Blockchain Applications

Lecture 7: Networks, Sentiment and Asset Classes

Lecture 7: Networks, Sentiment and Asset Classes

Daniel Traian PELE

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

Lecture Question and Outline

Slide 2 in the PDF

The lecture asks four related questions. How are crypto assets linked to each other? Are they an asset class of their own? Do protocol events move their prices? And does the sentiment of social-media posts carry information about the market?

The route follows these questions. The first section treats the market as a network: coins are the nodes and their co-movements the links, measured by correlations, a minimum spanning tree and the spillovers of Diebold & Yilmaz (2012). The second section describes each asset by the statistical features of its returns and asks whether crypto assets form a separate group, with a factor analysis and a classifier. The third runs an event study around 13 upgrades of the Ethereum protocol. The fourth describes the BitMood sentiment index, built from Facebook posts, and the standard a trading back-test must meet before its results can be trusted.

The data are daily prices from EOD Historical Data (EODHD) of 19 coins and 14 traditional assets, January 2021 – October 2026, and the dates of the Ethereum upgrades from 2019 to 2025.

In brief

  • Question: crypto assets do not move alone. How are they linked, are they an asset class of their own, and do protocol events and social media move prices?
  • Outline
    • networks: correlation trees, connectedness, Diebold–Yilmaz spillovers
    • cryptocurrencies as an asset class: factor analysis and classification
    • event studies: Ethereum upgrades
    • sentiment from social media: the BitMood index
  • Data: End-of-Day Historical Data (EODHD) daily prices of 19 coins and 14 traditional assets, January 2021 – October 2026; Ether upgrade dates 2019–2025

Learning Outcomes

Slide 3 in the PDF
  • Build a correlation distance, a minimum spanning tree and an absorption ratio, and interpret hubs and connectedness
  • Read a Diebold–Yilmaz spillover table: total, TO, FROM and NET spillovers
  • Explain how factor analysis and classification separate cryptos from other assets
  • Run and interpret a market-model event study: abnormal return (AR), cumulative abnormal return (CAR), individual and cross-sectional tests
  • Describe how a text-sentiment index is built and why trading back-tests need out-of-sample selection

Abbreviations (1/2)

Slide 4 in the PDF
ACF autocorrelation function
ADA Cardano
AI artificial intelligence
AIC Akaike information criterion
ALGO Algorand
AR absorption ratio (networks)
AR abnormal return (event studies)
ATOM Cosmos
BH Benjamini–Hochberg (adjusted p-values)
Bi-LSTM bidirectional LSTM
BMP Boehmer–Musumeci–Poulsen (test)
BNB BNB (Binance)
bp basis point (0.01%)
BTC Bitcoin
CAPM capital asset pricing model
CAR cumulative abnormal return
CNN convolutional neural network
CS compound score (sentiment)
CS cross-sectional (in \(t_{CS}\))
CSAD cross-sectional absolute deviation
CTE conditional tail expectation
DAI Digital Art Index
DOGE Dogecoin
DOT Polkadot
EIP Ethereum Improvement Proposal
EODHD End-of-Day Historical Data (data provider)
ETC Ethereum Classic
ETF exchange-traded fund
ETH Ether (Ethereum)
FABIZ Faculty of Business Administration in Foreign Languages, ASE
FIGARCH fractionally integrated GARCH
FIL Filecoin
FinBERT BERT language model for financial text
FTX FTX crypto exchange (collapsed November 2022)

Abbreviations (2/2)

Slide 5 in the PDF
FX foreign exchange
GARCH generalised autoregressive conditional heteroskedasticity
GRU gated recurrent unit
HAC heteroskedasticity and autocorrelation consistent
KMO Kaiser–Meyer–Olkin (sampling adequacy)
LDA latent Dirichlet allocation
LINK Chainlink
LSTM long short-term memory (network)
LTC Litecoin
MA moving average
MST minimum spanning tree
NFT non-fungible token
n.s. not significant
PC1, PC2 first, second principal component
RF random forest
RV realised volatility
RMSE root mean squared error
S&P 500 Standard & Poor's 500 stock index
SC synthetic control
SCAR standardised cumulative abnormal return
sd standard deviation
SE standard error
SHAP Shapley additive explanations
SI reaction-intensity score of a post (BitMood)
SOL Solana
SR Sharpe ratio
SVM support vector machine
TRX Tron
USDT Tether (stablecoin)
VAR vector autoregression
VaR value at risk
vol volatility
WC engagement-weighted compound score (BitMood)
XMR Monero
XRP XRP Ledger token

Reading and Tools

Slide 6 in the PDF

The main reading on asset classes is the article Are cryptos becoming alternative assets? by Pele, Wesselhöfft, Härdle, Kolossiatis & Yatracos (2023) in the European Journal of Finance. The network tools come from Mantegna (1999), who turned correlations into distances and trees, and from Diebold & Yilmaz (2012), who defined spillover indices from a vector autoregression. The herding test of the Appendix goes back to Christie & Huang (1995) and Chang, Cheng & Khorana (2000).

For event studies, the standard survey is MacKinlay (1997) in the Journal of Economic Literature; the study of Ethereum upgrades by Lin, Găman, Wang & Pele (2026) is a working paper. The sentiment index is described in BitMood: AI analysis of Bitcoin trends via Facebook emotions by Conda, Găman, Bâg, Mazurencu-Marinescu-Pele, Pele & Härdle (2026) in Digital Finance. Five video units of the Quantinar platform accompany the lecture. The methods are available in Python: `statsmodels` estimates vector autoregressions, `networkx` builds networks and spanning trees, and `scikit-learn` computes principal components and clusters; the notebook of Lecture 7 applies them to the data of the lecture.

In brief

Using a generalized vector autoregressive framework in which forecast-error variance decompositions are invariant to the variable ordering, we propose measures of both the total and directional volatility spillovers.

— Diebold & Yilmaz (2012)

Lectures 2 and 4 studied one coin at a time: its returns, its volatility and its tail risk. An investor, however, rarely holds a single coin, and a risk manager must know what happens to the whole book when one coin falls. Crypto assets do not move alone. The first question of this lecture is therefore how strongly the coins are linked, which coins sit at the centre of the market and which at its edge, and how these links change through time.

The second question is whether crypto assets form an asset class of their own, or whether some of them, Bitcoin in particular, now behave like stocks, bonds or currencies. The answer matters for portfolio construction and for regulation, because an asset class is treated as one block in an allocation. The third and fourth questions concern information: do announced protocol events, here the upgrades of Ethereum, move the price of Ether, and does the mood of social-media posts help to forecast the market?

The data are daily closing prices of 19 coins and 14 traditional assets (equities, Treasuries, gold, silver, oil and foreign-exchange rates) from January 2021 to October 2026, from the data provider EOD Historical Data (EODHD), the dates of 13 Ethereum upgrades from 2019 to 2025, and published results from two recent studies on upgrades and on Facebook sentiment.

How connected is the crypto market? Two measures computed on 90-day windows of daily log returns of 19 coins, from April 2021 to 22 October 2026: the mean of all pairwise correlations (dark blue) and the absorption ratio, the share of the total variance explained by the first principal component (orange). Both move between about 0.36 and 0.86 and track each other closely.
How connected is the crypto market? Two measures computed on 90-day windows of daily log returns of 19 coins, from April 2021 to 22 October 2026: the mean of all pairwise correlations (dark blue) and the absorption ratio, the share of the total variance explained by the first principal component (orange). Both move between about 0.36 and 0.86 and track each other closely.

What you will be able to do

  • Build a correlation distance, a minimum spanning tree and an absorption ratio, and interpret hubs and connectedness.
  • Read a Diebold–Yilmaz spillover table: total, TO, FROM and NET spillovers.
  • Explain how factor analysis and classification separate cryptos from other assets.
  • Run and interpret a market-model event study: abnormal return (AR), cumulative abnormal return (CAR), individual and cross-sectional tests.
  • Describe how a text-sentiment index is built and why trading back-tests need out-of-sample selection.

The Appendix holds optional material on herding and on price indices for digital art.

Lecture plan

  1. Crypto Markets as NetworksFour meanings of a link; correlation distance and the minimum spanning tree; the absorption ratio; Diebold–Yilmaz spillovers.
  2. Cryptocurrencies as an Asset ClassA map of 33 assets by the statistical features of their returns; the factor analysis and classification of Pele et al. (2023).
  3. Event Studies: Ethereum UpgradesThe market-model event study; abnormal returns around 13 upgrades; attention, activity and volatility.
  4. Sentiment from Social MediaThe BitMood index from Facebook posts; a contrarian trading rule; volume forecasts; the minimum standard for a back-test.
  5. SummaryThe main numbers of the lecture and five interpretation exercises.
  6. Appendix: Further MaterialHerding and the CSAD test; quality-adjusted price indices for digital art.

Crypto Markets as Networks

A network has nodes and links. Here the nodes are coins and the links describe how their returns move together. The section first separates four different meanings of a "link", then builds a tree from correlations, measures how much of the market is driven by one common factor, and finally measures which coins transmit shocks to the others.

Four Measures of Linkage between Crypto Assets

Slide 7 in the PDF

The word "link" is used for several different measures, and they answer different questions. A tail link, from Lecture 6, means that coin \(j\) has a nonzero slope in the 5% Quantile Lasso regression of coin \(i\): it concerns the lower tail of the distribution only and has a direction, from \(j\) to \(i\). A correlation \(\rho_{ij}\) means that the daily returns of \(i\) and \(j\) move together on the same day; it uses the whole distribution and has no direction.

An edge of the minimum spanning tree (MST) is a correlation that belongs to the \(N-1\) strongest links needed to connect all \(N\) coins; it also has no direction. A spillover \(\tilde\theta_{ij}\) in the sense of Diebold and Yilmaz is the share of the forecast-error variance of coin \(i\) over the next \(H\) days that is due to shocks in coin \(j\); it has a direction, from \(j\) to \(i\).

In the rest of the lecture, "link" means a correlation or an MST edge, and "spillover" means a Diebold–Yilmaz share. None of the four measures shows that one coin causes another to move. Each describes a statistical co-movement, which can come from a common driver, such as news that affects the whole market.

In brief

measurea link \(i\)–\(j\) meansdirection, part of the distribution
tail link (Lecture 6)coin \(j\) has a nonzero slope in the 5% Quantile Lasso of coin \(i\)\(j \rightarrow i\); lower tail only
correlation \(\rho_{ij}\)returns move together on the same daynone (symmetric); whole distribution
MST edge\(\rho_{ij}\) is among the \(N-1\) strongest links that connect all coinsnone; whole distribution
spillover \(\tilde\theta_{ij}\) (Diebold–Yilmaz)share of \(i\)'s forecast-error variance due to shocks in \(j\)\(j \rightarrow i\); variance over \(H\) days
  • In this lecture, “link” means a correlation or an MST edge; “spillover” means a Diebold–Yilmaz share
  • None of the four measures shows that one coin causes another to move

Building a Network from Correlations

Slide 8 in the PDF

The network is built from the daily log returns of 19 coins over 2,121 days, January 2021 – October 2026. The Pearson correlation \(\rho_{ij}\) between two return series measures their linear co-movement, from \(-1\) to 1. It is not a distance: a higher correlation means two coins are closer. Mantegna (1999) therefore transforms it into a correlation distance,

\[d_{ij} = \sqrt{2\,(1 - \rho_{ij})}\]
(7.1)

which is a proper distance (a metric): it is 0 when \(\rho = 1\), \(\sqrt 2 \approx 1.41\) when \(\rho = 0\) and 2 when \(\rho = -1\).

With 19 coins there are 171 pairs, too many links to read. The minimum spanning tree keeps only \(N - 1 = 18\) of them: the set of edges that connects all coins, contains no loop, and has the smallest total distance,

\[\mathrm{MST} = \arg\min_{T \text{ spanning tree}} \sum_{(i,j) \in T} d_{ij}\]
(7.2)

The tree keeps the strongest links and removes the rest. A coin with many edges in the tree, that is with a high degree, is a hub: many other coins are attached to the market through it.

Two examples show the scale. Bitcoin (BTC) and Ether (ETH) have \(\rho = 0.823\), so formula (7.1) gives \(d = \sqrt{2 \times 0.177} = 0.595\), a short distance. Dogecoin (DOGE) and Tron (TRX) have \(\rho = 0.349\), the lowest correlation in the sample, and \(d = 1.141\), close to the distance of two uncorrelated coins.

In brief

  • Nodes: 19 coins; edges: dependence of daily log returns, January 2021 – October 2026 (EODHD, 2 121 days)
  • Pearson correlation \(\rho_{ij}\) becomes a metric distance (Mantegna, 1999):
    \[ d_{ij} = \sqrt{2\,(1 - \rho_{ij})}, \qquad \rho = 1 \Rightarrow d = 0, \quad \rho = 0 \Rightarrow d = \sqrt2 \]
  • Minimum spanning tree (MST): the \(N - 1\) edges that connect all coins with the smallest total distance
    \[ \mathrm{MST} = \arg\min_{T \text{ spanning tree}} \sum_{(i,j) \in T} d_{ij} \]
    • keeps the strongest links and removes the rest; a hub = a coin with many MST edges (high degree)
  • Worked example: BTC–ETH, \(\rho = 0.823\): \(d = \sqrt{2 \cdot 0.177} = 0.595\); DOGE–TRX, \(\rho = 0.349\): \(d = 1.141\)

Illustration of a Correlation of 0.82

Slide 9 in the PDF

A correlation coefficient is easier to judge when one sees the data behind it. Each dot in the two panels is one day between January 2021 and October 2026: its horizontal position is the daily log return of Bitcoin, its vertical position the return of a second coin. The left panel uses Ether, the coin most correlated with Bitcoin; the right panel uses Tron (TRX), the least correlated. The red line is the least-squares line of the second coin's return on Bitcoin's.

For Bitcoin and Ether the correlation is \(\rho = 0.82\) and the cloud is narrow. The slope of 1.11 means that Ether moves on average by 1.11% for each 1% move of Bitcoin, so Ether amplifies the market's moves. Both coins move in the same direction on 82% of the days. For Bitcoin and Tron, \(\rho = 0.52\): the cloud is wide, the slope is 0.67, and the two coins move in the same direction on 69% of the days. The square of the correlation is the share of the second coin's daily variance that a straight line in Bitcoin's return explains: 0.68 for Ether, but only 0.27 for Tron.

The two correlations become the distances of (7.1): \(d = \sqrt{2 \times 0.177} = 0.60\) for Bitcoin and Ether and \(d = \sqrt{2 \times 0.479} = 0.98\) for Bitcoin and Tron. The minimum spanning tree is built from the shortest of these distances, so a pair like Bitcoin and Ether is a natural edge, while Tron is attached to the tree through whichever coin is closest to it.

In brief

Illustration of a Correlation of 0.82
  • Each dot is one day, January 2021 – October 2026; left: BTC and ETH, the highest correlation with BTC; right: BTC and TRX, the lowest
  • BTC–ETH: \(\rho = 0.82\); ETH moves by 1.11% on average per 1% move of BTC (least-squares slope); both coins move in the same direction on 82% of days
  • BTC–TRX: \(\rho = 0.52\), slope 0.67, same direction on 69% of days; \(\rho^2 = 0.27\): BTC explains about a quarter of TRX's daily variance
  • Distances: \(d = 0.60\) for BTC–ETH and \(d = 0.98\) for BTC–TRX; the MST prefers the short ones
  • Source: own computation on EODHD daily prices of 19 coins, January 2021 – 22 October 2026
BA_L7_pairs

Correlation Matrix and Minimum Spanning Tree

Slide 10 in the PDF
    Question for the room The ▶ button builds the correlation matrix and the minimum spanning tree for each calendar year, 2021 to 2026. Which coin is the hub (the coin with the most MST edges) in most of these years?
    Answer

    Polkadot (DOT): it is the hub in 2022 (degree 5), 2023 (degree 9), 2024 (5) and 2025 (5). Litecoin is the hub in 2021 (5); in 2026, to 23 October, Bitcoin, Ether and Ethereum Classic share the highest degree (4). Over the full sample ETH and DOT tie with degree 6. The MST keeps the 18 strongest links that connect all coins, so a hub is a coin that is the nearest neighbour of many others, not the largest coin: Bitcoin has degree 1 in 2021 and 2 in 2025. The mean pairwise correlation is highest in 2022 (0.73) and lowest in 2024 (0.55).

    The left panel is a heat map of the 19 × 19 correlation matrix of daily log returns, January 2021 – October 2026: darker red means a higher correlation. Almost the whole matrix is red. The mean pairwise correlation is 0.60, the lowest is 0.35 (DOGE–TRX), and the correlation with Bitcoin ranges from 0.52 for Tron to 0.82 for Ether. Lighter rows and columns, such as those of DOGE, TRX and Monero (XMR), mark the coins that move most independently of the others.

    The right panel draws the minimum spanning tree. Each circle is a coin, its size grows with its degree, and Bitcoin and Ether are in red. The tree has two hubs with six edges each. Ether is linked to Bitcoin, BNB, Solana (SOL), Litecoin (LTC), Chainlink (LINK) and Ethereum Classic (ETC); Polkadot (DOT) is linked to Cardano (ADA), Cosmos (ATOM), Filecoin (FIL), Algorand (ALGO), Tron and Chainlink. Chainlink connects the two hubs.

    Bitcoin itself has only two edges, to Ether and to Monero. A low degree does not make a coin unimportant: Bitcoin is highly correlated with every coin, but the tree keeps only the single strongest path to each coin, and for most coins that path runs through Ether or Polkadot. The spillover table below gives a different view of Bitcoin's role.

    In brief

    • Mean pairwise correlation 0.60 (minimum 0.35, DOGE–TRX); correlation with BTC from 0.52 (TRX) to 0.82 (ETH)
    • Two hubs: ETH (degree 6: BTC, BNB, SOL, LTC, LINK, ETC) and DOT (degree 6: ADA, ATOM, FIL, ALGO, TRX, LINK); BTC links only to ETH and XMR
    • Source: own computation on EODHD daily prices of 19 coins, January 2021 – 22 October 2026
    BA_L7_mst

    The Absorption Ratio: a Worked Example

    Slide 11 in the PDF

    A tree shows the structure of the links; a single number for their overall strength comes from principal components. The principal components of the correlation matrix of \(N\) coins are uncorrelated portfolios of the standardised returns. The eigenvalue \(\lambda_k\) is the variance of the \(k\)-th component; the eigenvalues are ordered, \(\lambda_1 \geq \lambda_2 \geq \dots \geq \lambda_N \geq 0\), and add up to \(N\), the total variance of \(N\) standardised series.

    The absorption ratio (AR) is the share of the total variance taken by the first component:

    \[\mathrm{AR} = \frac{\lambda_1}{\sum_{k=1}^N \lambda_k} = \frac{\lambda_1}{N}\]
    (7.3)

    It lies between \(1/N\), when the coins are uncorrelated and no component dominates, and 1, when one factor drives all of them.

    A simple case shows how the ratio depends on the correlations. If every pair of the \(N = 19\) coins has the same correlation \(\rho\), the first eigenvalue is \(\lambda_1 = 1 + (N-1)\rho\) and the other 18 eigenvalues equal \(1 - \rho\). With \(\rho = 0.60\), close to the mean correlation of the sample, \(\lambda_1 = 1 + 18 \times 0.60 = 11.8\) and \(\mathrm{AR} = 11.8/19 = 0.62\). With \(\rho = 0.85\), \(\lambda_1 = 16.3\) and \(\mathrm{AR} = 0.86\). With \(\rho = 0\), \(\mathrm{AR} = 1/19 = 0.05\): there is no common factor.

    A high absorption ratio means that one common factor, in practice the movement of the whole crypto market, drives most coins. Spreading money over many coins then reduces risk only a little, because most of the variance is common to all of them.

    In brief

    • Principal components of a correlation matrix of \(N\) coins: uncorrelated portfolios of the standardised returns
      • eigenvalue \(\lambda_k\) = variance of the \(k\)-th component; \(\lambda_1 \geq \lambda_2 \geq \dots \geq \lambda_N \geq 0\) and \(\sum_k \lambda_k = N\)
    • Absorption ratio (AR): share of the total variance taken by the first component, between \(1/N\) and 1
      \[ \mathrm{AR} = \frac{\lambda_1}{\sum_{k=1}^N \lambda_k} = \frac{\lambda_1}{N} \]
    • Worked example: \(N = 19\) coins, every pairwise correlation equal to \(\rho\)
      • \(\lambda_1 = 1 + (N-1)\rho\); the other 18 eigenvalues equal \(1 - \rho\)
      • \(\rho = 0.60\): \(\lambda_1 = 1 + 18 \cdot 0.60 = 11.8\), \(\;\mathrm{AR} = 11.8/19 = 0.62\)
      • \(\rho = 0.85\): \(\lambda_1 = 16.3\), \(\;\mathrm{AR} = 0.86\)
      • \(\rho = 0\): \(\mathrm{AR} = 1/19 = 0.05\), no common factor
    • A high AR means one common factor drives most coins: diversification across coins is weak

    Eigenvalues of the Correlation Matrix and the Absorption Ratio

    Slide 12 in the PDF
      Question for the room The ▶ button computes the eigenvalues of the correlation matrix for each calendar year, 2021 to 2026. In which year does the first component take the largest share of the variance?
      Answer

      2022: λ₁ = 14.12 and the absorption ratio λ₁/19 is 0.743, against 0.592 in 2023, the lowest year. 2022 is the year of the Terra and FTX collapses, when the mean pairwise correlation reached 0.725. In every year the formula (1 + 18ρ̄)/19 from the mean correlation is within 0.021 of the absorption ratio, so the year with the highest mean correlation is also the year with the strongest common factor. Only one eigenvalue exceeds 1 in 2022 and 2025; in 2024 three do (λ₂ = 1.27), a sign of a second, weaker factor.

      The worked example assumed that all pairs have the same correlation. The left panel checks this idea on the actual correlation matrix of the 19 coins, January 2021 – October 2026. The dark blue bars are its 19 eigenvalues, from the largest to the smallest; the orange dots are the eigenvalues of the equicorrelation matrix with the same mean correlation, \(\bar\rho = 0.60\).

      The first eigenvalue is \(\lambda_1 = 11.84\); the second is only 0.79, and no other eigenvalue exceeds 1. The equicorrelation formula predicts \(\lambda_1 = 1 + 18 \times 0.60 = 11.74\) and \(1 - 0.60 = 0.40\) for each of the others, close to the actual values. The absorption ratio of the full sample is \(11.84/19 = 0.62\): one common factor carries almost two thirds of the variance of 19 standardised return series, and the remaining 18 components share the rest.

      The right panel repeats the calculation for each calendar year. The absorption ratio was lowest in 2023 (0.59) and highest in 2022 (0.74), the year of the Terra and FTX collapses, when the whole market fell together. In every year the simple formula \((1 + 18\bar\rho)/19\), which uses only the mean correlation of the year, is within 0.021 of the absorption ratio. For this market the mean pairwise correlation is therefore a good summary of how strong the common factor is.

      In brief

      • Left: the 19 eigenvalues of the full-sample correlation matrix. \(\lambda_1 = 11.84\), the next one 0.79; only \(\lambda_1\) exceeds 1
      • The equicorrelation formula with the mean correlation \(\bar\rho = 0.60\) predicts \(\lambda_1 = 11.74\) and 0.40 for each of the others
      • Right: the absorption ratio by calendar year, from 0.59 (2023) to 0.74 (2022, the year of Terra and FTX)
      • \((1 + 18\bar\rho)/19\) is within 0.021 of the absorption ratio in every year: the mean correlation already tells how strong the common factor is
      • Source: own computation on EODHD daily prices of 19 coins, January 2021 – 22 October 2026
      BA_L7_eigen

      Rolling Correlation and Absorption Ratio, 2021–2026

      Slide 13 in the PDF
        Question for the room The ▶ button lengthens the window from 30 to 360 days. What happens to the highest mean pairwise correlation over all windows?
        Answer

        It falls, from 0.90 (30-day window ending 15 June 2021) to 0.85 (90 days, 14 August 2021) and 0.75 (360 days). A long window averages the days after a crash with calmer days, so the peaks of co-movement are smoothed away; the lowest value rises from 0.28 to 0.49. The average absorption ratio hardly moves, from 0.69 to 0.67: the window changes how sharply the episodes show, not the overall level of connectedness.

        The chart computes both measures on rolling windows: for each date, the correlation matrix of the previous 90 days of daily log returns of the 19 coins gives the mean pairwise correlation (dark blue) and the absorption ratio \(\mathrm{AR}_t\) (orange). The last window ends on 22 October 2026.

        The mean correlation was lowest, 0.36, in April 2021 and highest, 0.85, in mid-August 2021, after the market crash of May 2021, when all coins fell together. In the last window it was 0.54. The absorption ratio averaged 0.68 over all windows and was 0.59 in the last one. The two lines move almost in parallel, as the worked example suggests: when the average correlation rises, so does the share of the first component.

        For an investor the practical reading is simple. When one factor explains around 80% of the variance, as in parts of 2021 and 2022, holding many coins instead of one adds very little diversification; the benefit is largest in the periods when the lines are low.

        In brief

        • 90-day windows of daily log returns, 19 coins; the last window ends on 22 October 2026
        • Mean pairwise correlation: from 0.36 (April 2021) to 0.85 (mid-August 2021, after the May crash); 0.54 in the last window
        • Absorption ratio \(\mathrm{AR}_t\): 0.68 on average over all windows, 0.59 in the last window
        • When one factor explains 80% of the variance, “diversifying across coins” buys almost nothing
        • Source: own computation on EODHD daily prices of 19 coins, January 2021 – 22 October 2026
        BA_L7_connect

        Spillovers: Diebold–Yilmaz

        Slide 14 in the PDF

        Correlations are symmetric and describe the same day. Diebold & Yilmaz (2012) measure instead how shocks to one coin affect the future returns of the others. The starting point is a vector autoregression (VAR) of order \(p\): each of the \(N\) return series is regressed on \(p\) past values of all series,

        \[\mathbf r_t = \sum_{l=1}^{p}\Phi_l \mathbf r_{t-l} + \boldsymbol\varepsilon_t\]
        (7.4)

        where \(\mathbf r_t\) is the vector of returns on day \(t\), \(\Phi_l\) are \(N \times N\) coefficient matrices and \(\boldsymbol\varepsilon_t\) are the shocks, with covariance matrix \(\Sigma\). Rewritten as a moving average (MA), \(\mathbf r_t = \sum_{h \geq 0} A_h \boldsymbol\varepsilon_{t-h}\), the model shows how a shock today feeds into the returns of the next days.

        The forecast-error variance decomposition splits the uncertainty of an \(H\)-day forecast of coin \(i\) into parts due to the shocks of each coin \(j\). The generalised version of Pesaran and Shin, used by Diebold and Yilmaz, does not depend on the order in which the variables enter the model:

        \[\theta_{ij}(H) = \frac{\sigma_{jj}^{-1}\sum_{h=0}^{H-1}\big(\mathbf e_i' A_h \Sigma \mathbf e_j\big)^2}{\sum_{h=0}^{H-1}\mathbf e_i' A_h \Sigma A_h' \mathbf e_i}, \qquad \tilde\theta_{ij} = \frac{\theta_{ij}}{\sum_j \theta_{ij}}\]
        (7.5)

        Here \(\mathbf e_i\) is a vector with 1 in position \(i\) and 0 elsewhere, and \(\sigma_{jj}\) is the variance of the shocks to coin \(j\). The raw shares \(\theta_{ij}\) need not add up to 1 across \(j\), so each row is normalised: \(\tilde\theta_{ij}\) is the share of coin \(i\)'s forecast-error variance due to shocks in coin \(j\).

        Four indices summarise the table of shares, in percent:

        \[S = \frac{100}{N}\sum_{i \neq j}\tilde\theta_{ij}, \quad \mathrm{TO}_j = 100\sum_{i \neq j}\tilde\theta_{ij}, \quad \mathrm{FROM}_i = 100\sum_{j \neq i}\tilde\theta_{ij}, \quad \mathrm{NET}_i = \mathrm{TO}_i - \mathrm{FROM}_i\]
        (7.6)

        The total spillover index \(S\) is the average share of forecast-error variance that comes from other coins. \(\mathrm{TO}_j\) is what coin \(j\) sends to all others, \(\mathrm{FROM}_i\) what coin \(i\) receives from all others, and \(\mathrm{NET}_i\) the difference: a positive NET marks a net transmitter of shocks, a negative NET a net receiver.

        In brief

        • VAR(\(p\)) on \(N\) return series, \(\mathbf r_t = \sum_{l=1}^{p}\Phi_l \mathbf r_{t-l} + \boldsymbol\varepsilon_t\), moving-average (MA) form \(\mathbf r_t = \sum_{h \geq 0} A_h \boldsymbol\varepsilon_{t-h}\)
        • Generalised forecast-error variance decomposition (Pesaran–Shin), horizon \(H\):
          \[ \theta_{ij}(H) = \frac{\sigma_{jj}^{-1}\sum_{h=0}^{H-1}\big(\mathbf e_i' A_h \Sigma \mathbf e_j\big)^2}{\sum_{h=0}^{H-1}\mathbf e_i' A_h \Sigma A_h' \mathbf e_i}, \qquad \tilde\theta_{ij} = \frac{\theta_{ij}}{\sum_j \theta_{ij}} \]
          • \(\tilde\theta_{ij}\): share of coin \(i\)'s forecast-error variance due to shocks in coin \(j\); independent of variable ordering
        • Indices, in % (\(\tilde\theta_{ij}\) is a share between 0 and 1, so each sum is multiplied by 100):
          \[ S = \frac{100}{N}\sum_{i \neq j}\tilde\theta_{ij}, \quad \mathrm{TO}_j = 100\sum_{i \neq j}\tilde\theta_{ij}, \quad \mathrm{FROM}_i = 100\sum_{j \neq i}\tilde\theta_{ij}, \quad \mathrm{NET}_i = \mathrm{TO}_i - \mathrm{FROM}_i \]
          • TO\(_j\): what coin \(j\) sends to all others; FROM\(_i\): what coin \(i\) receives from all others

        Spillover Table, 8 Coins, 2021–2026

        Slide 15 in the PDF

        The table applies formulas (7.4) – (7.6) to eight large coins: Bitcoin, Ether, XRP, BNB, Solana, Cardano, Dogecoin and Litecoin, with daily log returns from 2021 to 2026. The lag order \(p = 5\) is chosen by the Akaike information criterion (AIC), a rule that balances the fit of the model against the number of its parameters, and the forecast horizon is \(H = 10\) days.

        Each row is a receiver of shocks and each column a source. A cell is read as a percentage: in row BTC, column ETH, the value 16.0 means that 16% of the 10-day forecast-error variance of Bitcoin is due to shocks in Ether. The diagonal is the part due to a coin's own shocks; for Bitcoin it is 23.3%. The FROM column adds up each row without the diagonal, and the TO row adds up each column without the diagonal.

        The total spillover index is the mean of the FROM column, \(S = 72.9\%\): on average almost three quarters of a coin's forecast uncertainty comes from the other seven coins, a highly connected system. The NET row shows who transmits. Ether has \(\mathrm{NET} = 93.2 - 77.5 = +15.7\) and Bitcoin \(+11.0\), so they are the largest net transmitters; Cardano and Litecoin are small net transmitters (\(+5.0\) each).

        Dogecoin is the largest net receiver, with \(\mathrm{NET} = -18.6\). This is not because it receives much: its FROM value, 63.6, is the lowest in the table, and 36.4% of its variance comes from its own shocks. It is a net receiver because it sends far less, \(\mathrm{TO} = 45.0\). This answers part of the question raised by the tree: Bitcoin has few edges in the MST but is the second-largest transmitter of shocks.

        In brief

        receiver \(\downarrow\) / source \(\rightarrow\)BTCETHXRPBNBSOLADADOGELTCFROM
        BTC23.316.09.410.610.011.37.312.076.7
        ETH15.322.59.510.510.911.86.912.577.5
        XRP11.211.828.19.08.613.05.912.471.9
        BNB12.713.38.928.410.410.04.911.371.6
        SOL12.113.88.510.528.311.65.79.471.7
        ADA12.113.211.58.710.325.07.212.075.0
        DOGE11.411.17.36.07.110.436.410.263.6
        LTC12.814.011.09.98.311.97.125.075.0
        TO87.793.266.165.465.780.045.080.0\(S = 72.9\)
        NET\(+11.0\)\(+15.7\)\(-5.8\)\(-6.3\)\(-6.0\)\(+5.0\)\(-18.6\)\(+5.0\)
        • VAR(5) by Akaike information criterion (AIC), \(H = 10\) days. Row = receiver of shocks; column = source
        • Reading a cell: row BTC, column ETH \(= 16.0\): 16% of BTC's 10-day forecast-error variance is due to ETH shocks (ETH \(\rightarrow\) BTC)
        • Worked check: \(\mathrm{NET}_{ETH} = 93.2 - 77.5 = +15.7\); \(S\) = mean of the FROM column \(= 72.9\%\): a highly connected system
        • ETH and BTC are the largest net transmitters. DOGE is the largest net receiver (\(-18.6\)), although it receives the least (FROM 63.6): it sends far less (TO 45.0)
        • Source: own computation on EODHD daily prices of 8 coins, January 2021 – October 2026

        Spillovers through Time

        Slide 16 in the PDF
          Question for the room The ▶ button raises the forecast horizon H from 1 to 20 days. How does the total spillover index change?
          Answer

          It hardly changes: for the full-sample VAR(5) it is 72.7% at H = 1, 72.9% at H = 10 and 72.9% at H = 20; in the last 200-day window it moves from 80.3% to 80.4%. The lag coefficients of daily crypto returns are small, so almost all of the forecast-error variance comes from the same-day correlation of the shocks (Σ), which is already in the H = 1 decomposition. A high spillover index here measures common same-day moves, not shocks that travel from coin to coin over the following days.

          The left panel re-estimates the total spillover index on rolling windows of 200 days and \(H = 10\), with a VAR of order 2. The dotted lines mark the collapse of the TerraUSD stablecoin (May 2022), the bankruptcy of the FTX exchange (November 2022) and the launch of spot Bitcoin exchange-traded funds (ETFs) in January 2024. The right panel repeats the NET spillovers of the full-sample table: green bars are net transmitters, red bars net receivers.

          The index ranged from 65.2% in July 2021 to 83.5% in May 2026, and was 80.4% in the last window, ending on 22 October 2026. It rose into the Terra and FTX episodes and stayed above 80% for most of 2022. It fell during 2023 and into early 2024, close to its minimum of 2021, and from the spring of 2024 it climbed to the highest values of the sample.

          The rise after the launch of the ETFs is consistent with more common, institutional flows into the market, which would make coins react to the same news. The chart alone does not identify the cause: the timing matches, but other events of 2024–2026 could produce the same pattern.

          In brief

          • Rolling VAR(2), 200-day windows: from 65.2% (July 2021) to 83.5% (May 2026); 80.4% in the last window, ending 22 October 2026
          • Spillovers rose into Terra and FTX, fell during 2023, and climbed after the spot exchange-traded funds (ETFs)
          • The rise after the ETFs is consistent with more common (institutional) flows; the chart alone does not identify the cause
          • Source: own computation on EODHD daily prices of 19 coins, January 2021 – 22 October 2026
          BA_L7_spill

          Section summary

          Crypto returns form a dense network: the mean pairwise correlation is 0.60, Ether and Polkadot are the hubs of the minimum spanning tree, and one common factor explains about two thirds of the variance on average. Spillovers show direction: about 73% of a coin's forecast-error variance comes from other coins, Ether and Bitcoin are the largest net transmitters and Dogecoin the largest net receiver.

          Cryptocurrencies as an Asset Class

          An asset class is a group of assets with similar risk and return behaviour, such as equities or government bonds. The section asks whether crypto assets form such a group, using the statistical features of their returns rather than their technology. It first builds a small map of 33 assets for 2021–2026 and then reads the larger study of Pele, Wesselhöfft, Härdle, Kolossiatis & Yatracos (2023) on 906 assets.

          Classifying Assets by Their Statistical “Phenotype”

          Slide 17 in the PDF

          The idea of Pele, Wesselhöfft, Härdle, Kolossiatis & Yatracos (2023) is to describe every asset by the stylised facts of its returns, its statistical "phenotype", and then to ask whether crypto assets form their own group. The lecture applies a small version of this idea to 33 assets: 19 coins, 6 equities, a Treasury bond fund, gold, silver, oil and 4 foreign-exchange (FX) rates. The data are daily log returns from 2021 to 2026. Each asset is measured on its own trading days: coins trade every day, while gold, FX and the S&P 500 stock index trade from Monday to Friday. A closing price repeated over a holiday would add returns of exactly zero that are not market moves and would understate volatility.

          Each asset is described by six features: the annualised volatility; the skewness, which measures asymmetry; the logarithm of the kurtosis, which measures how heavy the tails are; the Hill tail index of the top 5% of returns, where a smaller value means heavier tails (Lecture 2); the lag-1 autocorrelation of absolute returns, \(\mathrm{ACF}|r_t|\), a measure of volatility clustering; and the 1% quantile of returns, in standard-deviation units. Each feature is standardised across the 33 assets to mean 0 and standard deviation 1, so that no feature dominates because of its units. Volatility remains one of the six features.

          A principal component analysis of the six standardised features gives a first component (PC1) that explains 62.7% of their variance and a second (PC2) that explains 18.5%. The loadings of PC1, the weights of the standardised features, are 0.40 for volatility, 0.39 for skewness, 0.46 for kurtosis, \(-0.44\) for the tail index, 0.46 for \(\mathrm{ACF}|r|\) and 0.27 for the 1% quantile. All features point in the same direction: higher volatility, more skew, heavier tails and more clustering. PC1 is therefore a "wildness" factor.

          The assets are then grouped by \(k\)-means clustering, which splits them into \(k\) groups so that each asset is close to the centre of its group. With \(k = 3\), the first cluster contains all 14 traditional assets and four coins: Bitcoin, Ether, Litecoin and Chainlink. The second contains the other 14 coins, and the third contains Dogecoin alone, whose skewness of 5.9 is far from every other asset.

          In brief

          • Idea of Pele et al. (2023): describe each asset by the stylised facts of its returns, then ask whether cryptos form their own group
          • Six features per asset, from daily log returns 2021–2026: annualised volatility, skewness, log kurtosis, Hill tail index (top 5%), lag-1 autocorrelation (ACF) of \(|r_t|\), 1% quantile in sd units
          • 33 assets: 19 coins, 6 equities, Treasuries, gold, silver, oil, 4 foreign exchange (FX) rates
          • Each asset is measured on its own trading days (coins every day, traditional assets Monday–Friday): a stale holiday close would add false zero returns
          • Each feature is standardised across the 33 assets (mean 0, sd 1), so that no feature dominates through its units
          • Principal components: first (PC1) explains 62.7%, second (PC2) 18.5%
            • PC1 loadings (weights of the standardised features): vol 0.40, skew 0.39, kurtosis 0.46, tail index \(-0.44\), ACF\(|r|\) 0.46, 1% quantile 0.27: a “wildness” factor
          • \(k\)-means, \(k = 3\):
            • cluster 1: all 14 traditional assets and BTC, ETH, LTC, LINK
            • cluster 2: 14 other coins; cluster 3: DOGE alone (skewness 5.9)
          • Source: own computation on EODHD daily prices of 19 coins and 14 traditional assets, January 2021 – October 2026

          Volatility and Tails of 33 Assets

          Slide 18 in the PDF

          Before the six features are combined, it helps to look at two of them directly. The chart places each asset by its annualised volatility (horizontal axis) and its Hill tail index (vertical axis), computed from daily log returns 2021–2026. A lower tail index means a heavier tail, that is, more frequent extreme returns. The colours give the conventional asset class.

          Volatility separates the groups completely. Every coin has an annualised volatility above 56%: Bitcoin is the calmest coin at 57%, Dogecoin the most volatile at 132%, and the median coin is at 95%. Every traditional asset is below 38%; the most volatile is the silver fund at 37%, and the exchange rates are near 7–9%.

          The tail index does not separate them. The coins range from 2.12 (Tron) to 3.66 (Chainlink), the traditional assets from 3.07 (silver) to 5.10 (Exxon), so the two ranges overlap. Bitcoin's tail index, 3.56, is even slightly higher than the 3.44 of the S&P 500: in this sample its tail is, if anything, lighter. This is why Bitcoin ends up next to the equities on the map once all six features are standardised: it is far apart from them only on volatility.

          In brief

          Volatility and Tails of 33 Assets
          • Two of the six features, daily log returns 2021–2026: annualised volatility (horizontal) and Hill tail index of the top 5% of \(|r_t|\) (vertical; lower = heavier tail)
          • Volatility separates the groups: every coin is above 56% a year (BTC 57%, DOGE 132%; median coin 95%); every traditional asset is below 38% (silver)
          • Tails overlap: coins 2.12 (TRX) to 3.66 (LINK), traditional assets 3.07 (silver) to 5.10 (Exxon); BTC (3.56) has a lighter tail than the S&P 500 (3.44)
          • Source: own computation on EODHD daily prices of 19 coins and 14 traditional assets, January 2021 – October 2026
          BA_L7_features

          Classification Map of 33 Assets

          Slide 19 in the PDF
            Question for the room The ▶ button raises the weight of volatility among the six standardised features from 0 (shape only) to 1 (all six features, as on the slide). How does the number of coins in the k-means group that holds the 14 traditional assets change?
            Answer

            It falls, from 8 coins to 4. With weight 0 to 0.15 the group holds BTC, ETH, LTC, LINK, DOT, AVAX, ATOM and ALGO; from 0.2 to 0.45, seven (AVAX leaves); from 0.5 on, only BTC, ETH, LTC and LINK. All 14 traditional assets stay in the same group at every weight. Volatility separates most altcoins from traditional assets; on the shape of the returns alone, the large coins resemble equities. PC1 explains 62.7% of the variance with all six features and 64.5% with shape only.

            The chart places each of the 33 assets on the first two principal components: the horizontal axis is PC1, the "wildness" factor (63% of the variance), the vertical axis is PC2 (19%). The colours give the conventional class: crypto in orange, equities in dark blue, the bond fund in green, commodities in brown and FX rates in purple.

            The traditional assets lie on the left, at negative PC1. Most coins form a cloud to the right of zero, and Dogecoin sits alone at the far right, near 6.5. Ether, Litecoin and Chainlink lie just left of zero, and Bitcoin lies among the traditional assets, next to the S&P 500. Its volatility of 57% a year is far above the 16% of the S&P 500, but volatility is one feature of six: its Hill tail index, 3.56, is close to the 3.44 of the S&P 500, and on the other five features, too, Bitcoin resembles equities.

            A second version of the map drops volatility and keeps only the five features of shape. With the same method, Bitcoin, Ether, Litecoin, Chainlink, Polkadot, Avalanche (AVAX), Cosmos and Algorand join the cluster of all 14 traditional assets; ten other coins form a second cluster and Dogecoin a third. The large coins thus resemble traditional assets in the shape of their returns, while most smaller coins remain a separate group.

            In brief

            • Bitcoin (vol 57%, Hill tail index 3.56) sits next to the S&P 500 (vol 16%, tail index 3.44): volatility is one feature of six, and on the other five BTC resembles equities
            • Shape only (volatility dropped, same method): BTC, ETH, LTC, LINK, DOT, AVAX, ATOM and ALGO join the cluster of all 14 traditional assets; 10 altcoins form a second cluster, DOGE a third
            • Most altcoins remain a separate cloud; DOGE is an outlier of its own
            • Source: own computation on EODHD daily prices of 19 coins and 14 traditional assets, January 2021 – October 2026
            BA_L7_classes

            Worked Example: Where Bitcoin Sits on PC1

            Slide 20 in the PDF

            The score of an asset on the first principal component is a weighted sum of its standardised features. The worked example computes it for Bitcoin by hand, in three steps.

            The first step standardises each feature across the 33 assets, \(z = (x - \text{mean})/\text{sd}\). Bitcoin's annualised volatility is 56.7%, the mean of the 33 assets is 62.3% and their standard deviation is 38.9%, so \(z = (56.7 - 62.3)/38.9 = -0.14\): Bitcoin is slightly less volatile than the average asset of this sample, because 19 of the 33 assets are coins. Its Hill tail index is 3.56 against a mean of 3.22 and a standard deviation of 0.61, so \(z = +0.57\). The other four \(z\)-values are \(-0.26\) for skewness, \(-0.69\) for log kurtosis, \(-0.50\) for the autocorrelation of \(|r_t|\) and \(-0.61\) for the 1% quantile.

            The second step multiplies each \(z\) by its PC1 loading and adds the products. The loadings are 0.40, 0.39, 0.46, \(-0.44\), 0.46 and 0.27; the products are \(-0.06\), \(-0.10\), \(-0.31\), \(-0.25\), \(-0.23\) and \(-0.16\), and their sum is \(-1.11\). The tail index enters with a negative loading, so Bitcoin's lighter-than-average tail lowers its score, as does its low kurtosis.

            The third step compares this score with those of other assets: the S&P 500 scores \(-1.25\), Ether \(-0.57\) and Dogecoin \(+6.46\). Bitcoin ranks 23rd of 33 on the "wildness" factor, closer to the S&P 500 than to Ether. The largest contributions come from kurtosis and the tail index, two features of shape: on the shape of its returns, Bitcoin is calmer than the average asset of the sample.

            In brief

            • Step 1, standardise each feature across the 33 assets: \(z = (x - \text{mean})/\text{sd}\)
              • volatility: \((56.7 - 62.3)/38.9 = -0.14\); Hill tail index: \((3.56 - 3.22)/0.61 = +0.57\)
              • the other four \(z\)-values of BTC: skewness \(-0.26\), log kurtosis \(-0.69\), ACF\(|r|\) \(-0.50\), 1% quantile \(-0.61\)
            • Step 2, weight each \(z\) by its PC1 loading and add: \(\;\mathrm{PC1} = \sum_k w_k z_k\)
            volskewkurtosistail indexACF\(|r|\)\(Q_{1\%}\)sum
            loading \(w_k\)0.400.390.46\(-0.44\)0.460.27
            \(z_k\) of BTC\(-0.14\)\(-0.26\)\(-0.69\)\(0.57\)\(-0.50\)\(-0.61\)
            \(w_k z_k\)\(-0.06\)\(-0.10\)\(-0.31\)\(-0.25\)\(-0.23\)\(-0.16\)\(\mathbf{-1.11}\)
            • Step 3, compare: S&P 500 \(-1.25\), ETH \(-0.57\), DOGE \(+6.46\); BTC ranks 23rd of 33 on the “wildness” factor
            • Kurtosis and the tail index contribute most: on shape, BTC is calmer than the average asset of the sample, which is why it sits next to equities
            • Source: own computation on EODHD daily prices of 19 coins and 14 traditional assets, January 2021 – October 2026

            Reading the Indicators of Pele et al. (2023)

            Slide 21 in the PDF

            The study of Pele, Wesselhöfft, Härdle, Kolossiatis & Yatracos (2023) uses several indicators that have not appeared earlier in the course. The stable \(\alpha \in (0, 2]\) is the tail parameter of a stable distribution fitted to all returns. The value \(\alpha = 2\) gives the Normal distribution; values below 2 mean infinite variance, and a smaller \(\alpha\) means heavier tails. The Hill tail index of Lecture 2, by contrast, is estimated from the largest returns only and is usually between 2 and 4 for financial returns. The two come from different models and different parts of the distribution, so their values cannot be compared.

            The Hurst exponent \(H \in (0, 1)\) measures memory in returns: 0.5 means no memory, values above 0.5 persistent trends and values below 0.5 mean reversion. The FIGARCH \(d \in [0, 1]\), from the fractionally integrated generalised autoregressive conditional heteroskedasticity (GARCH) model, measures memory in volatility. With \(d = 0\) the model is a GARCH, in which shocks to volatility fade geometrically; a larger \(d\) means that they fade more slowly, hyperbolically.

            Three statistics describe the analysis itself. A factor loading is the correlation between an indicator and a factor, from \(-1\) to 1; an indicator "belongs" to the factor on which it loads most. The Kaiser–Meyer–Olkin (KMO) measure, between 0 and 1, is the share of the correlation among the indicators that is common to all of them; values above 0.8 mean that the data are well suited to factor analysis. The Nagelkerke \(\tilde R^2\), between 0 and 1, is a pseudo-\(R^2\) for a logit model, rescaled so that 1 means perfect classification.

            In brief

            • Stable \(\alpha \in (0, 2]\): tail parameter of a stable distribution fitted to all returns; \(\alpha = 2\) is the Normal, \(\alpha < 2\) means infinite variance; smaller = heavier tails
            • Hill tail index (Lecture 2): estimated from the largest returns only; usually 2–4 for financial returns. Stable \(\alpha\) and Hill index come from different models and parts of the distribution: their values are not comparable
            • Hurst exponent \(H \in (0, 1)\): memory in returns; \(0.5\) = none, above 0.5 = persistent trends, below 0.5 = mean reversion
            • FIGARCH \(d \in [0, 1]\): memory in volatility; \(d = 0\) is GARCH (shocks fade geometrically), larger \(d\) = shocks fade more slowly (hyperbolically)
            • Factor loading: correlation between an indicator and a factor, from \(-1\) to 1; an indicator “belongs” to the factor it loads on most
            • Kaiser–Meyer–Olkin (KMO) \(\in [0, 1]\): share of the indicators' correlation that is common to all of them; above 0.8 = data well suited to factor analysis
            • Nagelkerke \(\tilde R^2 \in [0, 1]\): pseudo-\(R^2\) of a logit, rescaled so that 1 = perfect classification

            Pele et al. (2023): Cryptos as a Separate Asset Class

            Slide 22 in the PDF

            Pele, Wesselhöfft, Härdle, Kolossiatis & Yatracos (2023) study 906 assets: 234 cryptos, 635 stocks, 13 FX rates, 17 commodities, 5 bonds and 2 real-estate indices, with daily data from January 2014 to November 2020. Each asset is described by 24 indicators, computed on expanding windows: variance, skewness, kurtosis, the stable parameters \(\alpha\) and \(\gamma\), left and right quantiles and conditional tail expectations (CTE, the mean return beyond a quantile) at levels from 0.5% to 5%, the first-order autocorrelation \(\rho(1)\), the Hurst exponent and the FIGARCH \(d\).

            A factor analysis reduces the 24 indicators to a few common factors. It uses principal components with a Varimax rotation, which makes each indicator load mainly on one factor; with KMO \(= 0.92\) the data are well suited. Three factors explain 88% of the variance. The tail factor explains 76% and collects the quantiles, the CTEs, the variance and the stable parameters; the memory factor (Hurst, \(\rho(1)\), \(d\)) and the moment factor (skewness, kurtosis) explain 6% each.

            The authors then ask how well each factor separates cryptos from other assets, with a logit model \(P(\text{crypto}) = \Lambda(\beta_0 + \beta_1 F)\), where \(F\) is the factor score and \(\Lambda(u) = 1/(1+e^{-u})\). For the tail factor \(\hat\beta_1 = 15.45\), with a standard error (SE) of 4.44, and Nagelkerke \(\tilde R^2 = 0.992\): the tail factor alone classifies almost perfectly. For the moment factor \(\hat\beta_1 = -0.74\) with \(\tilde R^2 = 0.12\), and the memory factor is not significant.

            Two classifiers confirm the result. A support vector machine (SVM), which draws the boundary that best separates two groups, reaches 99.56% accuracy on two factors, and only Bitcoin and Tether (USDT) are misclassified; \(k\)-means with 6 clusters reaches 98.45%. The authors conclude that cryptos are a separate asset class, "mainly due to the tail factor", and that Bitcoin lies at the border.

            In brief

            • Pele, Wesselhöfft, Härdle, Kolossiatis & Yatracos (2023): 906 assets (234 cryptos, 635 stocks, 13 FX, 17 commodities, 5 bonds, 2 real-estate indices), daily, January 2014 – November 2020
            • 24 indicators per asset, on expanding windows: variance, skewness, kurtosis, stable \(\alpha\) and \(\gamma\), left and right quantiles and conditional tail expectations (CTE) at 0.5–5%, \(\rho(1)\), Hurst exponent, fractionally integrated GARCH (FIGARCH) \(d\)
            • Factor analysis (principal components, Varimax; Kaiser–Meyer–Olkin (KMO) \(= 0.92\)): 3 factors explain 88%
              • tail factor 76% (quantiles, CTEs, variance, stable parameters); memory 6% (Hurst, \(\rho(1)\), \(d\)); moment 6% (skewness, kurtosis)
            • Logit \(P(\text{crypto}) = \Lambda(\beta_0 + \beta_1 F)\): tail \(\hat\beta_1 = 15.45\) (standard error (SE) 4.44), Nagelkerke \(\tilde R^2 = 0.992\); moment \(-0.74\) (\(\tilde R^2 = 0.12\)); memory n.s.
            • Support vector machine (SVM) on two factors: 99.56% accuracy, only BTC and USDT misclassified; \(k\)-means (6 clusters): 98.45%
            • Conclusion: a separate asset class, “mainly due to the tail factor”; BTC lies at the border

            Indicator Averages by Asset Class (Pele et al., 2023)

            Slide 23 in the PDF

            The table reports the average of six indicators within each class, for daily returns in decimals over 2014–2020, from the study of Pele, Wesselhöfft, Härdle, Kolossiatis & Yatracos (2023) on 906 assets. For cryptos, the 1% daily quantile averaged over 234 cryptos is \(Q_{1\%} = -0.366\): a loss of 36.6% or more occurs on 1% of days; for stocks the figure is 5.4% and for FX rates 1.3%. The \(\mathrm{CTE}_{1\%}\), the mean loss on those worst days, is 56.4% for cryptos and 8.5% for stocks.

            The variance of cryptos, \(26.442 \times 10^{-3}\), is about 60 times that of stocks, \(0.400 \times 10^{-3}\). Their stable \(\alpha\), 1.34, is lower than that of stocks, 1.60, which means heavier tails within the stable model. Cryptos are also the only class with positive average skewness (0.686), while every traditional class is skewed to the left. The FIGARCH \(d\) of cryptos, 0.553, is close to that of bonds (0.545): long memory in volatility does not separate them.

            The authors also use projection pursuit, a search for directions in the space of indicators along which two groups separate. Without the stable \(\alpha\), no direction separates cryptos perfectly; with it, 181,618 directions do. For portfolios, the average 1% value at risk (VaR) of an equal-weight portfolio, computed by historical simulation, is 3.01% for classical assets, 16.34% for cryptos and 4.37% for a mixed portfolio.

            The data of 2021–2026 in the previous slides are consistent with "Bitcoin at the border": on standardised features, and on shape features alone, Bitcoin and Ether group with traditional assets, while most altcoins stay apart.

            In brief

            CryptosStocksBondsFXCommoditiesReal estate
            \(\sigma^2 \cdot 10^3\)26.4420.4000.7450.0280.3880.146
            stable \(\alpha\)1.3421.6021.4601.7221.6621.696
            \(Q_{1\%}\)\(-0.366\)\(-0.054\)\(-0.057\)\(-0.013\)\(-0.050\)\(-0.034\)
            \(\mathrm{CTE}_{1\%}\)\(-0.564\)\(-0.085\)\(-0.097\)\(-0.017\)\(-0.074\)\(-0.052\)
            skewness\(0.686\)\(-0.560\)\(-0.829\)\(-0.937\)\(-0.495\)\(-1.142\)
            FIGARCH \(d\)0.5530.2890.5450.4070.4260.510
            • Daily returns in decimals, 2014–2020: crypto \(Q_{1\%} = -0.366\) is a 1% daily quantile of \(-36.6\%\), the average over 234 cryptos
            • Crypto variance is about 60 times that of stocks; stable \(\alpha = 1.34\) against 1.60: heavier tails within the stable model
            • Separation by projection pursuit: without stable \(\alpha\) no direction separates cryptos perfectly; with it, 181 618 directions do
            • Average 1% value at risk (VaR) of an equal-weight portfolio (historical): classical 3.01%, cryptos 16.34%, mixed 4.37%
            • 2021–2026 (previous slides): on standardised features, and on shape features alone, BTC and ETH group with traditional assets, consistent with “BTC at the border”; most altcoins stay apart

            Section summary

            By the shape of their returns, and above all by their tails, crypto assets form a separate class: in Pele, Wesselhöfft, Härdle, Kolossiatis & Yatracos (2023) the tail factor explains 76% of the variance of 24 indicators and separates cryptos from other assets with 99.6% accuracy. The largest coins, Bitcoin and Ether, lie at the border and in 2021–2026 group with traditional assets, while most altcoins and Dogecoin remain apart.

            Event Studies: Ethereum Upgrades

            An event study measures how a price reacts to a dated event by comparing the actual return with the return that would have been expected without the event. The section explains the method, applies it to 13 upgrades of the Ethereum protocol and compares the result with a study that looks at attention, activity and volatility as well as prices.

            The Market-Model Event Study

            Slide 24 in the PDF

            The method, surveyed by MacKinlay (1997), has three steps. The first defines the normal return, the return expected without the event. Here it comes from a market model estimated on the window from 250 to 31 days before the event, 220 days that end well before any news can leak:

            \[R_t = \alpha + \beta R_{m,t} + u_t\]
            (7.7)

            where \(R_t\) is the return of the asset and \(R_{m,t}\) the return of the market. The second step computes the abnormal return (AR), the actual return minus the normal return, and adds it up over an event window from day \(\tau_1\) to day \(\tau_2\) to obtain the cumulative abnormal return (CAR):

            \[\mathrm{AR}_t = R_t - \hat\alpha - \hat\beta R_{m,t}, \qquad \mathrm{CAR}(\tau_1, \tau_2) = \sum_{t=\tau_1}^{\tau_2}\mathrm{AR}_t\]
            (7.8)

            The third step tests whether the CAR differs from zero. For one event, \(t = \mathrm{CAR}(\tau_1,\tau_2) / \sqrt{L\,\hat\sigma_u^2}\), where \(L = \tau_2 - \tau_1 + 1\) is the length of the window and \(\hat\sigma_u^2\) the residual variance of the market model. This simplified version ignores the estimation error of \(\hat\alpha\) and \(\hat\beta\). The error is small when the estimation window is long: with 220 estimation days, \(|t|\) is overstated by only 0.7–1.7% for the Ether upgrades.

            For many events \(e = 1, \dots, N\), three tests ask whether the average reaction is zero:

            \[t_{CS} = \frac{\overline{\mathrm{CAR}}}{s_{\mathrm{CAR}}/\sqrt N}, \qquad \mathrm{SCAR}_e = \frac{\mathrm{CAR}_e}{\hat\sigma_{u,e}\sqrt L}, \qquad t_{BMP} = \frac{\overline{\mathrm{SCAR}}}{s_{\mathrm{SCAR}}/\sqrt N}\]
            (7.9)

            The cross-sectional (CS) test divides the mean CAR by its standard error across events, with \(s_{\mathrm{CAR}}\) the standard deviation of the CARs. The simplified Boehmer–Musumeci–Poulsen (BMP) test first standardises each CAR by its own residual volatility (the standardised CAR, SCAR), so that events in calm and in volatile periods weigh equally; the forecast-error correction of the original test is left out. The sign test counts the positive CARs, which under the null hypothesis of no effect follow a binomial distribution \(\mathrm{Bin}(N, 0.5)\).

            The application studies Ether against an equal-weighted index of 11 other coins as the market, with daily data. There are 13 upgrade families, from Istanbul in 2019 to Fusaka in 2025, each with two dates: the formal schedule (when the date was fixed) and the implementation (when the upgrade went live). The CS, BMP and sign tests assume that events are independent. When events cluster in time, this is only approximately true. For Ether, two schedule dates, Istanbul and Muir Glacier, are 2 days apart, and 17 of the 26 estimation windows contain another event day.

            In brief

            • Normal return from a market model estimated on \([-250, -31]\) days before the event:
              \[ R_t = \alpha + \beta R_{m,t} + u_t, \qquad \mathrm{AR}_t = R_t - \hat\alpha - \hat\beta R_{m,t}, \qquad \mathrm{CAR}(\tau_1, \tau_2) = \sum_{t=\tau_1}^{\tau_2}\mathrm{AR}_t \]
            • One event: \(\;t = \mathrm{CAR}(\tau_1,\tau_2) / \sqrt{L\,\hat\sigma_u^2}\), \(\;L = \tau_2 - \tau_1 + 1\); simplified: it ignores the estimation error of \(\hat\alpha, \hat\beta\), which is small for a long window (220 days: \(|t|\) overstated by 0.7–1.7% for the Ether upgrades)
            • Many events \(e = 1, \dots, N\): cross-sectional (CS) test, and a simplified Boehmer–Musumeci–Poulsen (BMP) test on standardised CARs (SCAR, without the forecast-error correction)
              \[ t_{CS} = \frac{\overline{\mathrm{CAR}}}{s_{\mathrm{CAR}}/\sqrt N}, \qquad \mathrm{SCAR}_e = \frac{\mathrm{CAR}_e}{\hat\sigma_{u,e}\sqrt L}, \qquad t_{BMP} = \frac{\overline{\mathrm{SCAR}}}{s_{\mathrm{SCAR}}/\sqrt N} \]
            • Sign test: number of positive CARs \(\sim \mathrm{Bin}(N, 0.5)\) under \(H_0\)
            • Application: Ether against an equal-weighted index of 11 other coins, daily; 13 upgrade families (Istanbul 2019 to Fusaka 2025), two dates each: formal schedule and implementation
            • Event clustering: the CS, BMP and sign tests assume independent events; for Ether, two schedule dates are 2 days apart (Istanbul, Muir Glacier), and 17 of 26 estimation windows contain another event day

            Worked Example: the Merge, Step by Step

            Slide 25 in the PDF

            The Merge of 15 September 2022 moved Ethereum from proof of work to proof of stake. The worked example computes its cumulative abnormal return by hand. The market model, fitted on days \(-250\) to \(-31\) before the event, gives \(\hat\alpha = 0.064\%\), \(\hat\beta = 1.054\) and a residual standard deviation \(\hat\sigma_u = 2.08\%\) a day. In that window Ether moved almost one for one with the index of 11 other coins.

            The first step computes the abnormal return of each day as the actual return minus the normal return predicted by the model, \(\mathrm{AR}_t = R_t - \hat\alpha - \hat\beta R_{m,t}\). On 14 September, Ether rose 3.36% and the index 1.66%; the model predicted \(0.064 + 1.054 \times 1.66 = 1.81\%\), so \(\mathrm{AR} = 1.54\%\). On the event day Ether fell 10.51% while the index fell 2.85%; the model predicted \(0.064 - 1.054 \times 2.85 = -2.94\%\), so \(\mathrm{AR} = -10.51 + 2.94 = -7.56\%\). On 16 September Ether fell 2.70% while the index rose 2.22%, so \(\mathrm{AR} = -2.70 - 0.064 - 1.054 \times 2.22 = -5.10\%\).

            The second step adds the three abnormal returns: \(\mathrm{CAR}[-1, 1] = 1.54 - 7.56 - 5.10 = -11.13\%\). The third step computes its standard error under the null hypothesis of no effect. Three independent daily residuals with standard deviation 2.08% add up to a standard deviation of \(2.08 \times \sqrt 3 = 3.61\%\). The fourth step divides: \(t = -11.13/3.61 = -3.09\). Since \(|t| > 1.96\), the abnormal return of the Merge is significant at 5% on its own.

            The calculation shows where the result comes from. Only a small part of Ether's fall on the event day is explained by the market; the rest, and the continued fall on the next day while the market rose, is specific to Ether.

            In brief

            • Implementation date: 15 September 2022. Market model fitted on days \([-250, -31]\): \(\hat\alpha = 0.064\%\), \(\hat\beta = 1.054\), residual sd \(\hat\sigma_u = 2.08\%\)
            • Step 1, abnormal return each day: \(\mathrm{AR}_t = R_t - \hat\alpha - \hat\beta R_{m,t}\) (returns in %)
            dayETH \(R_t\)index \(R_{m,t}\)\(\mathrm{AR}_t\)calculation
            \(-1\) (14 Sep)3.361.661.54\(3.36 - 0.064 - 1.054 \times 1.66\)
            \(0\) (15 Sep)\(-10.51\)\(-2.85\)\(-7.56\)\(-10.51 - 0.064 + 1.054 \times 2.85\)
            \(+1\) (16 Sep)\(-2.70\)2.22\(-5.10\)\(-2.70 - 0.064 - 1.054 \times 2.22\)
            • Step 2, cumulate: \(\mathrm{CAR}[-1, 1] = 1.54 - 7.56 - 5.10 = -11.13\%\)
            • Step 3, standard error over \(L = 3\) days: \(\hat\sigma_u\sqrt{3} = 2.08 \times 1.732 = 3.61\%\)
            • Step 4, test: \(t = -11.13/3.61 = -3.09\); \(|t| > 1.96\), so this single CAR is significant at 5%
            • On the day itself the market model predicts \(0.064 - 1.054 \times 2.85 = -2.94\%\) for ETH; the rest of the \(-10.51\%\), that is \(-7.56\%\), is abnormal
            • Source: own computation on EODHD daily prices of Ether and 11 other coins; upgrade dates 2019–2025

            CARs around 13 Upgrades

            Slide 26 in the PDF

            The table lists, for each upgrade, the CAR over the three days around the implementation date, \([-1, 1]\), and its \(t\) statistic. Most CARs are small and not significant. Two are significant at 5% and they have opposite signs: the Merge of September 2022, which moved Ethereum from proof of work to proof of stake, has \(\mathrm{CAR} = -11.1\%\) (\(t = -3.09\)), and Pectra in May 2025 has \(+11.3\%\) (\(t = 2.92\)). Berlin (\(-9.9\%\)) and London (\(+8.7\%\)), the upgrade that introduced the fee mechanism EIP-1559 (EIP: Ethereum Improvement Proposal), are large but not significant at 5% (\(t = -1.79\) and \(t = 1.35\)).

            The cross-sectional tests ask whether the upgrades move the price on average. At implementation, the mean \(\overline{\mathrm{CAR}}[-1,1] = -0.18\%\), with \(t_{CS} = -0.10\) (p-value 0.92) and \(t_{BMP} = -0.20\). Seven of 13 CARs are positive, which gives a sign-test p-value of 1.00. Over the longer window \([0, 5]\) the mean CAR is \(-0.21\%\) (p 0.94). At the schedule dates the mean \(\overline{\mathrm{CAR}}[-1,1]\) is \(-0.59\%\) (p 0.63), and the paired difference between implementation and schedule is \(+0.41\%\) (p 0.84).

            We do not reject the hypothesis of a zero average reaction. The large CARs of the Merge and Pectra have opposite signs, and no common revaluation of Ether is detectable in these windows. Individual upgrades may still matter, but 13 events with large and noisy abnormal returns cannot establish a common effect.

            In brief

            implementationCAR\([-1,1]\)\(t\)
            Istanbul (Dec 2019)\(0.4\%\)\(0.12\)
            Muir Glacier\(-0.3\%\)\(-0.10\)
            Beacon Chain\(1.3\%\)\(0.41\)
            Berlin\(-9.9\%\)\(-1.79\)
            London (EIP-1559)\(8.7\%\)\(1.35\)
            Altair\(4.4\%\)\(0.84\)
            Arrow Glacier\(-4.3\%\)\(-0.85\)
            Gray Glacier\(-4.6\%\)\(-1.48\)
            Merge (Sep 2022)\(\mathbf{-11.1\%}\)\(\mathbf{-3.09}\)
            Shapella\(3.3\%\)\(1.17\)
            Dencun\(-3.1\%\)\(-1.16\)
            Pectra (May 2025)\(\mathbf{11.3\%}\)\(\mathbf{2.92}\)
            Fusaka (Dec 2025)\(1.7\%\)\(0.54\)
            • Cross-section, implementation:
              • \(\overline{\mathrm{CAR}}[-1,1] = -0.18\%\), \(t_{CS} = -0.10\) (p 0.92), \(t_{BMP} = -0.20\)
              • 7 of 13 positive (sign test p \(= 1.00\))
              • \(\overline{\mathrm{CAR}}[0,5] = -0.21\%\) (p 0.94)
            • Schedule dates: \(\overline{\mathrm{CAR}}[-1,1] = -0.59\%\) (p 0.63)
            • Implementation minus schedule, paired: \(+0.41\%\) (p 0.84)
            • Merge and Pectra have large CARs of opposite sign; no common revaluation of Ether is detectable in these windows
            • Source: own computation on EODHD daily prices of Ether and 11 other coins; upgrade dates 2019–2025

            Significance of Individual CARs

            Slide 27 in the PDF
              Question for the room The ▶ button extends the window of all 26 CARs from [−1, 0] to [−1, +10]; the 5% band of each event widens with √L. How many of the 26 dates lie outside their own band as the window grows?
              Answer

              Between 2 and 5 at every window: 5 for [−1, 0], 3 for [−1, +1], then 2 to 4. The Merge and Pectra implementation dates are outside in every window from [−1, +1] on (the Merge reaches −23.3% at [−1, +10], Pectra +24.8%); the others enter and leave. With 26 dates and a 5% test, about 1.3 would lie outside by chance in each window. A longer window adds noise to every CAR and widens every band by the same factor √L, so it does not create new significant events.

              The chart shows all 26 cumulative abnormal returns \(\mathrm{CAR}[-1, +1]\): for each upgrade family, the schedule date in dark blue and the implementation date in purple. The red whiskers around zero mark the 5% band of each event, \(\pm 1.96\,\hat\sigma_u\sqrt 3\), computed from that event's own market model. A bar that reaches beyond its whisker is significant at 5%.

              The bands differ from event to event because the residual volatility does. The narrowest is \(\pm 5.2\%\), for the schedule date of Dencun in early 2024, a calm period; the widest is \(\pm 12.6\%\), for the implementation of London in August 2021, when the whole market was very volatile. The same CAR can therefore be significant for one event and not for another.

              Three of the 26 bars leave their band: the implementation dates of the Merge (\(-11.1\%\)) and of Pectra (\(+11.3\%\)), and the schedule date of Gray Glacier (\(-10.0\%\)). If no upgrade had any effect, about \(26 \times 0.05 = 1.3\) dates would leave their band by chance, so three is not far from what chance produces. Berlin's implementation CAR of \(-9.9\%\) is almost as large as the Merge's but stays inside its band of \(\pm 10.9\%\), because Ether was more volatile before Berlin.

              In brief

              • Bars: CAR\([-1, +1]\) of each upgrade, schedule (blue) and implementation (purple) dates. Red whiskers: \(\pm 1.96\,\hat\sigma_u\sqrt3\), the 5% band of that event
              • The band differs by event because \(\hat\sigma_u\) does: \(\pm 5.2\%\) (Dencun schedule) to \(\pm 12.6\%\) (London, in the volatile 2021 market)
              • Outside the band: Merge and Pectra (implementation), Gray Glacier (schedule): 3 of 26 dates, against 1.3 expected by chance at 5%
              • Berlin (\(-9.9\%\)) is almost as large as the Merge but stays inside its wider band
              • Source: own computation on EODHD daily prices of Ether and 11 other coins; upgrade dates 2019–2025
              BA_L7_cars

              Worked Example: Testing 13 CARs Together

              Slide 28 in the PDF

              The cross-sectional test asks whether the upgrades move the price on average. The worked example applies it to the 13 implementation CARs\([-1, 1]\), in %: 0.39, \(-0.31\), 1.28, \(-9.91\), 8.67, 4.35, \(-4.27\), \(-4.59\), \(-11.13\), 3.32, \(-3.14\), 11.25 and 1.73.

              The first step is the mean: the CARs add up to \(-2.36\), so \(\overline{\mathrm{CAR}} = -2.36/13 = -0.18\%\). The second step measures the spread across events, the sample standard deviation \(s_{\mathrm{CAR}} = 6.50\%\), and divides it by \(\sqrt{13}\) to obtain the standard error of the mean, 1.80%. The third step forms \(t_{CS} = -0.18/1.80 = -0.10\). With 12 degrees of freedom the 5% critical value is 2.18, and the two-sided p-value is 0.92: we do not reject a zero mean effect.

              The fourth step is the sign test, which uses only the signs. Seven of the 13 CARs are positive. If the upgrades had no effect, the number of positive CARs would follow a binomial distribution with 13 trials and probability 0.5, whose mean is 6.5; seven is as close to the centre as an integer can be, and the p-value is 1.00.

              Two individual CARs, the Merge and Pectra, are significant on their own, but with opposite signs, and they cancel in the mean. The cross-sectional test uses the spread of the CARs across events rather than the residual volatility of one market model. It therefore allows each upgrade to have its own true effect and asks only whether these effects are zero on average.

              In brief

              • The 13 implementation CARs\([-1, 1]\) in %: \(0.39, -0.31, 1.28, -9.91, 8.67, 4.35, -4.27, -4.59, -11.13, 3.32, -3.14, 11.25, 1.73\)
              • Step 1, mean: \(\overline{\mathrm{CAR}} = -2.36/13 = -0.18\%\)
              • Step 2, spread across events: \(s_{\mathrm{CAR}} = 6.50\%\); standard error of the mean \(= 6.50/\sqrt{13} = 1.80\%\)
              • Step 3, cross-sectional test: \(t_{CS} = -0.18/1.80 = -0.10\); critical value \(t_{12,\,0.975} = 2.18\); p \(= 0.92\): we do not reject a zero mean
              • Step 4, sign test: 7 of 13 CARs are positive; under \(H_0\) the count is \(\mathrm{Bin}(13, 0.5)\) with mean 6.5, so 7 is as central as possible: p \(= 1.00\)
              • Two individual CARs are significant (Merge \(-11.1\%\), Pectra \(+11.3\%\)), but they cancel: large, opposite-signed reactions are compatible with a zero average effect
              • The test needs \(s_{\mathrm{CAR}}\) from the cross-section, not \(\hat\sigma_u\): it allows each upgrade to have its own true effect
              • Source: own computation on EODHD daily prices of Ether and 11 other coins; upgrade dates 2019–2025

              Abnormal Returns around the Event Day

              Slide 29 in the PDF
                Question for the room The ▶ button extends the CAR window of the 13 implementation dates from [−1, 0] to [−1, +10]. Does the mean CAR move away from zero as the window lengthens?
                Answer

                No. The mean CAR[−1, τ₂] stays between −0.25% and +1.51%, and the cross-sectional test never comes near rejection: |t| ≤ 0.53, p ≥ 0.61 for every τ₂. The individual paths fan out as the window lengthens, because each CAR adds more daily noise (residual sd 1.5–3.7% a day), but the mean does not drift. The schedule dates give the same picture: the lowest p-value is 0.17, for [−1, +4].

                The chart plots the cumulative abnormal return of Ether from 10 days before to 10 days after each event, for the schedule dates (left) and the implementation dates (right). Each thin purple line is one upgrade family, the thick red line is the mean CAR across the 13 families, and the dotted vertical line marks the event day.

                The individual paths fan out in both directions, some to more than \(+25\%\) and others below \(-30\%\) within ten days, while the mean path stays close to zero in both panels. This is the picture behind the non-significant cross-sectional tests: large reactions exist, but they do not share a sign.

                The market model behind the chart differs moderately across events. The estimated beta of Ether to the coin index ranges from 0.75 (Dencun) to 1.20 (Fusaka), and the residual standard deviation from 1.5% to 3.7% a day. A daily residual volatility of 3.7% gives a three-day standard deviation of about \(3.7\% \times \sqrt 3 \approx 6.4\%\), which explains why CARs of 9–10% need not be significant.

                In brief

                • Individual paths fan out in both directions; the mean path stays near zero
                • Estimated betas of Ether to the coin index: 0.75 (Dencun) to 1.20 (Fusaka); residual sd 1.5–3.7% a day
                • Source: own computation on EODHD daily prices of Ether and 11 other coins; upgrade dates 2019–2025
                BA_L7_event

                Lin et al. (2026): Effects of Ethereum Upgrades on Market Outcomes

                Slide 30 in the PDF

                Lin, Găman, Wang & Pele (2026) study the same 13 upgrade families, but look at outcomes beyond returns: internet search attention, the number of active addresses on the chain and volatility. They compare Ether with other coins in two ways. The cross-asset difference-in-differences estimate \(\hat\delta_{e,s}\) is the change of an outcome for Ether around event \(e\) minus the average change for \(J_e\) comparison ("donor") coins:

                \[\hat\delta_{e,s} = \Delta y_{e,s,ETH} - \frac{1}{J_e}\sum_{j=1}^{J_e}\Delta y_{e,s,j}, \qquad \hat y^{SC}_{e,t} = \sum_j w_{e,j}\,y_{j,t},\; w_{e,j} \geq 0,\; \textstyle\sum_j w_{e,j} = 1\]
                (7.10)

                The synthetic control (SC) \(\hat y^{SC}_{e,t}\) builds a weighted average of donor coins that matches Ether before the event; the gap after the event is the estimated effect. The tests are exact sign-flip tests at the level of upgrade families, and the p-values are adjusted for testing many outcomes with the Benjamini–Hochberg (BH) method.

                The estimates are in log points, where 0.01 is about 1%. Google search attention rises already on the schedule day (0.184, BH p 0.034) and more on the implementation day (0.350, BH p 0.004), which is \(e^{0.350} - 1 = +42\%\). Active addresses fall at implementation, by 0.145 log points or 13.5%, in the synthetic-control comparison (BH p 0.020). The median hourly log realised volatility (RV) in the implementation hours rises by 0.316, or 37% (BH p 0.007). The share of jumps in price variation over 24 hours, measured with bipower variation, does not change (estimate near 0, BH p 0.99), and the paired CAR between implementation and schedule is \(-0.0098\) with a raw p-value of 0.375: no repricing is detected.

                An upgrade therefore goes with more attention, more volatility and fewer active addresses, but not with a detectable price change. Two common interpretations go beyond these estimates. A non-significant CAR does not prove that upgrades are "priced in", and fewer active addresses do not show that "users pause" during an upgrade: both are hypotheses consistent with the numbers, not conclusions from them.

                In brief

                • Lin, Găman, Wang & Pele (2026): 13 upgrade families; outcomes beyond returns; cross-asset difference-in-differences and synthetic control (SC) against donor coins
                  \[ \hat\delta_{e,s} = \Delta y_{e,s,ETH} - \frac{1}{J_e}\sum_{j=1}^{J_e}\Delta y_{e,s,j}, \qquad \hat y^{SC}_{e,t} = \sum_j w_{e,j}\,y_{j,t},\; w_{e,j} \geq 0,\; \textstyle\sum_j w_{e,j} = 1 \]
                • Exact sign-flip tests at family level, Benjamini–Hochberg adjusted p-values (BH p)
                outcome (log points)estimateBH pmeaning
                Google attention, schedule day\(0.184\)\(0.034\)search starts at scheduling
                Google attention, implementation day\(0.350\)\(0.004\)\(+42\%\)
                active addresses, implementation (synthetic control)\(-0.145\)\(0.020\)\(-13.5\%\)
                median hourly log realised volatility (RV), implementation hours\(0.316\)\(0.007\)\(+37\%\) volatility
                jump share, bipower variation (24 h)\(\approx 0\)\(0.99\)no lasting change
                paired CAR, implementation minus schedule\(-0.0098\)raw p \(0.375\)no repricing detected
                • The event goes with more attention and volatility and fewer active addresses
                • A non-significant CAR does not prove that upgrades are “priced in”, and fewer addresses do not show that “users pause”: both are hypotheses consistent with these estimates

                Section summary

                Over 13 Ethereum upgrades, abnormal returns are large for some events, of both signs, and the mean CAR is not distinguishable from zero. Upgrades go with measurable changes in attention, on-chain activity and volatility, which a study limited to prices would miss.

                Sentiment from Social Media

                A sentiment index turns a large amount of text, such as news or social-media posts, into a daily number that measures how positive or negative the mood is. The section describes how the BitMood index is built from Facebook posts, reads its two applications, a trading rule and volume forecasts, and sets out the minimum standard that such results must meet.

                BitMood: Building a Sentiment Index

                Slide 31 in the PDF

                Conda, Găman, Bâg, Mazurencu-Marinescu-Pele, Pele & Härdle (2026) build the index from 120,000 English Facebook posts about Bitcoin, a 5% monthly sample of about 2.3 million posts collected with the CrowdTangle tool, from 3 January 2015 to 15 October 2023. Each post is scored by FinBERT, a language model trained on financial text, which gives the probabilities that the post is positive, neutral or negative, \(S^+\), \(S^0\) and \(S^-\). The compound score (CS) of post \(i\) is \(\mathrm{CS}_i = S^+_i - S^-_i\).

                Posts with many interactions reach more readers, so the daily index weights each post by its total interactions \(T_i\). The daily weighted probabilities and the engagement-weighted compound score (WC) of day \(d\) are

                \[P^{\pm,0}_d = \frac{\sum_{i \in d} S^{\pm,0}_i T_i}{\sum_{i \in d} T_i}, \qquad \mathrm{WC}_d = \frac{P^+_d - P^-_d}{P^+_d + P^-_d + P^0_d} \in [-1, 1]\]
                (7.11)

                The index is \(+1\) when all weighted posts are certainly positive, \(-1\) when all are certainly negative and 0 when positive and negative weights balance.

                A second measure, the reaction-intensity score (SI) of a post, uses the Facebook reactions with weights from the literature: \(\mathrm{SI}_i = \sum_k(\log N_{i,k} + 1)\,w_k\), where \(N_{i,k}\) is the number of reactions of type \(k\); the weights are 0.1 for a like, 0.05 for a comment, 0.2 for a share, 0.5 for love, wow and haha, and \(-0.5\) for sad and angry. Because \(\log 0\) is undefined and the article does not say how posts with no reactions of a type are treated, a reader who reproduces the score must choose a rule; \(\log(1 + N_{i,k})\) is a common fix.

                A topic model, latent Dirichlet allocation (LDA), groups the words of the posts into themes: investment (42.8% of the tokens), trading and blockchain (24.6%), gaming (14.3%), links to other social media (13%) and business (5.3%).

                In brief

                • Conda, Găman, Bâg, Mazurencu-Marinescu-Pele, Pele & Härdle (2026): 120 000 English Facebook posts on Bitcoin (5% monthly sample of about 2.3 million, CrowdTangle), 3 January 2015 – 15 October 2023
                • FinBERT gives each post probabilities \(S^+, S^0, S^-\); compound score (CS) \(\mathrm{CS}_{i} = S^+_{i} - S^-_{i}\)
                • Engagement-weighted compound score (WC), daily, with \(T_i\) = total interactions of post \(i\):
                  \[ P^{\pm,0}_d = \frac{\sum_{i \in d} S^{\pm,0}_i T_i}{\sum_{i \in d} T_i}, \qquad \mathrm{WC}_d = \frac{P^+_d - P^-_d}{P^+_d + P^-_d + P^0_d} \in [-1, 1] \]
                • Reaction-intensity score (SI) of a post (weights from the literature): \(\mathrm{SI}_i = \sum_k(\log N_{i,k} + 1)\,w_k\), \(N_{i,k}\) = number of reactions of type \(k\); like 0.1, comment 0.05, share 0.2, love/wow/haha 0.5, sad/angry \(-0.5\)
                  • \(\log 0\) is undefined and the article does not say how posts with no reactions of a type are treated; \(\log(1 + N_{i,k})\) is a common fix
                • Topics (latent Dirichlet allocation, LDA): investment 42.8%, trading and blockchain 24.6%, gaming 14.3%, social-media links 13%, business 5.3% of tokens

                BitMood: a Contrarian Trading Rule

                Slide 32 in the PDF

                The first application is a trading rule. Following an earlier rule of Preis et al., the signal compares today's value of a series \(f\) with its average over the previous \(\delta t\) days,

                \[\Delta f(t, \delta t) = f(t) - \frac{1}{\delta t}\sum_{k=1}^{\delta t} f(t-k)\]
                (7.12)

                The rule is contrarian: it buys Bitcoin when sentiment has fallen below its recent average and sells when it has risen. Performance is measured by the Sharpe ratio (SR), the mean return of the strategy divided by the standard deviation of its returns; Lecture 8 annualises it by multiplying the daily ratio by \(\sqrt{365}\).

                The best configuration reported for the test period, 1 January – 15 October 2023, uses a window of \(\delta t = 3\) days, the compound sentiment and a transaction cost of \(10^{-5}\). It reaches a cumulative return of 1.321 and a Sharpe ratio of 0.210. Three details limit the reading of these numbers. The article does not say whether the return is a simple return, a log return or a wealth multiple, nor whether the SR is daily or annualised. A cost of \(10^{-5}\) of the amount traded is 0.1 basis point (bp, 0.01%) per trade, far below typical exchange fees.

                The main caveat concerns selection. The best of many combinations, 7 windows times 6 cost levels times several sentiment series, was chosen on the same 2023 period on which it is scored. The reported performance is then the best of many tries, which is higher than what the chosen rule can be expected to deliver on new data; Lecture 8 measures this effect.

                In brief

                • Contrarian rule (after Preis et al.): \(\Delta f(t, \delta t) = f(t) - \frac{1}{\delta t}\sum_{k=1}^{\delta t} f(t-k)\); buy when sentiment fell, sell when it rose
                • Sharpe ratio (SR): mean return of the strategy divided by the standard deviation of its returns; Lecture 8 annualises it with \(\sqrt{365}\)
                • Best configuration, test period 1 January – 15 October 2023: \(\delta t = 3\) days, compound sentiment, cost \(10^{-5}\)
                  • cumulative return 1.321 and Sharpe ratio 0.210
                  • the article does not say whether the return is simple, log or a wealth multiple, or whether the SR is daily or annualised
                  • a cost of \(10^{-5}\) as a fraction of the amount traded is 0.1 basis point (bp) per trade, far below typical exchange fees
                • Caveat: the best of many window–feature–cost combinations (7 windows \(\times\) 6 costs \(\times\) several sentiment series) was chosen on the same 2023 period on which it is scored

                BitMood: Volume Forecasts

                Slide 33 in the PDF

                The second application forecasts the daily trading volume of Bitcoin from January 2022 to October 2023, with sentiment among the inputs. Seven models are compared: a linear regression, a random forest (RF), an average of many decision trees, and five neural networks: a stacked long short-term memory (LSTM) network, a convolutional neural network (CNN), a combined LSTM-CNN, a gated recurrent unit (GRU) network and a bidirectional LSTM (Bi-LSTM). The accuracy measure is the root mean squared error (RMSE) of the scaled volume, the square root of the mean squared forecast error; lower is better.

                The correct reading uses the two rows in order: select on the validation period, report on the test period. The linear regression has the lowest validation RMSE, 0.1423, so it is the model to choose; its test RMSE is 0.1935, only 0.0005 above the best test value, 0.1930 of the Bi-LSTM. Picking the Bi-LSTM because of its test value would use the test data for selection. In this comparison the simplest model is as accurate as the most complex ones.

                The Shapley additive explanations (SHAP) of the linear model, which split each forecast into contributions of the inputs, are dominated by lagged volume and negative sentiment. This does not show that sentiment improves the forecast: every model in the table uses sentiment, and none is compared with a model that uses lagged volume alone. SHAP describes how a model uses its inputs; it does not replace that comparison. The study also gives no evidence on returns.

                In brief

                • Daily Bitcoin trading volume, January 2022 – October 2023; root mean squared error (RMSE) of scaled volume for linear regression, random forest (RF), long short-term memory (LSTM), convolutional (CNN) and gated recurrent unit (GRU) networks:
                Lin. reg.RFstacked LSTMCNNLSTM-CNNGRUBi-LSTM
                validation\(\mathbf{0.1423}\)0.17670.23230.21790.25010.18400.1701
                test\(\mathbf{0.1935}\)0.23550.24530.25780.27660.19590.1930
                • Select on validation, report on test: the linear regression has the lowest validation RMSE (0.1423, bold) and a test RMSE of 0.1935, 0.0005 above the best test value (Bi-LSTM)
                • Lagged volume and negative sentiment dominate the Shapley additive explanations (SHAP) of the linear model
                • Every model in the table uses sentiment: the table does not show that sentiment improves on lagged volume alone, and SHAP does not replace that test
                • No evidence here on returns

                A Lagged-Volume Benchmark for BitMood

                Slide 34 in the PDF

                A sentiment model of trading volume is useful only if it beats a forecast built from past volume alone. The chart shows how strong that benchmark is over the BitMood volume sample. The left panel plots the logarithm of daily Bitcoin trading volume from EODHD, 1 January 2022 – 15 October 2023 (653 days), in orange, with its 30-day mean in dark blue. The right panel plots today's log volume against yesterday's, with the fitted line of an autoregression of order 1, AR(1).

                Volume is highly persistent. The autocorrelation of log volume is 0.78 at a lag of one day, 0.74 at seven days and still 0.39 at thirty days. The seven-day value reflects a weekly pattern: log volume is 0.33 below its mean on Saturdays and 0.28 below it on Sundays, so the weekday alone carries information. The AR(1) on yesterday's log volume explains 61% of the variance of today's (\(R^2 = 0.61\)).

                The root mean squared error (RMSE) of simple forecasts, in logs, makes the point concrete: 0.50 when every day is forecast by the sample mean, 0.33 for "the same as yesterday" and 0.31 for the AR(1). A sentiment model has to lower the error beyond these values, on the same test days. The BitMood table reports RMSEs of scaled volume, not of log volume, so the numbers are not directly comparable; more importantly, it does not report the same model without sentiment, which is the comparison that would show what sentiment adds.

                In brief

                A Lagged-Volume Benchmark for BitMood
                • Log of daily Bitcoin trading volume (EODHD), 1 January 2022 – 15 October 2023, the BitMood volume sample; 653 days
                • Autocorrelation of log volume: 0.78 at lag 1, 0.74 at lag 7 (weekly pattern: log volume is 0.33 below its mean on Saturdays and 0.28 on Sundays), 0.39 at lag 30
                • An autoregression of order 1, AR(1), on yesterday's log volume explains 61% of the variance of today's (\(R^2\)); root mean squared error (RMSE) in logs: 0.50 with the sample mean, 0.33 with “same as yesterday”, 0.31 with the AR(1)
                • Sentiment must lower the error beyond this benchmark; the BitMood table does not report that comparison
                • Source: own computation on EODHD daily Bitcoin volume, 1 January 2022 – 15 October 2023
                BA_L7_volume

                Minimum Standards for Back-Test Evaluation

                Slide 35 in the PDF

                A back-test evaluates a forecasting model or trading rule on past data as if it had been used at the time. The two BitMood applications show the most common weaknesses, and they lead to a minimum standard. The data are split into three periods, in time order. The training period estimates the parameters. The validation period chooses the model, the window, the threshold or the cost setting; Lecture 5 chose its alarm threshold this way. The test period scores the chosen configuration once, and nothing is changed after seeing the result.

                The model must be compared with a benchmark that does not use the new information, scored on the same test days. For BitMood this is the same volume model without the sentiment features. The comparison is reported as a difference in RMSE with a confidence interval, for example with the Diebold–Mariano test of Lecture 8.

                Finally, the number of configurations tried must be reported, because the best of many tries looks better than it is. Lecture 8 applies this standard to machine-learning forecasts of Bitcoin and adds the corrections for trying many rules.

                In brief

                • Three periods, in time order:
                  • training: estimate the parameters
                  • validation: choose the model, window, threshold or cost setting (Lecture 5: alarm threshold chosen on validation)
                  • test: score the chosen configuration once; nothing is changed after seeing it
                • A benchmark without the new information, scored on the same test days
                  • for BitMood: the same volume model without the sentiment features
                  • report the RMSE difference with a confidence interval (for example, a Diebold–Mariano test, Lecture 8)
                • Report the number of configurations tried
                • Lecture 8 applies this standard to machine-learning forecasts of Bitcoin and adds the corrections for trying many rules

                Section summary

                Text sentiment can be measured at scale with language models and weighted by engagement. The published results show that sentiment models forecast volume about as well as a linear regression, but they do not test whether sentiment adds to lagged volume, and the trading result was selected on its own test period.

                Summary

                Crypto returns form a dense network. The mean pairwise correlation of 19 coins over 2021–2026 is 0.60, Ether and Polkadot are the hubs of the minimum spanning tree, and the first principal component explains on average 68% of the variance in 90-day windows (59% in October 2026). The spillover analysis adds direction: 73% of the forecast-error variance of eight large coins comes from other coins, Ether and Bitcoin are the largest net transmitters and Dogecoin the largest net receiver; the rolling index reached 83.5% in May 2026.

                By their tails, cryptos form a separate asset class: in Pele, Wesselhöfft, Härdle, Kolossiatis & Yatracos (2023) the tail factor explains 76% of the variance of the indicators and an SVM classifies 99.6% of assets correctly. In 2021–2026, on standardised features and on shape alone, Bitcoin and Ether group with classical assets.

                Ethereum upgrades produce large CARs of both signs, and the mean CAR is not distinguishable from zero; Lin, Găman, Wang & Pele (2026) report effects on attention, activity and volatility instead. Sentiment models forecast volume about as well as a linear regression, whether sentiment adds to lagged volume is not tested, and trading claims need selection on validation data.

                In brief

                • Crypto returns form a dense network: mean correlation 0.60; ETH and DOT are the hubs of the minimum spanning tree; one factor explains 68% of the variance on average (59% in October 2026)
                • Spillovers: 73% of forecast-error variance comes from other coins; ETH and BTC are the largest net transmitters, DOGE the largest net receiver; the index reached 83.5% in May 2026
                • Cryptos are a separate asset class through their tails (Pele et al., 2023: tail factor 76%, SVM 99.6%); on standardised features, and on shape alone, BTC and ETH group with classical assets in 2021–2026
                • Ethereum upgrades: large CARs of both signs, mean CAR not distinguishable from zero; Lin et al. (2026) report effects on attention, activity and volatility
                • Sentiment models forecast volume about as well as a linear regression; whether sentiment adds to lagged volume is not tested; trading claims need selection on validation data

                Interpretation Exercises

                Slide 37 in the PDF

                The five exercises use the numbers of the lecture. Each asks what a measure shows and what it does not show: the role of Bitcoin in the tree and in the spillover table, the diversification benefit at different absorption ratios, the meaning of a negative NET spillover, the conclusions from two significant CARs among thirteen, and the questions to ask about a reported Sharpe ratio.

                In brief

                1. The MST shows BTC with only two links while ETH has six. Does this mean BTC is less important for the market? Use the spillover table in your answer.
                2. The absorption ratio rises from 0.45 to 0.85. What happens to the benefit of holding ten coins instead of one?
                3. In the spillover table, \(\mathrm{NET}_{DOGE} = -18.6\). Explain what this number measures and what it does not prove.
                4. The Merge has CAR\([-1,1] = -11.1\%\) (\(t = -3.09\)) and Pectra \(+11.3\%\) (\(t = 2.92\)), while the mean over 13 upgrades is \(-0.18\%\) (p 0.92). What can and cannot be concluded?
                5. A sentiment strategy reports a Sharpe ratio of 0.21 in 2023. Which two questions would you ask before using it?
                • Next: Lecture 8, machine learning, artificial intelligence (AI) and their limits; exam review

                Appendix: Further Material

                The Appendix holds optional material for self-study. It covers a test for herding among crypto investors and the construction of price indices for unique assets such as digital art.

                Appendix: Further Material

                Slide 38 in the PDF
                Optional material for self-study.
                • Herding: the CSAD Test
                • Rolling-Window Estimates of Herding
                • Hedonic Price Indices for Digital Art
                • Worked Example: Composition Bias in a Boom (simulated)

                Herding: the CSAD Test

                Slide 39 in the PDF

                Herding means that investors follow the market instead of their own information. The test of Christie & Huang (1995) and Chang, Cheng & Khorana (2000) uses the cross-sectional absolute deviation (CSAD), the average distance of the coin returns from the market return on day \(t\), \(\mathrm{CSAD}_t = N^{-1}\sum_{i=1}^N |R_{i,t} - R_{m,t}|\). Under rational pricing, as in the capital asset pricing model (CAPM), dispersion rises linearly with the size of the market move \(|R_{m,t}|\); under herding, coins move together in large moves and dispersion rises less. The regression \(\mathrm{CSAD}_t = \alpha + \gamma_1 |R_{m,t}| + \gamma_2 R_{m,t}^2 + \varepsilon_t\) therefore tests \(H_0: \gamma_2 = 0\) against herding, \(\gamma_2 < 0\). On 19 coins with an equal-weighted market return, \(n = 2{,}121\) days and heteroskedasticity- and autocorrelation-consistent (HAC) standard errors with 10 lags, the full sample gives \(\hat\gamma_1 = 0.254\) (\(t = 10.76\)) and \(\hat\gamma_2 = -0.0028\) (\(t = -2.88\)), with \(R^2 = 0.248\); on up days and on down days separately \(\hat\gamma_2\) is not significant (\(t = 0.03\) and \(-0.12\)).

                In brief

                • Cross-sectional absolute deviation of returns around the market return \(R_{m,t}\):
                  \[ \mathrm{CSAD}_t = \frac1N\sum_{i=1}^N |R_{i,t} - R_{m,t}| \]
                • Rational pricing (capital asset pricing model, CAPM): dispersion rises linearly with \(|R_{m,t}|\). Herding: in big moves, coins move together and dispersion rises less
                  \[ \mathrm{CSAD}_t = \alpha + \gamma_1 |R_{m,t}| + \gamma_2 R_{m,t}^2 + \varepsilon_t, \qquad H_0: \gamma_2 = 0, \quad \text{herding: } \gamma_2 < 0 \]
                • Estimates for 19 coins, January 2021 – October 2026, equal-weighted \(R_m\), \(n = 2\,121\), heteroskedasticity and autocorrelation consistent (HAC(10)) errors:
                sample\(\hat\gamma_1\) (\(t\))\(\hat\gamma_2\) (\(t\))\(R^2\)
                all days\(0.254\) (\(10.76\))\(-0.0028\) (\(-2.88\))\(0.248\)
                up days (\(R_m > 0\))\(0.0003\) (\(0.03\))
                down days\(-0.0001\) (\(-0.12\))

                Rolling-Window Estimates of Herding

                Slide 40 in the PDF

                The left panel plots each day's CSAD against the equal-weighted market return, with the fitted regression in red; the right panel shows the \(t\) statistic of \(\gamma_2\) in rolling 365-day windows, with the dashed lines at \(\pm 1.96\). The full-sample curve is concave, with its vertex at \(|R_m| = \hat\gamma_1/(2|\hat\gamma_2|) \approx 45\%\), and the curvature rests on the 16 days with \(|R_m| > 15\%\). In the rolling windows, \(\gamma_2\) is significantly negative in only 1.7% of windows and significantly positive, which indicates anti-herding, in 39%. Herding in crypto is an episode rather than a permanent trait, and a full-sample coefficient can hide the opposite behaviour during most of the sample.

                In brief

                Rolling-Window Estimates of Herding
                • Full sample: \(\hat\gamma_2 < 0\) significant; the fitted curve is concave (vertex at \(|R_m| = \hat\gamma_1/2|\hat\gamma_2| \approx 45\%\)), and the curvature rests on the 16 days with \(|R_m| > 15\%\)
                • 365-day windows: \(\gamma_2\) significantly negative in only 1.7% of windows, significantly positive (anti-herding) in 39%
                • Interpretation: herding in crypto is an episode, not a trait; a full-sample coefficient can hide the opposite behaviour most of the time
                • Source: own computation on EODHD daily prices of 19 coins, January 2021 – October 2026
                BA_L7_csad

                Hedonic Price Indices for Digital Art

                Slide 41 in the PDF

                Each non-fungible token (NFT) or digital artwork is unique and trades rarely, so the works sold in one quarter differ from those sold in the previous quarter. The median sale price then mixes price changes with composition changes. A time-dummy hedonic regression separates the two: the log price of sale \(n\) is regressed on dummies \(D_{n,t}\) for the quarter of sale and on the characteristics \(\mathbf x_n\) of the work (rarity, artist, size, collection), \(\ln P_n = \sum_{t=1}^{T}\delta_t D_{n,t} + \mathbf x_n'\boldsymbol\beta + \varepsilon_n\). The coefficient \(\hat\delta_t\) is the quality-adjusted log price level of quarter \(t\), and the index is \(I_t = 100\,e^{\hat\delta_t}\). The same method is used for the Digital Art Index (DAI), presented in a Quantinar unit, and for real-estate and art markets in general.

                In brief

                • Each non-fungible token (NFT) or digital artwork is unique and trades rarely; the works sold this quarter differ from those sold last quarter
                • Median sale price mixes two things: price changes and composition changes
                • Time-dummy hedonic regression on all sales \(n\) in quarters \(t = 0, \dots, T\):
                  \[ \ln P_{n} = \sum_{t=1}^{T}\delta_t D_{n,t} + \mathbf x_n'\boldsymbol\beta + \varepsilon_n, \qquad I_t = 100\,e^{\hat\delta_t} \]
                  • \(\mathbf x_n\): characteristics (rarity, artist, size, collection); \(\hat\delta_t\): quality-adjusted log price level
                • Used for the Digital Art Index (Quantinar unit) and for real estate and art markets in general

                Worked Example: Composition Bias in a Boom (simulated)

                Slide 42 in the PDF

                The simulation has 3,600 sales over 12 quarters; in the boom, quarters 2 to 5, more low-rarity works by unknown artists are sold. The chart shows the true index (green), the hedonic index (dark blue, with its 95% interval) and the median-price index (red, dashed), all equal to 100 in quarter 0. The hedonic regression recovers the effects of the characteristics, 2.52 for rarity and 1.20 for a known artist against true values of 2.5 and 1.2, and tracks the true index with an RMSE of 0.09 in logs. The median index has an RMSE of 0.25 and at the peak shows 315 instead of 448, 30% too low: the boom looks smaller because cheaper works flood the market.

                In brief

                Worked Example: Composition Bias in a Boom (simulated)
                • 3 600 simulated sales, 12 quarters; in the boom (quarters 2–5) more low-rarity, unknown-artist works are sold
                • Hedonic regression recovers the characteristics (rarity 2.52, known artist 1.20; true 2.5 and 1.2) and the index: RMSE 0.09 in logs
                • Median index: RMSE 0.25; at the peak it shows 315 instead of 448 (30% too low): the boom looks smaller because cheaper works flood the market
                BA_L7_hedonic

                References

                • Chang, E. C., Cheng, J. W. & Khorana, A. (2000). An Examination of Herd Behavior in Equity Markets: An International Perspective. Journal of Banking & Finance, 24(10), 1651–1679. https://doi.org/10.1016/S0378-4266(99)00096-5
                • Christie, W. G. & Huang, R. D. (1995). Following the Pied Piper: Do Individual Returns Herd around the Market? Financial Analysts Journal, 51(4), 31–37. https://doi.org/10.2469/faj.v51.n4.1918
                • Conda, A. I., Găman, Ș., Bâg, R. C., Mazurencu-Marinescu-Pele, M., Pele, D. T. & Härdle, W. K. (2026). BitMood: AI Analysis of Bitcoin Trends via Facebook Emotions. Digital Finance, 8(3), article 42. https://doi.org/10.1007/s42521-026-00205-3
                • Diebold, F. X. & Yilmaz, K. (2012). Better to Give than to Receive: Predictive Directional Measurement of Volatility Spillovers. International Journal of Forecasting, 28(1), 57–66. https://doi.org/10.1016/j.ijforecast.2011.02.006
                • Lin, Găman, Wang & Pele (2026). When Does the Market React to Protocol Innovation? Evidence from the Ethereum Upgrade Lifecycle. Working paper.
                • MacKinlay, A. C. (1997). Event Studies in Economics and Finance. Journal of Economic Literature, 35(1), 13–39. https://www.jstor.org/stable/2729691
                • Mantegna, R. N. (1999). Hierarchical Structure in Financial Markets. The European Physical Journal B, 11(1), 193–197. https://doi.org/10.1007/s100510050929
                • Pele, D. T., Wesselhöfft, N., Härdle, W. K., Kolossiatis, M. & Yatracos, Y. G. (2023). Are Cryptos Becoming Alternative Assets? The European Journal of Finance, 29(10), 1064–1105. https://doi.org/10.1080/1351847X.2021.1960403