19% in Bitcoin and 81% in cash. A protection floor is checked every day:

                    \[\mathrm{Floor}_t = (\text{lock-in} + \text{buffer}) \times \max_{s \in [t-1\text{y},\,t]} I_s, \qquad \text{lock-in} = 80\%, \ \text{buffer} = 1\%\]
                    (3.20)

                    The index is rebalanced when the gap between the floor and the cash position exceeds the buffer: it sells Bitcoin after falls and buys after rises. A fee of 0.3% of the traded amount is charged. The index value is the value of the two parts, \(I_t = \sum_i W_{it}\,P_{it} - A_t\), with \(W_{it}\) the units of Bitcoin and of cash and \(A_t\) an adjustment for forks or changes of method.

                    We rebuilt the index with our reading of the rule: the cash is kept equal to the floor, so the index starts with 81% in cash, as published, and the cash earns the effective federal funds rate (FRED: DFF).

                    In brief

                    • LIXX Risk-Adjusted Bitcoin & USD Overnight Interest Index LIXX (2023), Găman, Będowska-Sójka, Pele & Härdle (2026)
                      • two parts: Bitcoin, and USD cash that earns the overnight rate; start: 19% Bitcoin, 81% cash
                    • Protection floor, checked every day:
                      \[ \mathrm{Floor}_t = (\text{lock-in} + \text{buffer}) \times \max_{s \in [t-1\text{y},\,t]} I_s, \qquad \text{lock-in} = 80\%, \ \text{buffer} = 1\% \]
                      • rebalance when the gap between the floor and the cash exceeds the buffer: sell Bitcoin after falls, buy after rises
                      • a fee of 0.3% of the traded amount
                    • Index value: \(I_t = \sum_i W_{it}\,P_{it} - A_t\)
                      • \(W_{it}\): units of Bitcoin and of cash; \(A_t\): adjustment for forks or changes of method
                    • Our reading of the rule, on real data: cash \(=\) floor, so the start is 81% cash, as published; cash earns the effective federal funds rate FRED: DFF

                    The Protection Floor on Real Data

                    Slide 52 in the PDF

                    The chart compares our floor index with Bitcoin bought and held, both 1,000 on 31 March 2018, to 30 October 2026. The floor index grew by 9.9% a year, with a volatility of 9.1% and a maximum drawdown of \(-13.9\%\); Bitcoin grew by 32.3% a year, with a volatility of 61.6% and a maximum drawdown of \(-76.6\%\).

                    The right panel shows why. The weight of Bitcoin stayed between 6.1% and 20.0%, and the rule rebalanced 66 times in 8.6 years. The floor protects the investor in crypto winters but gives up most of the growth: it is a different product, not a safer version of Bitcoin.

                    In brief

                    The Protection Floor on Real Data
                    • 31 March 2018 – 30 October 2026: floor index CAGR 9.9%, volatility 9.1%, maximum drawdown \(-13.9\%\); Bitcoin: 32.3%, 61.6%, \(-76.6\%\)
                    • Weight of Bitcoin between 6.1% and 20.0%; 66 rebalancings in 8.6 years
                    BA_L3_floor

                    Smart Beta and Fundamentals

                    Slide 53 in the PDF

                    The Coinbase Size Tilt Index (COINTILT) is a smart-beta index (MarketVector, 2025b). It weights each coin by the square root of its market value,

                    \[w_i = \frac{\sqrt{\mathrm{MV}_i}}{\sum_{j=1}^{n}\sqrt{\mathrm{MV}_j}}\]
                    (3.21)

                    which tilts the index towards smaller coins; there is no cap. A coin is eligible if it holds at least 0.25% of the market value of the MarketVector Digital Assets 100 index, has traded for 180 days with at least 25 million USD of volume a day, and can be held by custodians approved by Coinbase.

                    The MarketVector Token Terminal Fundamental Index (MVTT10F) is a fundamentals index (MarketVector, 2024). Its weights combine factor weights,

                    \[w_i = \sum_{k=1}^{m} a_k\,w_{k,i}, \qquad w_{k,i} = \frac{k_i}{\sum_{j=1}^{n} k_j}\]
                    (3.22)

                    with two on-chain factors: fees, with \(a = 80\%\), and active users, with \(a = 20\%\), each the monthly mean of the square roots of the daily values. The index holds ten coins: the top seven by the factor ranking, then three from ranks 8 to 13; meme and privacy coins are excluded. Both indices use the same Laspeyres calculation as BITX and CRIX; only the weights change.

                    In brief

                    • Smart beta: Coinbase Size Tilt Index (COINTILT) MarketVector (2025b)
                      \[ w_i = \frac{\sqrt{\mathrm{MV}_i}}{\sum_{j=1}^{n}\sqrt{\mathrm{MV}_j}} \]
                      • the square root tilts the index towards smaller coins; no cap
                      • eligibility: at least 0.25% of the market value of the MarketVector Digital Assets 100, 180 days of trading, 25 million USD a day of volume, custody by Coinbase-approved custodians
                    • Fundamentals: MarketVector Token Terminal Fundamental Index (MVTT10F) MarketVector (2024)
                      \[ w_i = \sum_{k=1}^{m} a_k\,w_{k,i}, \qquad w_{k,i} = \frac{k_i}{\sum_{j=1}^{n} k_j} \]
                      • two on-chain factors \(k\): fees, \(a = 80\%\), and active users, \(a = 20\%\); each the monthly mean of the square roots of the daily values
                      • ten coins: the top seven by the factor ranking, then three from ranks 8 to 13; meme and privacy coins excluded
                    • Both use the same Laspeyres calculation as BITX and CRIX; only the weights change

                    Worked Example: Square-Root Weights

                    Slide 54 in the PDF

                    On 30 September 2026 the market values of our ten coins were, in billion USD, Bitcoin 1,678.7, Ethereum 327.7 and the other eight 81.7 together. With market-value weights, Bitcoin has \(1{,}678.7/2{,}088.1 = 80.8\%\) and Ethereum 15.8%.

                    With square-root weights, the square roots are \(\sqrt{1{,}678.7} = 41.0\) and \(\sqrt{327.7} = 18.1\), and their sum over the ten coins is 80.9. Bitcoin gets \(41.0/80.9 = 50.7\%\), Ethereum \(18.1/80.9 = 22.4\%\), and the other eight 26.9% together. The effective number of coins, \(1/\mathrm{HHI}\), is 1.48 with market-value weights, 3.15 with square-root weights and 10 with equal weights. The square root does not change the ranking of the coins; it compresses the distance between large and small ones.

                    In brief

                    • Ten coins, 30 September 2026, market values in billion USD: BTC 1 678.7, ETH 327.7, the other eight 81.7 together
                    • Market-value weights: BTC \(1\,678.7/2\,088.1 = 80.8\%\), ETH 15.8%
                    • Square roots: \(\sqrt{1\,678.7} = 41.0\), \(\sqrt{327.7} = 18.1\); sum over the ten coins 80.9
                      • BTC \(41.0/80.9 = 50.7\%\), ETH \(18.1/80.9 = 22.4\%\), the other eight 26.9% together
                    • Effective number of coins, \(1/\mathrm{HHI}\): 1.48 with market-value weights, 3.15 with square-root weights, 10 with equal weights
                    • The square root does not change the ranking; it compresses the distance between large and small coins

                    Three Weighting Schemes for Ten Coins

                    Slide 55 in the PDF

                    The chart applies three weightings to the same ten coins, rebalanced monthly from 31 March 2018 to 30 October 2026: market value, as in BITX; the square root of market value, as in COINTILT; and equal weights. The market-value index grew by 28.0% a year with a volatility of 64.6% and a maximum drawdown of \(-77.9\%\); the square-root index by 26.7%, 69.7% and \(-83.4\%\); the equal-weight index by 32.8%, 80.0% and \(-84.5\%\).

                    The daily returns of the square-root index correlate 0.980 with those of the market-value index, those of the equal-weight index 0.880. Giving more weight to small coins raises volatility and deepens drawdowns; whether it raises growth depends on the period. The right panel shows the main difference: Bitcoin is about four fifths of the market-value index, half of the square-root index and a tenth of the equal-weight index.

                    In brief

                    Three Weighting Schemes for Ten Coins
                    • Monthly rebalancing, 31 March 2018 – 30 October 2026; CAGR / volatility / maximum drawdown: market value 28.0% / 64.6% / \(-77.9\%\); square root 26.7% / 69.7% / \(-83.4\%\); equal 32.8% / 80.0% / \(-84.5\%\)
                    • Correlation of daily returns with the market-value index: 0.980 (square root), 0.880 (equal weights)
                    BA_L3_weighting

                    Sentiment: the Fear and Greed Index

                    Slide 56 in the PDF

                    The CMC Crypto Fear and Greed Index of CoinMarketCap gives one daily score from 0, extreme fear, to 100, extreme greed (CoinMarketCap, 2025; Găman, Będowska-Sójka, Pele & Härdle, 2026):

                    \[\text{Score}_t = \sum_{k=1}^{m} w_k\,s_{kt}, \qquad s_{kt} \in [0, 100]\]
                    (3.23)

                    Each component is scaled to 0–100 and weighted: volatility, from the implied volatility of Bitcoin and Ether options, 25%; market momentum and volume 25%; social media 15%; surveys 15%; Bitcoin dominance 10%; and search trends 10%. The score falls into bands: 0–24 extreme fear, 25–49 fear, 50 neutral, 51–74 greed and 75–100 extreme greed. It is not an investible index: there is no basket and no divisor, only a summary of signals, and the choice of weights is the provider's judgement.

                    In brief

                    • CMC Crypto Fear and Greed Index CoinMarketCap (2025), Găman, Będowska-Sójka, Pele & Härdle (2026): one daily score from 0 (extreme fear) to 100 (extreme greed)
                      \[ \text{Score}_t = \sum_{k=1}^{m} w_k\,s_{kt}, \qquad s_{kt} \in [0, 100] \]
                    • Components and weights \(w_k\)
                      • volatility (implied volatility of Bitcoin and Ether options) 25%; market momentum and volume 25%
                      • social media 15%; surveys 15%; Bitcoin dominance 10%; search trends 10%
                    • Bands: 0–24 extreme fear, 25–49 fear, 50 neutral, 51–74 greed, 75–100 extreme greed
                    • Not an investible index: no basket, no divisor; a summary of signals, and the choice of weights is the provider's judgement

                    Broad Market Index Families

                    Slide 57 in the PDF

                    CRIX is one of several crypto index families; the table describes the main ones from their providers' documents, without performance figures. The Royalton CRIX chooses its number of constituents with an information criterion, uses market-value weights, rebalances monthly and reselects quarterly (thecrix.de). The S&P Cryptocurrency Broad Digital Market index of S&P Dow Jones Indices, launched in 2021, covers more than 200 coins (S&P Dow Jones Indices: Cryptocurrency Indices). The Bloomberg Galaxy Crypto Index (BGCI) holds the largest eligible coins with a weight cap of 40% and a floor of 1%, and is reconstituted monthly (Bloomberg, 2020). The CoinDesk 20 (CD20) holds the 20 largest coins excluding stablecoins and memecoins, with a cap of 30% for the largest constituent and 20% for the others, and is reviewed quarterly (CoinDesk Indices, 2025). The Nasdaq Crypto Index (NCI) is designed as a benchmark for a significant part of the digital-asset market (Nasdaq: NCI).

                    The comparison shows the main design choices of a crypto index: a fixed or a data-driven number of constituents, a cap on the largest weight or none, and the frequency of review. Methodologies change over time, so the current version should be checked with each provider.

                    In brief

                    Indexconstituents and weightssource
                    Royalton CRIXnumber chosen by an information criterion; market-value weights; monthly rebalancing, quarterly reselectionthecrix.de
                    S&P Cryptocurrency Broad Digital Marketover 200 coins (S&P Dow Jones Indices, since 2021)S&P Dow Jones Indices: Cryptocurrency Indices
                    Bloomberg Galaxy Crypto Index (BGCI)largest eligible coins; weight cap 40%, floor 1%; monthlyBloomberg (2020)
                    CoinDesk 20 (CD20)largest 20 coins, no stablecoins or memecoins; cap 30% for the largest, 20% for the others; quarterlyCoinDesk Indices (2025)
                    Nasdaq Crypto Index (NCI)benchmark of a significant part of the digital-asset marketNasdaq: NCI
                    • Descriptive only, from the providers' documents; methodologies change, check the current version
                    • The main design choices: fixed or data-driven number of constituents, cap or no cap, frequency of review

                    Section summary

                    Crypto indices come in nine types, from market and sector indices to futures, volatility, yield, smart-beta, fundamentals and sentiment indices. Most share the Laspeyres backbone and differ in universe, weights, data and rules. On the same ten coins, Bitcoin is 80.8%, 50.7% or 10% of the index depending on the weighting, and a protection floor cut the maximum drawdown from \(-76.6\%\) to \(-13.9\%\) at the cost of most of the growth.

                    Summary

                    The crypto market is concentrated: Bitcoin holds 82% of the value of our ten coins, an effective number of 1.4 coins. A list of today's winners overstates past returns: bought in 2018, the top five of 2026 earned 17.6% a year in a backtest, against 7.1% for the five coins an investor could have chosen at the time.

                    CRIX is a Laspeyres market-value index with a divisor, a liquidity rule and a number of constituents chosen by the AIC, smoothed over six months. Nine types of crypto indices share the same backbone and differ in universe, weights, data and rules: on the same ten coins Bitcoin is 80.8%, 50.7% or 10% of the index, and a protection floor cut the drawdown from \(-76.6\%\) to \(-13.9\%\). Our ten-coin replica tracks CRIX with a correlation of 0.990 and a tracking error of 9.3% a year; Bitcoin alone reaches 0.971 and 15.6%.

                    From 2018 to 2026 CRIX grew by 24.0% a year with a volatility of 65.0% and a maximum drawdown of 77.4%; its record of 21,523.34 came on 6 October 2025. VCRIX follows the 30-day realised volatility of CRIX with a correlation of 0.94, and a GARCH(1,1) model has a persistence of 0.998. A spot Bitcoin fund has a tracking difference of \(-0.28\%\) a year and a tracking error of 3.5% when prices are synchronous.

                    In brief

                    • The crypto market is concentrated: Bitcoin holds 82% of our ten coins, an effective number of 1.4 coins
                    • A list of today's winners overstates past returns: 17.6% a year against 7.1% for the coins chosen ex ante
                    • CRIX: a Laspeyres market-value index with a divisor; liquidity rule; number of constituents by the AIC, smoothed over six months
                    • Nine types of crypto indices share the same backbone and differ in universe, weights, data and rules
                    • Same ten coins, three weightings: Bitcoin 80.8%, 50.7% or 10% of the index; a protection floor cut the drawdown from \(-76.6\%\) to \(-13.9\%\)
                    • Our ten-coin replica tracks CRIX with correlation 0.990 and tracking error 9.3% a year; Bitcoin alone: 0.971 and 15.6%
                    • CRIX 2018–2026: CAGR 24.0%, volatility 65.0%, maximum drawdown \(-77.4\%\); record 21 523.34 on 6 October 2025
                    • VCRIX follows the 30-day realised volatility (correlation 0.94); GARCH persistence 0.998
                    • A spot Bitcoin fund: tracking difference \(-0.28\%\) a year, tracking error 3.5% when prices are synchronous

                    Review Quiz (1)

                    Slide 59 in the PDF

                    The first three questions repeat the mechanics of the index: the divisor follows from (3.4), the first-rise rule reads the AIC values from left to right and stops at the first increase, and square-root weights give more weight to smaller coins without changing their ranking. The practice quiz of Lecture 3 on the course website has more questions of this kind, drawn at random.

                    In brief

                    1. On a rebalancing day the market value of the new basket is 2 045.41 billion USD and the index stands at 9 521.15. The new divisor is
                      • 9 521.15   (b) 214.83 million   (c) 2 045.41 billion   (d) 0.0047
                    2. \(\mathrm{AIC}(k)\): \(-946.2\), \(-953.3\), \(-953.6\), \(-952.3\) for \(k = 2, 3, 4, 5\). The first-rise rule chooses
                      • \(k = 2\)   (b) \(k = 3\)   (c) \(k = 4\)   (d) \(k = 5\)
                    3. An index weights each coin by the square root of its market value. Compared with market-value weights it
                      • changes the ranking of coins   (b) gives more weight to smaller coins   (c) removes Bitcoin   (d) fixes the number of coins at ten
                    • Answers: 1 (b), 2 (c), 3 (b); more questions in the practice quiz of Lecture 3 on the course website

                    Review Quiz (2)

                    Slide 60 in the PDF

                    The second set asks for interpretation: what tracking difference and tracking error mean for an investor, why VCRIX drops sharply exactly 30 days after a crash, and what the gap between an ex-post and an ex-ante basket measures.

                    In brief

                    • A fund has a tracking difference of \(-0.28\%\) a year and a tracking error of 3.5%. Which statement is right?
                      • it lags its index by about 3.5% every year   (b) it lags by about 0.28% a year on average and deviates by about 3.5% a year around that
                    • VCRIX jumps on 13 March 2020 and drops sharply 30 days later, on 12 April 2020. The drop is explained by
                      • a calm day on 12 April   (b) the crash day leaving the 30-day window
                    • The ex-post top 5 of 2026, bought in 2018, earned 17.6% a year; the ex-ante top 5 earned 7.1%. The difference measures
                      • the skill of the index provider   (b) look-ahead and survivorship bias
                    • Answers: 4 (b), 5 (b), 6 (b)

                    Interpretation Exercises

                    Slide 61 in the PDF

                    The exam asks for interpretation, not calculation, as in these six exercises. Each gives numbers from the lecture and asks what they mean for an index provider, a fund or an investor, and what they do not show.

                    In brief

                    1. Bitcoin holds 82.2% of the value of ten coins; \(1/\mathrm{HHI} = 1.44\). A colleague proposes an equally weighted index “to represent the market better”. Comment on representativeness and investibility.
                    2. The AIC curve of the last quarter falls by 284 points from \(k = 1\) to \(k = 2\) and by 0.3 from \(k = 3\) to \(k = 4\). What does this say about the number of constituents and about the stability of the choice? Why smooth the curve over six months?
                    3. Our replica: correlation 0.990 and tracking error 9.3% against CRIX; capped below 50%: 0.979 and 14.5%. What do these numbers say about the published index? What do they not prove?
                    4. CRIX: CAGR 24.0%, volatility 65.0%; Bitcoin: 29.4% and 61.7% (2018–2026). Would a broad crypto index have helped a Bitcoin investor? In which year was the answer different?
                    5. A Bitcoin-and-cash index with a protection floor: CAGR 9.9%, maximum drawdown \(-13.9\%\); Bitcoin: 32.3% and \(-76.6\%\). For which investor is the floor index the better benchmark? What does the investor give up?
                    6. IBIT: tracking difference \(-0.28\%\) a year; tracking error 3.5% at 16:00 New York and 20.6% at midnight UTC. Which number would you report to an investor, and why?

                    References (1)

                    Slide 62 in the PDF
                    • Akaike, H. (1974). A new look at the statistical model identification. IEEE Transactions on Automatic Control 19(6), 716–723. doi:10.1109/TAC.1974.1100705
                    • Bollerslev, T. (1986). Generalized autoregressive conditional heteroskedasticity. Journal of Econometrics 31(3), 307–327. doi:10.1016/0304-4076(86)90063-1
                    • Brown, S. J., Goetzmann, W., Ibbotson, R. G., Ross, S. A. (1992). Survivorship bias in performance studies. Review of Financial Studies 5(4), 553–580. doi:10.1093/rfs/5.4.553
                    • Corsi, F. (2009). A simple approximate long-memory model of realized volatility. Journal of Financial Econometrics 7(2), 174–196. doi:10.1093/jjfinec/nbp001
                    • Găman, Ș., Będowska-Sójka, B., Pele, D. T., Härdle, W. K. (2026). White paper on cryptocurrency indices. IDA Institute Digital Assets, working paper.
                    • Härdle, W. K. (2026). System and method for managing an investible cryptocurrency index fund. US Patent 12,784,107 B2, granted 3 November 2026; assignee Royalton Partners AG. PDF
                    • Kim, A., Trimborn, S., Härdle, W. K. (2021). VCRIX: a volatility index for crypto-currencies. International Review of Financial Analysis 78, 101915. doi:10.1016/j.irfa.2021.101915

                    References (2)

                    Slide 63 in the PDF
                    • Petukhina, A., Trimborn, S., Härdle, W. K., Elendner, H. (2021). Investing with cryptocurrencies: evaluating their potential for portfolio allocation strategies. Quantitative Finance 21(11), 1825–1853. doi:10.1080/14697688.2021.1880023
                    • Trimborn, S., Härdle, W. K. (2018). CRIX an index for cryptocurrencies. Journal of Empirical Finance 49, 107–122. doi:10.1016/j.jempfin.2018.08.004
                    • Trimborn, S., Li, M., Härdle, W. K. (2020). Investing with cryptocurrencies: a liquidity constrained investment approach. Journal of Financial Econometrics 18(2), 280–306. doi:10.1093/jjfinec/nbz016
                    • Whaley, R. E. (2000). The investor fear gauge. Journal of Portfolio Management 26(3), 12–17. doi:10.3905/jpm.2000.319728

                    References (3)

                    Slide 64 in the PDF

                    Appendix: Further Material

                    The appendix documents the data and explains the kernel density estimator with leave-one-out used in the information criterion.

                    Appendix: Data Notes

                    Slide 65 in the PDF

                    CRIX and VCRIX were read from the chart data of thecrix.de on 7 November 2026: 3,321 daily CRIX values from 6 September 2017 to 6 November 2026, with 28 missing days in 2026 (22 March, 14 May to 7 June, 31 August and 5 November). Returns are computed only between consecutive days; a return over a gap is not used.

                    Before 16 March 2018 CRIX moves one day ahead of the EODHD Bitcoin close: in 2017 the correlation of its daily returns is 0.90 with the next day's Bitcoin return and 0.06 with the same day's. thecrix.de notes that a different data source was used for 2016 to 2018. The market values of the replica are EODHD closes times the Coin Metrics current supply (`SplyCur`), and volumes are EODHD volumes in USD. The replica is a teaching replication of the published rules on a small universe, not the official index.

                    In brief

                    • CRIX and VCRIX were read from the chart data of thecrix.de on 7 November 2026
                      • 3 321 daily CRIX values, 6 September 2017 – 6 November 2026; 28 missing days in 2026 (22 March, 14 May – 7 June, 31 August, 5 November)
                      • returns are computed only between consecutive days; a return over a gap is not used
                    • Before 16 March 2018 CRIX moves one day ahead of the EODHD Bitcoin close (correlation of daily returns 0.90 with the next day's Bitcoin return in 2017, 0.06 with the same day); thecrix.de notes a different data source for 2016–2018
                    • Market values of the replica: EODHD close \(\times\) Coin Metrics current supply (SplyCur); volumes: EODHD, USD
                    • The replica is a teaching replication of the published rules on a small universe, not the official index

                    Appendix: Kernel Density and Leave-One-Out

                    Slide 66 in the PDF

                    A kernel density estimate smooths a sample \(x_1, \dots, x_n\) into a density by placing a small bump, the kernel \(K\), on each observation:

                    \[\hat f(x) = \frac{1}{nh}\sum_{s=1}^{n} K\Big(\frac{x - x_s}{h}\Big), \qquad K(u) = \varphi(u) = \frac{e^{-u^2/2}}{\sqrt{2\pi}}\]
                    (3.24)

                    The bandwidth \(h\) sets the width of the bumps. Evaluating \(\hat f\) at its own data overstates the likelihood, because each point sits on top of its own bump. The leave-one-out estimate drops the point itself: \(\hat f_{-t}(x_t) = \frac{1}{(n-1)h}\sum_{s \ne t} K\big((x_t - x_s)/h\big)\).

                    The bandwidth follows Silverman's rule, \(h = 0.9\,\min(\hat\sigma, \widehat{\mathrm{IQR}}/1.349)\,n^{-1/5}\). In the last quarter of the replica, with \(n = 90\) residuals, \(h = 0.00069\) in log-return units, or 0.069 percentage points.

                    In brief

                    • Kernel density estimate of a sample \(x_1, \dots, x_n\) with kernel \(K\) and bandwidth \(h\):
                      \[ \hat f(x) = \frac{1}{nh}\sum_{s=1}^{n} K\Big(\frac{x - x_s}{h}\Big), \qquad K(u) = \varphi(u) = \frac{e^{-u^2/2}}{\sqrt{2\pi}} \]
                    • Evaluating \(\hat f\) at its own data overstates the likelihood: each point sits on its own bump
                    • Leave-one-out: drop the point itself, \(\hat f_{-t}(x_t) = \frac{1}{(n-1)h}\sum_{s \ne t} K\big((x_t - x_s)/h\big)\)
                    • Bandwidth (Silverman's rule): \(h = 0.9\,\min(\hat\sigma, \widehat{\mathrm{IQR}}/1.349)\,n^{-1/5}\)
                      • last quarter of the replica: \(n = 90\) residuals, \(h = 0.00069\) in log-return units (0.069 percentage points)

                    References