Blockchain Applications

Lecture 2: Stylized Facts, Efficiency and Bubbles

Lecture 2: Stylized Facts, Efficiency and Bubbles

Daniel Traian PELE

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

Lecture Question and Outline

Slide 2 in the PDF

The lecture asks how crypto prices behave and what that behaviour means for a business that holds, accepts or prices crypto assets. A treasurer, an exchange or a fund cannot work with general statements such as "crypto is risky"; each needs numbers: how large a typical daily move is, how often an extreme move occurs, and whether the risk changes over time.

The route has five steps. The first turns prices into returns on a market that trades seven days a week. The second describes the shape of the return distribution, with its heavy tails and frequent crashes. The third follows volatility through time and compares the behaviour of Bitcoin with that of equities and gold. The fourth tests whether past prices predict returns, the question of market efficiency. The fifth studies bubbles, drawdowns and network value through Metcalfe's law. Lecture 4 then builds models for the facts found here: generalised autoregressive conditional heteroskedasticity (GARCH) models for volatility, and value at risk (VaR) and expected shortfall (ES) for tail risk.

In brief

  • Question: how do crypto prices behave, and what does that mean for a business that holds, accepts or prices crypto?
    • a treasurer, an exchange or a fund needs numbers, not narratives
  • Outline
    • prices and returns on a 7-day market: Lecture 1 described levels and trends; today the same prices become returns
    • the shape of returns: heavy tails, rare but frequent crashes
    • volatility through time, and the link with equities and gold
    • market efficiency: can past prices predict returns?
    • bubbles, drawdowns and network value (Metcalfe's law)
  • Lecture 4: models for volatility and tail risk: generalised autoregressive conditional heteroskedasticity (GARCH), value at risk (VaR), expected shortfall (ES)

Learning Outcomes

Slide 3 in the PDF
  • Compute log returns and annualise volatility with the right number of days
  • Read a summary table: volatility, skewness, kurtosis, worst day
  • Use a histogram and a quantile–quantile (QQ) plot to show why the Normal distribution understates crash risk
  • Interpret rolling volatility and rolling correlation charts
  • Explain weak-form efficiency and read a variance-ratio test
  • Measure drawdowns and judge bubble and network-value claims with data

Abbreviations

Slide 4 in the PDF
ACF autocorrelation function
CC BY-SA 3.0 DE Creative Commons Attribution-ShareAlike 3.0 Germany licence
CI confidence interval
COVID-19 coronavirus disease 2019
CRIX CRypto IndeX (Trimborn & Härdle (2018))
EODHD EOD (end-of-day) Historical Data, data vendor
ES expected shortfall
EWMA exponentially weighted moving average
FABIZ Faculty of Business Administration in Foreign Languages, ASE
FTX crypto exchange (company name), bankrupt in November 2022
GARCH generalised autoregressive conditional heteroskedasticity
i.i.d. independent and identically distributed
IDA Institute Digital Assets
JB Jarque–Bera (test statistic)
LPPL log-periodic power law
LPPLS log-periodic power law singularity
MBB Master Blockchain for Business
QQ quantile–quantile
S&P 500 Standard & Poor's 500 US stock index
sd standard deviation
SE standard error
SFM Statistics of Financial Markets (master course; Quantlet prefix)
USD US dollar
UTC Coordinated Universal Time
VaR value at risk
VR variance ratio

Reading and Tools

Slide 5 in the PDF

Two papers are the core reading. Trimborn & Härdle (2018) construct CRIX, the value-weighted market index for cryptocurrencies already used in Lecture 1. Peterson (2018) applies Metcalfe's law, the idea that the value of a network grows with the square of its users, to Bitcoin; the last section of the lecture tests this idea on the data. For further reading, Pele et al. (2023) ask whether crypto assets are becoming alternative assets, a class of their own next to equities, bonds and commodities.

Two video units of the Quantinar course Blockchain Analytics accompany the lecture. The data are the daily closes of Bitcoin, Ethereum, the S&P 500 and gold from 2018 to 18 September 2026, from EODHD, and the daily on-chain transactions of Bitcoin from 2014 to 2025, from the ETH Chronicles project used in Lecture 1. The notebook of Lecture 2 in the course repository computes every number and chart of the lecture.

In brief

A market in which prices always “fully reflect” available information is called “efficient.”

— Fama (1970)

Lecture 1 described crypto markets through levels: market values, supplies, transactions and their growth over a decade. A business that holds, accepts or prices crypto assets needs a different kind of number. A treasurer who keeps bitcoin on the balance sheet must know how much the position can move in a day or a year; an exchange must know how often extreme days occur; a fund must know whether past prices contain information about future returns. All of these questions are about returns, the changes of prices, and about the statistical behaviour of those changes.

The lecture therefore turns the prices of Lecture 1 into daily returns and studies their properties. It measures the volatility and the shape of the return distribution, shows that the Normal distribution understates the frequency of crashes, follows volatility and the correlation with equities through time, and tests whether past returns help to predict future ones. The last part asks two valuation questions: how far prices can fall from a peak, and whether bubble models or network activity can explain the price.

The data are daily closing prices of Bitcoin, Ethereum, the S&P 500 stock index and gold from 1 January 2018 to 18 September 2026, from the data vendor EOD Historical Data (EODHD), with longer samples from 2014 for some tests, and the monthly on-chain transactions of Bitcoin from 2014 to 2025 used in Lecture 1.

Daily log returns, 1 January 2018 – 18 September 2026. Left: the histogram of Bitcoin's daily returns in % with the Normal density of the same mean and standard deviation. Middle and right: quantile–quantile plots of the standardised returns of Bitcoin and of the S&P 500 against the quantiles of the Normal distribution; points on the red line would mean Normal returns.
Daily log returns, 1 January 2018 – 18 September 2026. Left: the histogram of Bitcoin's daily returns in % with the Normal density of the same mean and standard deviation. Middle and right: quantile–quantile plots of the standardised returns of Bitcoin and of the S&P 500 against the quantiles of the Normal distribution; points on the red line would mean Normal returns.

What you will be able to do

  • Compute log returns and annualise volatility with the right number of days.
  • Read a summary table: volatility, skewness, kurtosis, worst day.
  • Use a histogram and a quantile–quantile (QQ) plot to show why the Normal distribution understates crash risk.
  • Interpret rolling volatility and rolling correlation charts.
  • Explain weak-form efficiency and read a variance-ratio test.
  • Measure drawdowns and judge bubble and network-value claims with data.

The Appendix holds optional worked examples and checks.

Lecture plan

  1. Prices and ReturnsLog returns, the price and return path of Bitcoin, and the annualisation of volatility on a market that trades every day.
  2. The Shape of ReturnsMean, volatility, skewness and kurtosis; the Jarque–Bera test; four assets compared; how rare a crash is, against the Normal distribution.
  3. Volatility and CorrelationRolling volatility, autocorrelation and the Ljung–Box test, volatility clusters and calm after calm, the correlation of Bitcoin with equities.
  4. Market EfficiencyWeak-form efficiency, the variance ratio and its robust test, sub-periods, efficiency through time and its limits.
  5. Bubbles, Crashes and Network ValueDrawdowns and the gain needed to recover, the log-periodic power law (LPPL) bubble model and Metcalfe's law of network value, in levels and in changes.
  6. SummaryThe main numbers of the lecture and five interpretation exercises.
  7. Appendix: Further MaterialTail indices, crisis episodes, safe havens, LPPL estimation, rolling estimates and worked examples.

Prices and Returns

Prices of crypto assets range from a few cents to tens of thousands of US dollars (USD) and grow by large factors over time, so their levels cannot be compared across assets or periods. Returns can. The section defines the daily log return and shows how to turn the variability of daily returns into an annual figure, a step where a wrong calendar produces a large error.

Log Returns and Annualisation

Slide 6 in the PDF

The daily log return in percent is

\[r_t = 100\,(\ln P_t - \ln P_{t-1})\]
(2.1)

where \(P_t\) is the closing price on day \(t\). Log returns add up over time, as Lecture 1 showed, and for small changes they are close to the percentage change of the price. The volatility of an asset is the standard deviation \(s\) of its daily returns: it measures how far a typical day's return lies from the average.

Volatility is usually quoted per year. If the daily returns are independent, the variance of a sum of \(A\) daily returns is \(A\) times the daily variance, so the annual volatility is

\[\sigma_a = s\sqrt{A}\]
(2.2)

where \(A\) is the number of returns in a year. Crypto trades every day of the year, so \(A = 365\); stock exchanges are closed at weekends and on holidays, so for equities \(A \approx 252\) trading days.

For Bitcoin from 2018 to 2026 the daily standard deviation is \(s = 3.355\%\). The right calculation is \(\sqrt{365} \times 3.355 = 19.10 \times 3.355 = 64.1\%\) a year. Using the equity calendar gives \(\sqrt{252} \times 3.355 = 53.3\%\), an understatement of 17%. The rule is to annualise each asset with its own calendar and to compare assets only after this step.

In brief

  • Daily log return in %: \(r_t = 100\,(\ln P_t - \ln P_{t-1})\); \(P_t\): the close on day \(t\)
  • Annual volatility from daily data: \(\sigma_a = s\sqrt{A}\)
    • \(s\): the standard deviation of daily returns; \(A\): the number of returns per year
    • crypto trades every day: \(A = 365\); stocks: \(A \approx 252\) trading days
  • Worked example: Bitcoin 2018–2026, \(s = 3.355\%\) a day
    • right: \(\sqrt{365} \times 3.355 = 19.10 \times 3.355 = 64.1\%\) a year
    • wrong: \(\sqrt{252} \times 3.355 = 53.3\%\), an understatement of 17%
  • Rule: annualise each asset with its own calendar; compare assets only after that

Bitcoin Prices and Log Returns, 2018–2026

Slide 7 in the PDF

The chart shows the same data twice. The top panel is the daily closing price of Bitcoin in USD on a log scale, from 1 January 2018 to 18 September 2026; the bottom panel is the daily log return of formula (2.1), in percent, for the same 3,183 days, with dotted lines at \(\pm 5\%\).

The two panels tell different stories. The price trends and swings: it fell below 4,000 USD at the end of 2018, rose above 60,000 USD in 2021 and above 120,000 USD in 2025. The returns have no trend at all; they stay around zero, with bursts of large moves in some periods and quiet stretches in others. On 340 days, 10.7% of the sample, the return was beyond \(\pm 5\%\).

The extremes are single days. The worst, 12 March 2020, was a log return of \(-46.5\%\), during the COVID-19 crash; the best, 8 February 2021, was \(+17.2\%\). The two panels are tied by the additivity of log returns: the sum of the 3,183 daily returns, 174.3, equals \(100\ln(P_{\text{end}}/P_{\text{start}})\), the log return of the whole period.

In brief

Bitcoin Prices and Log Returns, 2018–2026
  • Top: daily close in USD, log scale; bottom: daily log returns in %, 1 January 2018 – 18 September 2026, \(n = 3\,183\)
  • The price trends and swings; the returns stay around zero, with bursts: 340 days (10.7%) beyond \(\pm 5\%\)
  • Worst day 12 March 2020: \(-46.5\%\); best day 8 February 2021: \(+17.2\%\)
  • Log returns add up: the sum of the 3 183 daily returns, 174.3, equals \(100 \ln(P_{\text{end}}/P_{\text{start}})\)
BA_L2_returns_path

The Shape of Returns

Volatility measures the typical size of a daily move, but not how often extreme moves occur. The section describes the whole distribution of daily returns with four moments, tests whether the distribution is Normal, compares four assets and computes how rare a crash would be if returns were Normal.

Moments of the Return Distribution

Slide 8 in the PDF

Four summary statistics describe a distribution of \(n\) daily returns \(r_1, \dots, r_n\). The mean \(\bar r\) is the average return and the standard deviation \(s\) the volatility, as in the previous section. The skewness \(\hat S\) is the average of the cubed standardised returns \((r_t - \bar r)/s\): it is 0 for a symmetric distribution and negative when large losses are more frequent or larger than large gains. The excess kurtosis \(\hat K\) is the average of the standardised returns to the fourth power, minus 3: it is 0 for the Normal distribution and positive when extreme days, in either direction, are more frequent than the Normal predicts. The formulas are on the slide.

The Jarque–Bera (JB) test combines skewness and kurtosis into one test of normality:

\[\mathrm{JB} = \frac{n}{6}\Big(\hat S^2 + \frac{\hat K^2}{4}\Big)\]
(2.3)

Under the null hypothesis that returns are Normal, JB follows approximately a chi-squared distribution with 2 degrees of freedom, \(\chi^2_2\), whose 5% critical value is 5.99. Normality is rejected at the 5% level when JB exceeds 5.99.

For daily returns from 2018 to 2026 the statistic is 29,620 for Bitcoin, 16,593 for Ethereum, 20,462 for the S&P 500 and 10,041 for gold. All four are thousands of times larger than the critical value, so normality is rejected for every asset, equities and gold included.

In brief

  • For \(n\) daily returns \(r_1, \dots, r_n\):
    \[ \bar r = \frac1n\sum_{t=1}^n r_t, \qquad s^2 = \frac{1}{n-1}\sum_{t=1}^n (r_t - \bar r)^2 \]
    \[ \hat S = \frac1n\sum_{t=1}^n \Big(\frac{r_t - \bar r}{s}\Big)^3, \qquad \hat K = \frac1n\sum_{t=1}^n \Big(\frac{r_t - \bar r}{s}\Big)^4 - 3 \]
    • \(\bar r\): mean; \(s\): standard deviation (volatility); \(\hat S\): skewness; \(\hat K\): excess kurtosis
    • Normal distribution: \(S = 0\) and \(K = 0\)
  • Jarque–Bera (JB) test of normality:
    \[ \mathrm{JB} = \frac{n}{6}\Big(\hat S^2 + \frac{\hat K^2}{4}\Big) \sim \chi^2_2 \text{ under } H_0\text{: Normal} \]
    • reject normality at 5% if \(\mathrm{JB} > 5.99\)
    • 2018–2026: Bitcoin 29 620, Ethereum 16 593, S&P 500 20 462, gold 10 041: normality is rejected for all four

Worked Example: Jarque–Bera for Bitcoin

Slide 9 in the PDF

For Bitcoin from 2018 to 2026 the sample has \(n = 3{,}183\) daily returns, a skewness of \(\hat S = -0.945\) and an excess kurtosis of \(\hat K = 14.825\). Formula (2.3) gives \(3{,}183/6 = 530.5\) times the sum of \(0.945^2 = 0.893\) and \(14.825^2/4 = 54.945\), that is \(530.5 \times 55.838 \approx 29{,}620\). The p-value, the probability of a statistic at least this large if returns were Normal, is below \(10^{-300}\).

The decomposition shows where the rejection comes from: the kurtosis term contributes 54.9 of the 55.8 inside the brackets, the skewness term only 0.9. The problem with the Normal distribution is the tails, not the asymmetry.

The test needs a careful reading. With a sample of several thousand days, JB rejects even small departures from normality, so a tiny p-value alone says little about how large the departure is. The size of \(\hat K\), here almost 15, is the informative number: it says that extreme days are far more frequent than under the Normal.

In brief

  • Bitcoin, 2018–2026: \(n = 3\,183\), \(\hat S = -0.945\), \(\hat K = 14.825\)
    \[ \mathrm{JB} = \frac{3\,183}{6}\Big(0.945^2 + \frac{14.825^2}{4}\Big) = 530.5 \times (0.893 + 54.945) \approx 29\,620 \]
  • 5% critical value of \(\chi^2_2\): 5.99; p-value \(< 10^{-300}\)
  • Almost all of the statistic comes from kurtosis (54.9 of 55.8): the problem is the tails, not the asymmetry
  • Caution: with \(n\) in the thousands, JB rejects even mild deviations; the size of \(\hat K\) matters more than the p-value

Four Assets Side by Side, 2018–2026

Slide 10 in the PDF

The table compares the daily log returns of four assets from 1 January 2018 to 18 September 2026. Each asset is measured on its own calendar: Bitcoin and Ethereum have a return for every day, 3,183 in total, annualised with \(A = 365\); the S&P 500 and gold have Monday-to-Friday closes, 2,189 and 2,268 returns, annualised with \(A = 252\).

The columns give the number of returns \(n\), the daily standard deviation, the annual volatility from formula (2.2), the skewness, the excess kurtosis, the worst day as a log return in percent, and the share of days with a move larger than 5% in either direction.

Bitcoin's annual volatility is 64.1%, against 19.2% for the S&P 500: it is 3.3 times as volatile. Ethereum is more volatile still, at 84.3%, and gold is the calmest asset, at 16.9%. Bitcoin moves by more than 5% on 10.7% of days, about one day in nine, while the S&P 500 does so on 0.6% of its trading days and gold on 0.3%. The worst day was the same for both crypto assets, 12 March 2020, during the market crash at the start of the coronavirus disease 2019 (COVID-19) pandemic: \(-46.5\%\) for Bitcoin and \(-55.1\%\) for Ethereum.

All four assets have negative skewness, so large losses are more frequent or larger than large gains, and all four have large positive excess kurtosis, between 10.2 for gold and 14.9 for the S&P 500. Heavy tails are therefore not specific to crypto; what distinguishes crypto is the scale of the moves.

In brief

\(n\)sd (%)ann. vol. (%)skewnessexcess kurtosisworst day (log %)days \(|r| > 5\%\) (%)
Bitcoin3 183\(3.36\)\(64.1\)\(-0.95\)\(14.8\)\(-46.5\)\(10.7\)
Ethereum3 183\(4.41\)\(84.3\)\(-0.83\)\(11.1\)\(-55.1\)\(17.8\)
S&P 5002 189\(1.21\)\(19.2\)\(-0.65\)\(14.9\)\(-12.8\)\(0.6\)
Gold2 268\(1.07\)\(16.9\)\(-0.55\)\(10.2\)\(-11.1\)\(0.3\)
  • Daily log returns, 1 January 2018 – 18 September 2026, each asset on its own calendar
    • crypto: every day, \(A = 365\); S&P 500 and gold: Monday–Friday closes, \(A = 252\)
    • skewness: 0 for a symmetric distribution; negative: the left tail (losses) is longer
    • excess kurtosis: 0 for the Normal distribution; positive: more extreme days than the Normal predicts
  • Interpretation
    • Bitcoin is 3.3 times as volatile as the S&P 500; it moves by more than 5% on one day in nine
    • the worst day of Bitcoin and Ethereum was the same: 12 March 2020 (COVID-19 crash)
    • all four are skewed to the left and heavy-tailed

Fit of the Normal Distribution

Slide 11 in the PDF

The chart shows the departure from normality in two ways. The left panel is a histogram of Bitcoin's daily returns with the Normal density of the same mean and standard deviation drawn over it. Compared with this curve, the data have too many quiet days near zero, too many extreme days in both tails, and too few days in between.

The middle and right panels are quantile–quantile (QQ) plots. The returns are standardised, sorted from the smallest to the largest and plotted against the quantiles a Normal distribution would give for a sample of the same size. If the returns were Normal, the points would lie on the red line.

    Question for the room Each threshold is k of the asset's own standard deviations. With k = 5, which asset has more days beyond the threshold in 2018–2026, Bitcoin or the S&P 500?
    Answer

    The S&P 500: 9 days against 5 for Bitcoin, where the Normal distribution expects about 0.001 and 0.002 days. The threshold is measured in each asset's own standard deviation (1.21% a day for the S&P 500, 3.35% for Bitcoin), so the count compares the shape of the tails, not the size of the moves: the excess kurtosis is 14.9 for the S&P 500 and 14.8 for Bitcoin. In percent Bitcoin moves more: it exceeds 5% on 10.7% of the days, the S&P 500 on 0.6%.

    Both QQ plots follow the line in the centre and bend away from it at both ends: the lowest returns are much lower and the highest much higher than the Normal predicts. For Bitcoin the lowest point, 12 March 2020, lies near \(-14\) standard deviations; for the S&P 500 the lowest lies near \(-10\). Heavy tails are a feature of crypto and of equities.

    In brief

    • Left: too many quiet days in the centre and too many extreme days in the tails, too few in between
    • QQ plot: sorted returns against the quantiles a Normal distribution would give; Normal data lie on the red line
    • Both QQ plots bend away at the ends: heavy tails in crypto and in equities
    BA_L2_hist_qq

    Frequency of Crashes

    Slide 12 in the PDF

    To judge how extreme a day is, each return is standardised: \(z_t = (r_t - \bar r)/s\) counts how many standard deviations it lies from the mean. Under the Normal distribution, a move of more than 3 standard deviations in either direction has a probability of 0.27%, and a move of more than 4 standard deviations a probability of 0.0063%. In 3,183 days the Normal therefore predicts 0.2 days beyond 4 standard deviations, that is, most probably none.

    Bitcoin had 17 days beyond 4 standard deviations between 2018 and 2026. The worst day, 12 March 2020, had a log return of \(-46.5\%\), that is \(z = -13.9\). As a simple return, the change in the value held, this is \(e^{-0.465} - 1 = -37.2\%\). Under the Normal distribution the probability of a day at least this bad is about \(5 \times 10^{-44}\), which means once in \(2 \times 10^{43}\) days.

    The practical lesson is that a risk model built on the Normal distribution will be surprised many times by days it treats as practically impossible. Lecture 4 replaces the Normal with a heavy-tailed distribution and lets volatility change over time.

    In brief

    • Standardise: \(z_t = (r_t - \bar r)/s\), the number of standard deviations of each day
    • Under the Normal distribution
      • \(P(|z| > 3) = 0.27\%\); \(P(|z| > 4) = 0.0063\%\)
      • in 3 183 days we would expect 0.2 days beyond 4 standard deviations
    • Bitcoin, 2018–2026: 17 days beyond 4 standard deviations
    • Worst day, 12 March 2020: log return \(-46.5\%\), i.e. \(z = -13.9\)
      • as a simple return: \(e^{-0.465} - 1 = -37.2\%\) of the value held
      • under the Normal: \(P(z \le -13.9) \approx 5 \times 10^{-44}\), i.e. once in \(2 \times 10^{43}\) days
    • Business lesson: a risk model built on the Normal distribution will be surprised many times; Lecture 4 fixes this

    Crash Days: Observed against the Normal Distribution

    Slide 13 in the PDF

    The previous slide compared one threshold, 4 standard deviations, with the Normal distribution. The chart repeats the comparison for thresholds \(k = 2, 3, 4, 5\) and 6. For each \(k\), the first bar counts the days of 2018–2026 with \(|z_t| > k\); the second bar is the number a Normal distribution with the same mean and standard deviation would give, \(n \times P(|z| > k)\). The vertical axis is on a log scale, because the expected numbers fall from about a hundred to one in a million.

    At \(k = 2\) the Normal does well: Bitcoin had 185 such days against 145 expected, and the S&P 500 99 against 100. The moderate days are where the Normal distribution fits. Further out the gap opens quickly. At \(k = 4\) Bitcoin had 17 days against 0.20 expected, 84 times as many, and the S&P 500 15 against 0.14, 108 times as many. At \(k = 6\) Bitcoin had 1 day and the S&P 500 6 days, while the Normal expects one such day in 507 million.

    The ratio of observed to expected days rises with \(k\) for both assets. This is what a heavy tail means: the problem of the Normal distribution is not the typical day but the rare ones, which are exactly the days that matter for risk management. The S&P 500, with far smaller moves, has tails as heavy relative to its own volatility as Bitcoin's.

    In brief

    Crash Days: Observed against the Normal Distribution
    • Days with \(|z_t| > k\), \(k = 2, \dots, 6\), in 2018–2026, against \(n \times P(|z| > k)\) under the Normal distribution (log scale)
    • \(k = 2\): Bitcoin 185 days against 145 expected, S&P 500 99 against 100: the Normal fits the moderate days
    • \(k = 4\): Bitcoin 17 against 0.20 (84 times as many), S&P 500 15 against 0.14 (108 times)
    • \(k = 6\): Bitcoin 1 day, S&P 500 6 days; the Normal expects one such day in 507 million
    • The ratio of observed to expected grows with \(k\): the tails, not the centre, break the Normal distribution
    BA_L2_crash_counts

    Worked Example: Expected Crash Days under the Normal

    Slide 14 in the PDF

    The expected counts in the chart come from the Normal probabilities. For Bitcoin, 2018–2026, there are \(n = 3{,}183\) daily returns. At \(k = 3\), the probability of a move beyond 3 standard deviations in either direction is \(P(|z| > 3) = 2 \times 0.00135 = 0.00270\), so the Normal expects \(3{,}183 \times 0.00270 = 8.6\) days. The data have 59, seven times as many.

    At \(k = 4\), \(P(|z| > 4) = 2 \times 3.17 \times 10^{-5} = 6.33 \times 10^{-5}\), and the expected number is \(3{,}183 \times 6.33 \times 10^{-5} = 0.20\) days. The data have 17, 84 times as many.

    The same numbers can be stated as waiting times. Under the Normal, a day beyond 4 standard deviations comes on average once every \(1/(6.33 \times 10^{-5}) = 15{,}787\) days, about 43 years. In the data it came once every \(3{,}183/17 = 187\) days, about every six months. A risk limit calibrated on Normal tail probabilities treats a twice-a-year event as a once-in-two-generations event.

    In brief

    • Bitcoin, 2018–2026: \(n = 3\,183\) daily returns
    • Step 1, \(k = 3\): \(P(|z| > 3) = 2 \times 0.00135 = 0.00270\)
      • expected days \(= 3\,183 \times 0.00270 = 8.6\); observed 59, seven times as many
    • Step 2, \(k = 4\): \(P(|z| > 4) = 2 \times 3.17 \times 10^{-5} = 6.33 \times 10^{-5}\)
      • expected days \(= 3\,183 \times 6.33 \times 10^{-5} = 0.20\); observed 17, 84 times as many
    • Step 3, waiting time between two days with \(|z| > 4\)
      • under the Normal: \(1/(6.33 \times 10^{-5}) = 15\,787\) days, about 43 years
      • observed: \(3\,183/17 = 187\) days, about six months
    • Reading: a risk limit built on Normal tail probabilities treats a once-in-six-months event as a once-in-43-years event

    Section summary

    Daily crypto returns are far from Normal: the Jarque–Bera test rejects normality for all four assets, almost entirely because of the tails, and Bitcoin had 17 days beyond 4 standard deviations where the Normal predicts 0.2. Equities and gold also have heavy tails; crypto differs mainly by the size of its moves.

    Volatility and Correlation

    A single volatility number for a whole sample hides how risk changes over time. The section measures volatility on moving windows, tests whether returns and squared returns are autocorrelated, shows the resulting volatility clusters, and follows the correlation between Bitcoin and equities.

    Volatility Through Time

    Slide 15 in the PDF

    Rolling volatility is the standard deviation of returns on a window of recent days, recomputed each day as the window moves forward by one day. Here the window covers about one month: 30 daily returns for crypto, annualised with \(\sqrt{365}\), and 21 trading days for the S&P 500, annualised with \(\sqrt{252}\). The Appendix gives the formula and the precision of the estimate.

    The chart shows the three series from 2016 to September 2026, in percent per year. The median rolling volatility is 54% for Bitcoin, 77% for Ethereum and 12% for the S&P 500. All three reach their highest values in March–April 2020, during the COVID-19 crash, when even the S&P 500 came close to 100% a year for a short period.

      Question for the room With a 30-day window the median of Bitcoin's rolling volatility is 54% a year. The ▶ button lengthens the window from 7 to 365 days. What happens to the median?
      Answer

      It rises: 47% with 7 days, 54% with 30 days, 58% with 90 days and 66% with 365 days, close to the 64.1% of the whole sample. A short window rarely contains one of the few very large days, so most short-window estimates lie below the long-run volatility and a few lie far above it: the maximum falls from 364% (7 days) to 106% (365 days). The S&P 500 median moves the same way, from 11% to 15%.

      Bitcoin's volatility trends down over the decade: its peaks after 2023 are lower than those of 2016–2021. It nevertheless stays well above that of equities: Bitcoin's rolling volatility was below that of the S&P 500 on only 0.1% of the days. The chart also shows that high-volatility periods last for weeks or months rather than single days, a pattern the next two slides measure.

      In brief

      • Volatility on about one month: 30 daily returns \(\times\sqrt{365}\) for crypto, 21 returns \(\times\sqrt{252}\) for the S&P 500
      • Medians: 54% (Bitcoin), 77% (Ethereum), 12% (S&P 500); all three peak in March–April 2020
      • Bitcoin's volatility trends down, but it was below that of the S&P 500 on only 0.1% of the days
      SFM_ch14_asset_class

      Autocorrelation and the Ljung–Box Test

      Slide 16 in the PDF

      The sample autocorrelation at lag \(k\), \(\hat\rho(k)\), is the correlation between the return on a day and the return \(k\) days earlier; the formula is on the slide. If returns are independent, \(\hat\rho(k)\) lies within \(\pm 1.96/\sqrt n\) of zero with probability 95%. The autocorrelation function (ACF) is the list of \(\hat\rho(k)\) for \(k = 1, 2, \dots\). The ACF of the returns \(r_t\) asks whether the direction of a move is predictable; the ACF of the squared returns \(r_t^2\) asks whether the size of a move is predictable.

      The Ljung–Box test checks the first \(m\) autocorrelations together:

      \[Q(m) = n(n+2)\sum_{k=1}^{m}\frac{\hat\rho(k)^2}{n-k}\]
      (2.4)

      Under the null hypothesis of no autocorrelation, \(Q(m)\) follows approximately a \(\chi^2_m\) distribution, and large values reject.

      For Bitcoin's daily log returns from 18 September 2014 to 18 September 2026, \(n = 4{,}384\), with \(m = 10\) lags, the 5% critical value of \(\chi^2_{10}\) is 18.3. For the returns \(Q = 20.2\), with a p-value of 0.028: a rejection, but only just. For the squared returns \(Q = 212.7\), more than ten times the critical value. Returns have weak dependence; squared returns have strong dependence.

      In brief

      • Sample autocorrelation at lag \(k\):
        \[ \hat\rho(k) = \frac{\sum_{t=k+1}^{n}(r_t - \bar r)(r_{t-k} - \bar r)}{\sum_{t=1}^{n}(r_t - \bar r)^2} \]
        • under independence, \(\hat\rho(k)\) lies within \(\pm 1.96/\sqrt n\) with probability 95%
        • autocorrelation function (ACF) of \(r_t\): is the direction predictable? ACF of \(r_t^2\): is the size of the move predictable?
      • Ljung–Box test of the first \(m\) autocorrelations together:
        \[ Q(m) = n(n+2)\sum_{k=1}^{m}\frac{\hat\rho(k)^2}{n-k} \sim \chi^2_m \text{ under } H_0\text{: no autocorrelation} \]
        • Bitcoin daily log returns, 18 September 2014 – 18 September 2026, \(n = 4\,384\), \(m = 10\) lags
        • \(Q = 20.2\) (p-value 0.028) for \(r_t\); \(Q = 212.7\) for \(r_t^2\); the 5% critical value of \(\chi^2_{10}\) is 18.3
        • weak dependence in returns, strong dependence in squared returns

      Worked Example: Ljung–Box with Two Lags

      Slide 17 in the PDF

      Formula (2.4) can be computed by hand for \(m = 2\) lags: \(Q(2) = n(n+2)\big[\hat\rho(1)^2/(n-1) + \hat\rho(2)^2/(n-2)\big]\). For Bitcoin's daily log returns since 2014, \(n = 4{,}384\).

      For the returns, \(\hat\rho(1) = -0.0223\) and \(\hat\rho(2) = 0.0121\), so the squared autocorrelations are 0.000498 and 0.000147. Then \(Q(2) = 4{,}384 \times 4{,}386 \times (0.000498/4{,}383 + 0.000147/4{,}382) = 2.83\). The 5% critical value of \(\chi^2_2\) is 5.99, and the p-value is 0.24: we do not reject the hypothesis of no autocorrelation at two lags.

      For the squared returns, \(\hat\rho(1) = 0.1367\) and \(\hat\rho(2) = 0.0527\), and the same computation gives \(Q(2) = 94.2\), with a p-value of about \(4 \times 10^{-21}\): the hypothesis is rejected. Because \(n(n+2)/(n-1) \approx n\), each term is close to \(n\,\hat\rho(k)^2\), and the first term alone, \(4{,}384 \times 0.0187 = 81.9\), already decides the test.

      The same conclusion appears one lag at a time. The 95% band of a single autocorrelation is \(\pm 1.96/\sqrt{4{,}384} = \pm 0.030\); the lag-1 autocorrelation of the returns lies inside it, that of the squared returns more than four times outside.

      In brief

      • Bitcoin daily log returns since 2014, \(n = 4\,384\); \(Q(2) = n(n+2)\big[\hat\rho(1)^2/(n-1) + \hat\rho(2)^2/(n-2)\big]\)
      • Returns: \(\hat\rho(1) = -0.0223\), \(\hat\rho(2) = 0.0121\)
        \[ Q(2) = 4\,384 \times 4\,386 \times \Big(\frac{0.000498}{4\,383} + \frac{0.000147}{4\,382}\Big) = 2.83 \]
        • below 5.99, the 5% critical value of \(\chi^2_2\) (p-value 0.24): we do not reject “no autocorrelation”
      • Squared returns: \(\hat\rho(1) = 0.1367\), \(\hat\rho(2) = 0.0527\), so \(Q(2) = 94.2\) (p-value \(4 \times 10^{-21}\)): rejected
        • the first term alone, \(\approx n\,\hat\rho(1)^2 = 4\,384 \times 0.0187 = 81.9\), decides the test
      • One lag at a time: the 95% band is \(\pm 1.96/\sqrt{4\,384} = \pm 0.030\); \(\hat\rho(1)\) of returns lies inside it, that of squared returns far outside

      Volatility Clusters

      Slide 18 in the PDF
        Question for the room With m = 10 lags the Ljung–Box test rejects for Bitcoin returns: Q(10) = 20.2 against a critical value of 18.3. The ▶ button raises m from 1 to 30. For which m does the test on returns reject at 5%?
        Answer

        Only for m = 10 and 11 and for m = 21 to 29. Q(m) grows with m, but so does the 5% critical value of χ²(m): at m = 12, Q = 20.6 against 21.0; at m = 30, Q = 43.5 against 43.8 (p-value 0.052). For returns the verdict depends on the number of lags, so the evidence of a predictable direction is weak. For squared returns Q(m) runs from 82 (m = 1) to 296 (m = 30), and every m rejects.

        The chart shows the two autocorrelation functions for lags of 1 to 30 days; the dashed lines are the 95% band \(\pm 1.96/\sqrt n\) under independence. In the left panel, the autocorrelations of the returns are small and mostly inside the band, with a few bars just outside it: the direction of tomorrow's move is hard to predict from past moves.

        In the right panel, the autocorrelations of the squared returns are all positive and many lie above the band, the largest above 0.1 at lags 1 and 7. The size of tomorrow's move is therefore partly predictable: calm days tend to follow calm days, and turbulent days follow turbulent days.

        This pattern is called volatility clustering. It explains the long high-volatility periods in the previous chart and is the basis of the GARCH models of Lecture 4, which forecast tomorrow's variance from today's squared return and today's variance.

        In brief

        • Autocorrelation at lag \(k\): the correlation between today's value and the value \(k\) days ago; dashed: the 95% band of pure noise
        • Returns (left): almost no autocorrelation; the direction of tomorrow's move is hard to predict
        • Squared returns (right): clearly positive; the size of tomorrow's move is predictable: calm days follow calm days, turbulent days follow turbulent days
        • This is volatility clustering, the basis of the GARCH models of Lecture 4
        SFM_ch14_garch_efficiency

        Absolute Returns after Calm and Turbulent Days

        Slide 19 in the PDF
          Question for the room The ▶ button moves the lag L from 1 to 60 days: the days are sorted by the absolute return L days earlier. At L = 30, how does Bitcoin's mean absolute return after the most turbulent fifth compare with that after the calmest fifth?
          Answer

          Still higher: 2.39% against 2.07% a day, 1.15 times, against 1.49 times at L = 1; at L = 60 the ratio is 1.03. For the S&P 500 it is still 1.39 at L = 30 (1.02% against 0.73%) and 0.91 at L = 60. A turbulent day raises the expected size of moves for weeks: volatility has a long memory, which the GARCH models of Lecture 4 describe with a slowly decaying variance. Turbulence is not followed by unusual calm: the bars do not turn upside down.

          The autocorrelations of squared returns measure volatility clustering through a correlation coefficient. The chart shows the same fact more directly. The days of 2018–2026 are sorted into five groups of equal size, quintiles, by the absolute return of the previous day, from the calmest fifth to the most turbulent fifth. Each bar is the mean absolute return today in that group; the dashed line is the mean over all days.

          If yesterday's move carried no information, the five bars would all be at the dashed line. They are not. For Bitcoin, the mean absolute return is 1.86% after the calmest days and 2.77% after the most turbulent, 1.49 times as much; the mean over all days is 2.21%. For the S&P 500 the figures are 0.65% and 1.17%, 1.80 times as much, around a mean of 0.79%.

          The size of yesterday's move therefore says something about the size of today's move, although not about its direction. This is the volatility clustering of the previous chart, expressed in percent a day, and it is the property that the GARCH models of Lecture 4 use to forecast volatility.

          In brief

          • Days of 2018–2026 sorted into five equal groups (quintiles) by yesterday's absolute return; bars: today's mean absolute return
          • Bitcoin: 1.86% after the calmest days, 2.77% after the most turbulent, 1.49 times as much; all days 2.21%
          • S&P 500: 0.65% and 1.17%, 1.80 times as much; all days 0.79%
          • Yesterday's size says something about today's size, not about its sign: the volatility clustering seen in the ACF of \(r_t^2\)
          BA_L2_calm

          Correlation of Bitcoin with Equities

          Slide 20 in the PDF

          A rolling correlation measures the correlation of daily returns on a moving window, here 250 common trading days, about one year of equity trading. Because Bitcoin trades every day and the S&P 500 only on weekdays, the prices are first joined on the days when both are quoted and the returns are computed afterwards; Monday's Bitcoin return then covers Friday to Monday, like the equity return. The Appendix discusses this choice.

          The chart shows three rolling correlations: Bitcoin with the S&P 500, Ethereum with the S&P 500 and Bitcoin with gold; vertical lines mark the COVID-19 crash, the collapse of the stablecoin TerraUSD and its sister token Luna (Terra/Luna), the bankruptcy of the crypto exchange FTX, and the approval of spot Bitcoin exchange-traded funds (ETFs).

            Question for the room With 250 common days the highest rolling correlation of Bitcoin and the S&P 500 is 0.57 (February 2023). What is the highest value with a 30-day window?
            Answer

            0.79 (May 2022), and the lowest falls from −0.16 to −0.64. A correlation estimated on 30 days has a large sampling error (about 1/√30 = 0.18 when the true correlation is zero), so short windows give more extreme highs and lows; the 500-day window gives 0.47 and −0.07. The correlation from 2020 onwards, 0.39, does not depend on the window.

            Before 2020 Bitcoin had no link with the S&P 500: the average correlation was 0.02. From 2020 the average was 0.39, with a peak of 0.57 in February 2023, and Ethereum followed the same path. Since 2020 Bitcoin co-moves with equities in the way a risk asset does, an asset that rises when investors take more risk and falls when they reduce it; in 2022 both fell while interest rates rose. After 2020 the correlation of Bitcoin with gold stayed below its correlation with equities. The chart shows co-movement, not its cause.

            In brief

            • Correlation of daily returns over the last 250 common trading days, about one year (method: Appendix)
            • Before 2020: no link with the S&P 500 (0.02); after: 0.39 on average, peak 0.57 in February 2023
            • Since 2020 Bitcoin co-moves with equities, as a risk asset would; in 2022 both fell while interest rates rose
            SFM_ch14_correlation

            Section summary

            Volatility changes strongly over time and comes in clusters: squared returns are clearly autocorrelated even though returns themselves are hardly predictable. Since 2020 Bitcoin's returns have been correlated with equity returns, around 0.4 on average, so a crypto position adds to equity risk rather than offsetting it.

            Market Efficiency

            If past prices helped to predict returns, a trader could profit from the pattern. The section states what weak-form efficiency means, introduces the variance ratio as a test of one of its implications, applies the test to Bitcoin, Ethereum and the S&P 500 over the full sample and on moving windows, and lists what the test leaves open.

            Weak-Form Efficiency and the Variance Ratio

            Slide 21 in the PDF

            Fama (1970) classifies market efficiency by the information that prices reflect. In a weak-form efficient market, past prices do not help to predict future returns. The argument is competitive: if a pattern in past prices predicted returns, traders would buy or sell on it, and their trades would move prices until the pattern disappeared.

            The variance ratio (VR) of Lo & MacKinlay (1988) tests one implication of this idea. It compares the variance of returns over \(q\) days with \(q\) times the variance of daily returns:

            \[\mathrm{VR}(q) = \frac{\mathrm{Var}(r_t + r_{t-1} + \dots + r_{t-q+1})}{q\,\mathrm{Var}(r_t)}\]
            (2.5)

            If returns are uncorrelated, as under a random walk of log prices, the variance of a \(q\)-day return is exactly \(q\) times the daily variance and \(\mathrm{VR}(q) = 1\). If positive days tend to follow positive days (trends, or momentum), multi-day moves are larger than this and \(\mathrm{VR}(q) > 1\); if moves tend to reverse (mean reversion), \(\mathrm{VR}(q) < 1\).

            The test statistic \(Z^*(q)\) of Lo & MacKinlay (1988) compares the estimated ratio with 1 and remains valid when volatility clusters, as it does here; the Appendix gives the formula. The hypothesis \(\mathrm{VR}(q) = 1\) is rejected at the 5% level when \(|Z^*| > 1.96\). A non-rejection says that no linear autocorrelation of returns up to horizon \(q\) was detected, which is only one implication of weak-form efficiency.

            In brief

            • Fama (1970): in a weak-form efficient market, past prices do not help to predict future returns
              • if they did, traders would exploit the pattern until it disappeared
            • Variance ratio (VR) Lo & MacKinlay (1988): compare the variance of \(q\)-day returns with \(q\) times the daily variance
              \[ \mathrm{VR}(q) = \frac{\mathrm{Var}(r_t + r_{t-1} + \dots + r_{t-q+1})}{q\,\mathrm{Var}(r_t)} \]
              • random walk (no predictability): \(\mathrm{VR} = 1\)
              • \(\mathrm{VR} > 1\): trends (momentum); \(\mathrm{VR} < 1\): reversals (mean reversion)
              • a test statistic \(Z^*\), robust to volatility clustering, rejects \(\mathrm{VR}(q) = 1\) at 5% if \(|Z^*| > 1.96\)
            • The VR test checks one implication of weak-form efficiency: no linear autocorrelation of returns up to horizon \(q\)

            Variance-Ratio Tests

            Slide 22 in the PDF

            The table reports the variance ratios at horizons of 2, 5 and 10 days and the robust statistic \(Z^*(5)\) with its p-value, for daily log returns from 2014 (Ethereum from 2016) to 18 September 2026.

            All variance ratios lie close to 1. Bitcoin's are 0.977, 0.991 and 1.027, Ethereum's 0.988, 1.067 and 1.137. The S&P 500 has ratios below 1, from 0.875 at 2 days to 0.813 at 10 days, a sign of mild mean reversion in the point estimates. The robust test does not reject \(\mathrm{VR}(5) = 1\) at 5% for any asset: \(Z^*(5)\) is \(-0.19\) for Bitcoin (p-value 0.853), 1.17 for Ethereum (0.241) and \(-1.28\) for the S&P 500 (0.201). The same holds in each of the sub-periods 2014–2017, 2018–2021 and 2022–2026.

            The correct reading is limited. We do not detect the tested deviation from a random walk at these horizons, but a non-rejection does not prove that the market is efficient. Whether a trading strategy would be profitable after trading costs and an adjustment for risk is a separate question that this test does not answer.

            In brief

            VR(2)VR(5)VR(10)\(Z^*(5)\) (p-value)
            Bitcoin\(0.977\)\(0.991\)\(1.027\)\(-0.19\) (0.853)
            Ethereum\(0.988\)\(1.067\)\(1.137\)\(1.17\) (0.241)
            S&P 500\(0.875\)\(0.849\)\(0.813\)\(-1.28\) (0.201)
            • Daily log returns since 2014 (Ethereum since 2016) to 18 September 2026
            • Interpretation
              • all variance ratios are close to 1; no test rejects \(\mathrm{VR}(5) = 1\) at 5%
              • the same holds in each sub-period 2014–2017, 2018–2021, 2022–2026
              • we do not detect the tested deviation from a random walk at these horizons; this does not prove efficiency
              • profitability after trading costs and risk adjustment is a separate question
            SFM_ch14_garch_efficiency

            Variance Ratios by Sub-Period

            Slide 23 in the PDF
              Question for the room The ▶ button raises the horizon q from 2 to 30 days. Over these 29 horizons and the nine bars, 261 tests at 5%, how many times does a bar cross ±1.96?
              Answer

              10 times, 3.8% of the tests: the S&P 500 in 2018–2021 at q = 2 (Z* = −2.31), and Ethereum in 2016–2017 for q = 22 to 30 (Z* up to 2.48 at q = 30). Bitcoin never crosses. Under a random walk about 5% of 261 tests, 13, would reject by chance, and tests at neighbouring horizons are strongly dependent, so the nine rejections of Ethereum at q = 22–30 are one finding, not nine. Not rejecting at q = 5 does not settle every horizon, but these rejections are no evidence of weekly trends either.

              The table above stated that no sub-period rejects the random walk. The chart shows the numbers behind that statement: the robust statistic \(Z^*(5)\) for Bitcoin, Ethereum and the S&P 500 in three sub-periods, 2014–2017 (Ethereum from 2016), 2018–2021 and 2022 to 18 September 2026. The dashed lines at \(\pm 1.96\) are the 5% rejection limits.

              For Bitcoin, \(Z^*(5)\) is \(-0.23\), \(-0.25\) and \(+0.18\); for Ethereum \(+1.03\), \(-0.04\) and \(+0.20\); for the S&P 500 \(-0.68\), \(-0.95\) and \(-0.85\). No bar crosses the limits, so we do not reject the random walk in any sub-period for any of the three assets.

              The S&P 500 has variance ratios below 1 in every sub-period, \(\mathrm{VR}(5) = 0.92\), 0.78 and 0.92: the point estimates show the mild mean reversion typical of an index, but the robust test does not find it significant. Crypto returns have no consistent sign at all.

              In brief

              • Robust \(Z^*(5)\) in three sub-periods, to 18 September 2026 (Ethereum from 2016); dashed: \(\pm 1.96\)
              • Bitcoin \(-0.23\), \(-0.25\), \(+0.18\); Ethereum \(+1.03\), \(-0.04\), \(+0.20\); S&P 500 \(-0.68\), \(-0.95\), \(-0.85\)
              • No bar crosses \(\pm 1.96\): we do not reject the random walk in any sub-period
              • S&P 500: \(\mathrm{VR}(5)\) is 0.92, 0.78 and 0.92, below 1 each time; mild mean reversion of an index, not significant
              BA_L2_vr_periods

              Efficiency Through Time

              Slide 24 in the PDF
                Question for the room With q = 5 no 500-day window rejects VR(5) = 1 for any of the three assets. Switch the horizon to q = 2. Which asset has the largest share of rejecting windows?
                Answer

                The S&P 500: 24 of 120 windows (20%) reject VR(2) = 1, all with Z* below −1.96 (lowest −2.73); Ethereum 2 of 163, Bitcoin none of 185. VR(2) − 1 is close to the lag-1 autocorrelation, and the daily S&P 500 index has a negative one (VR(2) = 0.875 over 2014–2026). Volatility does not make returns predictable: Bitcoin, the most volatile asset, has no rejecting window.

                A test on the full sample could hide periods of predictability that cancel out. The chart therefore computes the robust \(Z^*(5)\) on moving windows of 500 days, one window every 21 days, for Bitcoin, Ethereum and the S&P 500; the dashed lines at \(\pm 1.96\) are the 5% rejection limits.

                No 500-day window rejects \(\mathrm{VR}(5) = 1\) for Bitcoin at 5%: the statistic stays between about \(-1.5\) and 1.1. The other two series also stay inside the limits, with the S&P 500 mostly below zero, in line with its variance ratios below 1.

                The result fits the literature. Early studies, such as Urquhart (2016), found Bitcoin inefficient in samples ending in 2013–2016, and Tran & Leirvik (2020) find that crypto markets became more efficient as they grew.

                In brief

                • Robust \(Z^*(5)\) on windows of 500 days, one window every 21 days; dashed: \(\pm 1.96\)
                • No 500-day window rejects \(\mathrm{VR}(5) = 1\) for Bitcoin at 5%
                • Early studies found Bitcoin inefficient before 2013–2016 Urquhart (2016); markets became more efficient as they grew Tran & Leirvik (2020)
                SFM_ch14_garch_efficiency

                Interpreting the Efficiency Tests

                Slide 25 in the PDF

                Not rejecting \(\mathrm{VR} = 1\) means that no linear predictability of returns from past returns was detected at the horizons tested. Other patterns are not tested: nonlinear dependence, such as the volatility clustering of the previous section, and predictability conditional on trading volume or news. Even full weak-form efficiency, moreover, says nothing about whether the price is right.

                A price can be unpredictable and still be far from any fundamental value. For a share, the fundamental value is the present value of its future dividends or profits. Bitcoin has no cash flows, no dividends and no profits to discount, so a "fundamental value" must come from its use: payments, settlement, collateral, or the size of its network.

                Two questions remain for the rest of the lecture. How far can the price fall from a peak, which is measured by drawdowns? Can a bubble be detected, and does network activity justify the price? The log-periodic power law (LPPL) and Metcalfe's law address the second question.

                In brief

                • Not rejecting \(\mathrm{VR} = 1\): no detectable linear predictability of returns from past returns at these horizons
                  • other patterns (nonlinear, or conditional on volume or news) are not tested
                  • even full weak-form efficiency says nothing about whether the price is right
                • A price can be unpredictable and still far from any fundamental value
                  • Bitcoin has no cash flows: no dividends, no profits to discount
                  • so “fundamental value” must come from use: payments, settlement, collateral, network size
                • Two questions remain
                  • how far can the price fall from a peak? (drawdowns)
                  • can we detect a bubble, and does network activity justify the price? (log-periodic power law (LPPL), Metcalfe)

                Section summary

                Variance-ratio tests do not reject a random walk for Bitcoin, Ethereum or the S&P 500 at horizons of 2 to 10 days, on the full sample, in sub-periods or on 500-day windows. This is evidence against simple linear predictability, not proof of efficiency, and it says nothing about whether the price level is justified.

                Bubbles, Crashes and Network Value

                The section turns from returns to price levels. It measures the depth and duration of falls from a peak, fits a model of speculative bubbles to the Bitcoin boom of 2017 and checks how stable its forecast is, and tests whether Bitcoin's market value follows the size of its network as Metcalfe's law states.

                Drawdown from the Running Maximum

                Slide 26 in the PDF

                The drawdown on day \(t\) is the loss from the highest price reached so far:

                \[D_t = \frac{P_t}{\max_{s \le t} P_s} - 1\]
                (2.6)

                It is 0 on a day with a new record high and negative otherwise. The maximum drawdown is the deepest fall over the sample, and the time under water is the number of days from a peak until the price sets a new record.

                Losses and gains are not symmetric. After a loss of \(d\), the price must rise by \(d/(1 - d)\) to return to its peak. A loss of 50% needs a gain of 100%, a loss of 75% a gain of 300%, and a loss of 83.4%, Bitcoin's maximum drawdown in the next chart, a gain of 502%.

                Managers care about drawdowns because a firm that holds crypto on its balance sheet reports the value of the position at the reporting date; a deep drawdown appears in the accounts, whatever the average return over the years.

                In brief

                • Drawdown on day \(t\): \(D_t = P_t/\max_{s \le t} P_s - 1\), the loss from the last record high
                  • maximum drawdown: the deepest fall; time under water: the days until a new record
                • A loss of \(d\) needs a gain of \(d/(1 - d)\) to recover
                  • \(d = 50\%\): \(+100\%\); \(d = 75\%\): \(+300\%\); \(d = 83.4\%\): \(+502\%\)
                • Why managers care: a firm holding crypto on its balance sheet reports the drawdown, not the average return

                The Gain Needed to Recover

                Slide 27 in the PDF

                The chart draws the relation between a drawdown \(d\) and the gain \(d/(1-d)\) that brings the price back to its record, as a red curve on a log scale. The dotted green line is the gain equal to the loss that intuition suggests. The points mark the maximum drawdowns of three assets, measured on daily closes since 2016.

                The S&P 500 fell by at most 33.9%, in March 2020; a gain of 51% recovered it. Bitcoin fell by 83.4% by December 2018, which required a gain of 502%, and Ethereum by 94.0% at the same time, which required 1,556%.

                The curve always lies above the dotted line and bends upward: every extra point of loss needs more than one point of gain, and beyond a loss of 50% the gain needed exceeds 100%. For deep falls the curve explodes, which is why the maximum drawdowns of crypto translate into years under water.

                In brief

                The Gain Needed to Recover
                • Red: the gain \(d/(1 - d)\) that recovers a drawdown \(d\), log scale; dotted green: a gain equal to the loss
                • Maximum drawdowns of the daily closes since 2016
                  • S&P 500: 33.9% (March 2020), recovered by a gain of 51%
                  • Bitcoin: 83.4% (December 2018), gain needed 502%; Ethereum: 94.0% (December 2018), 1 556%
                • Every extra point of loss needs more than one point of gain; beyond a loss of 50% the gain needed exceeds 100%
                BA_L2_recovery

                Worked Example: a Drawdown and Its Recovery

                Slide 28 in the PDF

                Bitcoin set a record of 67,567 USD on 8 November 2021 and fell to 15,787 USD on 21 November 2022. The first step is the depth of the fall, \(d = 1 - 15{,}787/67{,}567 = 76.6\%\). The second step is the gain needed to recover, \(d/(1-d) = 0.766/0.234 = 328\%\), which is the same as \(67{,}567/15{,}787 - 1\): the price had to multiply by 4.28.

                The third step is the time. The price set a new record on 4 March 2024, 847 days after the peak and 469 days after the trough. The fourth step measures the speed of the recovery: a factor of 4.28 in 469 days, or 1.28 years, is a growth rate of \(4.28^{1/1.28} - 1 = 210\%\) a year from the trough. At a steady 50% a year, the same recovery would have taken \(\ln 4.28/\ln 1.5 = 3.6\) years.

                The time under water therefore depends on how fast the price rises after the trough, which nobody knows at the trough. A treasurer who holds crypto through a drawdown must be able to wait for years, not months.

                In brief

                • Bitcoin: peak on 8 November 2021 at 67 567 USD; trough on 21 November 2022 at 15 787 USD
                • Step 1, depth: \(d = 1 - 15\,787/67\,567 = 76.6\%\)
                • Step 2, gain needed: \(d/(1 - d) = 0.766/0.234 = 328\%\), the same as \(67\,567/15\,787 - 1\)
                • Step 3, time: new record on 4 March 2024, 847 days after the peak and 469 days after the trough
                • Step 4, speed of the recovery: \(4.28^{1/1.28} - 1 = 210\%\) a year from the trough
                  • at a steady 50% a year the same recovery would take \(\ln 4.28/\ln 1.5 = 3.6\) years
                • Reading: the time under water depends on how fast prices rise after the trough, which no one knows at the trough

                Drawdowns of Bitcoin, Ethereum and the S&P 500

                Slide 29 in the PDF
                  Question for the room Bitcoin spent 76% of the days since 2016 more than 10% below its last record. On what share of the days was it more than 50% below its record?
                  Answer

                  34%, about one day in three; Ethereum 52%, the S&P 500 none (its maximum drawdown is −33.9%). The share falls much more slowly than the threshold rises, because Bitcoin's deep drawdowns lasted for long periods (2018–2020 and 2022–2023). Recovering from −50% needs a gain of 100%.

                  The chart plots the drawdown, formula (2.6), of Bitcoin, Ethereum and the S&P 500 from the daily closes since 2016. A line at 0 means the asset is at a record high; the depth below 0 is the percentage loss from the last record.

                  The maximum drawdowns were \(-83.4\%\) for Bitcoin, \(-94.0\%\) for Ethereum and \(-33.9\%\) for the S&P 500. The two crypto assets had two long periods far below their records, from 2018 to 2020 and from 2022 to 2023, while the S&P 500's deepest fall, in March 2020, was recovered within months.

                  The time spent below the record is as informative as the depth: Bitcoin spent 76% of the days more than 10% below its record high, the S&P 500 16%. Holding Bitcoin means being in a substantial loss from the peak most of the time.

                  In brief

                  • Daily closes since 2016; maximum drawdowns: Bitcoin \(-83.4\%\), Ethereum \(-94.0\%\), S&P 500 \(-33.9\%\)
                  • Bitcoin spent 76% of the days more than 10% below its record; the S&P 500 16%
                  SFM_ch14_bubbles_etf

                  Bubbles and Crashes

                  Slide 30 in the PDF

                  A bubble is a price far above any fundamental value, sustained by the expectation that the asset can be sold later at an even higher price. Cheah & Fry (2015) find that the price of Bitcoin shows speculative bubbles.

                  Johansen, Ledoit & Sornette (2000) model how such a bubble ends. Traders imitate each other, and imitation creates positive feedback: rising prices attract buyers, who push prices higher. Prices then grow faster than exponentially, which means that the growth rate itself grows. In the model, the hazard rate of a crash, the probability of a crash in the next instant given that none has happened yet, rises as the price approaches a critical time \(t_c\).

                  Markets are organised in a hierarchy, from individual traders to funds to whole markets, and imitation acts at each level. This hierarchy adds oscillations to the price path, and the oscillations speed up as \(t_c\) approaches. The result is the log-periodic power law (LPPL), written out on the next slide. The photograph shows Didier Sornette, a co-author of the model.

                  In brief

                  • Bubble: a price far above any fundamental value, sustained by the expectation of selling higher
                  • The idea of Johansen, Ledoit & Sornette (2000): imitation between traders creates positive feedback
                    • prices grow faster than exponentially: the growth rate itself grows
                    • the crash hazard rises as the price approaches a critical time \(t_c\)
                    • the hierarchy of traders (individuals, funds, markets) adds oscillations that speed up near \(t_c\)
                  • The result is the log-periodic power law (LPPL), next slide
                  Bubbles and Crashes Didier Sornette, co-author of the LPPL model Photo: Didier Sornette (2012); CC BY-SA 3.0 DE; Wikimedia Commons

                  The LPPL Model

                  Slide 31 in the PDF

                  The LPPL model describes the expected log price before the critical time, for \(t < t_c\):

                  \[\ln P_t = A + B\,(t_c - t)^m + C\,(t_c - t)^m \cos\!\big(\omega \ln(t_c - t) - \phi\big) + \varepsilon_t\]
                  (2.7)

                  where \(\varepsilon_t\) is the error. The critical time \(t_c\) is the end of the bubble regime implied by the fitted model, not a guaranteed crash date, and \(A\) is the log price the model reaches at \(t_c\). The term \(B(t_c - t)^m\) produces the super-exponential growth: with \(B < 0\) and \(0 < m < 1\), \((t_c - t)^m\) shrinks to 0 as \(t\) approaches \(t_c\), so the log price rises towards \(A\) at an increasing speed. The cosine term adds the oscillations, with amplitude \(C\), log-frequency \(\omega\) and phase \(\phi\).

                  The oscillations are log-periodic: the cosine depends on \(\ln(t_c - t)\), not on \(t\). Its peaks come at times \(t_k\) with \((t_c - t_{k+1})/(t_c - t_k) = e^{-2\pi/\omega}\), so the gaps between successive peaks shrink by a constant factor and the oscillations accelerate towards \(t_c\). The Appendix works this out for the Bitcoin fit.

                  The fit imposes ranges on the nonlinear parameters, \(0.1 \le m \le 0.9\) and \(6 \le \omega \le 13\), and requires \(t_c\) to fall after the last observation. After the fit, two conditions must hold to call the episode a bubble: \(\hat B < 0\), for super-exponential growth, and \(|\hat C|/|\hat B| < 1\), so that the oscillations do not dominate the growth.

                  In brief

                  • Log price before the critical time, \(t < t_c\); the model describes \(\mathrm{E}[\ln P_t]\), and \(\varepsilon_t\) is the error:
                    \[ \ln P_t = A + B\,(t_c - t)^m + C\,(t_c - t)^m \cos\!\big(\omega \ln(t_c - t) - \phi\big) + \varepsilon_t \]
                  • Parameters
                    • \(t_c\): critical time, the end of the bubble regime implied by the fitted model, not a guaranteed crash date; \(A\): the log price reached at \(t_c\)
                    • \(B < 0\) and \(0 < m < 1\): super-exponential growth; \((t_c - t)^m\) shrinks to 0 as \(t \to t_c\)
                    • \(C\): amplitude of the oscillations; \(\omega\): their log-frequency; \(\phi\): their phase
                  • Log-periodic: the peaks of the cosine come at times \(t_k\) with \((t_c - t_{k+1})/(t_c - t_k) = e^{-2\pi/\omega}\)
                    • the gaps between peaks shrink geometrically: the oscillations accelerate towards \(t_c\)
                  • Constraints in the estimation: \(0.1 \le m \le 0.9\), \(6 \le \omega \le 13\), \(t_c\) after the last observation
                  • Bubble filter on the fitted parameters: \(\hat B < 0\) and \(|\hat C|/|\hat B| < 1\)

                  Bitcoin 2017: an LPPL Fit

                  Slide 32 in the PDF

                  The chart fits formula (2.7) to the daily log close of Bitcoin from 1 January to 1 December 2017, \(n = 335\) days. The estimation has two steps, described in the Appendix: least squares for the linear parameters inside a search over \((t_c, m, \omega)\), started from many points because the objective has local minima.

                  The orange line is the Bitcoin price on a log scale, the blue line the LPPL fit, the green dotted line the end of the estimation window and the red dashed line the estimated critical time, \(\hat t_c = \) 25 December 2017. The actual peak came on 16 December 2017, at 19,497 USD, followed by a fall of 83%.

                    Question for the room The estimated tc is 25 December 2017, with R² = 0.982. The ▶ button moves tc from 2 December 2017 to 31 March 2018, with m and ω fixed. How does R² change?
                    Answer

                    Very little: it stays between 0.962 (tc = 31 March 2018) and 0.982 (25 December 2017); it is 0.975 for tc = 2 December and 0.981 for 1 January. A fit this good for every tc over four months cannot date the crash: the window determines the shape of the past rise, not the critical time. The actual peak, 16 December 2017, has no special place in this R² profile.

                    The fit is close, with \(R^2 = 0.98\): 98% of the variance of the log price in the window is explained by the model. The window, however, was chosen after the crash was known. A good retrospective fit is not a dated forecast: it does not show that the same model, fitted in real time, would have given the right date. The next slide repeats the fit with windows that end at different dates.

                    In brief

                    • Daily log close, window 1 January – 1 December 2017 (\(n = 335\)); two-step estimation, least squares inside a search over \((t_c, m, \omega)\) (Appendix)
                    • Estimated \(t_c\) = 25 December 2017; the actual peak came on 16 December 2017 (19 497 USD), followed by a fall of 83%
                    • \(R^2 = 0.98\) on a window chosen after the crash: a good retrospective fit, not a dated forecast (next slide)
                    BA_L2_lppl_btc2017

                    \(\hat t_c\) across Estimation Windows

                    Slide 33 in the PDF
                      Question for the room The ▶ button adds the eleven weekly windows one by one, from the window ending 1 October to the one ending 10 December 2017. Does the estimated tc move towards the actual peak, 16 December 2017, as the window end approaches it?
                      Answer

                      No. For the first nine windows the estimated tc stays between 1 and 7 January 2018, with m = 0.9 and ω = 13 at the bounds of the search; the window ending 3 December gives the only interior fit (tc = 28 December), and the window ending 10 December puts tc at the window end itself. More data near the peak did not narrow down the date: the estimates are driven by the bounds of the search.

                      The chart repeats the LPPL fit for eleven windows that all start on 1 January 2017 and end on successive weeks from 1 October to 10 December 2017. The horizontal axis is the end of the window and the vertical axis the estimated critical time \(\hat t_c\); the orange dashed line is the actual peak of 16 December 2017, and the green dotted line the window end, below which \(\hat t_c\) cannot fall.

                      Nine of the eleven fits, marked with red circles, sit at the bounds of the search, \(m = 0.9\) and \(\omega = 13\), with \(\hat t_c\) between 1 and 7 January 2018. A fit at a bound means that the data push the parameter beyond the allowed range, so the estimate is poorly identified.

                      Only the window ending on 3 December gives an interior fit, with \(\hat t_c\) = 28 December 2017. The window ending on 10 December collapses onto its own last day, a degenerate solution. In real time, therefore, the model would have produced a series of dates in late December and early January, and only one of the fits would have passed a check of its parameters.

                      In brief

                      • Eleven windows, all starting 1 January 2017, ending weekly from 1 October to 10 December 2017
                      • Nine of eleven fits sit at the bounds \(m = 0.9\), \(\omega = 13\) (red circles): the parameters are poorly identified; their \(\hat t_c\) lie between 1 and 7 January 2018
                      • Only the window ending 3 December gives an interior fit (\(\hat t_c\) = 28 December 2017)
                      • Window ending 10 December: \(\hat t_c\) equals the last observation, a degenerate solution
                      BA_L2_lppl_tc

                      Network Value: Metcalfe's Law

                      Slide 34 in the PDF

                      Metcalfe's law states that the value of a network grows with the square of the number of its users \(n\). The reasoning is that \(n\) users can form about \(n^2/2\) connections and that each connection adds value. The law has been used to value telecom and social networks, and Peterson (2018) applies it to Bitcoin.

                      As a regression on logarithms, the law reads

                      \[\ln M_t = a + b \ln n_t + u_t\]
                      (2.8)

                      where \(M_t\) is the market value, \(n_t\) a measure of use, here Bitcoin transactions per day, and \(u_t\) the error. Metcalfe's law predicts \(b = 2\). The coefficient \(b\) is an elasticity: 1% more use goes with \(b\)% more market value.

                      The data are monthly means from November 2014 to October 2025, 132 months, excluding the 11 erroneous days of 2023 described in Lecture 1. Two limits apply before any chart is drawn. Transactions are not users: one firm can send many transactions, and users of exchanges trade off the chain. And two series that both trend upward can give a high \(R^2\) without a stable link between them, a spurious regression.

                      In brief

                      • Metcalfe's law: the value of a network grows with the square of its users \(n\)
                        • \(n\) users can form about \(n^2/2\) connections; each connection adds value
                        • used to value telecom and social networks; applied to Bitcoin by Peterson (2018)
                      • As a regression on logs:
                        \[ \ln M_t = a + b \ln n_t + u_t, \qquad \text{Metcalfe: } b = 2 \]
                        • \(M_t\): market value; \(n_t\): a measure of use, here Bitcoin transactions per day; \(u_t\): the error
                        • \(b\) is an elasticity: 1% more use goes with \(b\)% more value
                      • Data: monthly means, November 2014 – October 2025, 132 months; the 11 erroneous days of 2023 (Lecture 1) excluded
                      • Limits, before any chart
                        • transactions are not users: one firm can send many transactions; exchange users trade off-chain
                        • two series with trends can give a high \(R^2\) without a stable link (spurious regression)

                      Market Value against Metcalfe Value

                      Slide 35 in the PDF

                      The left panel compares Bitcoin's monthly market value with a Metcalfe value, \(\exp(\hat c + 2\ln n_t)\): the slope is fixed at \(b = 2\), and the constant \(\hat c\) is set so that the ratio of market value to Metcalfe value is 1 on average in logs. Both series are in billions of USD on a log scale.

                      The right panel plots the ratio of the two, also on a log scale, with reference lines at 0.5, 1 and 2. The ratio was below 0.5 in 2015–2016 and above 4 in 2021 and in 2025, with a maximum of 9.1 in June 2025.

                        Question for the room The ▶ button moves the exponent b of the network value from 1 to 4.5. For which b is the spread of the log ratio, market value over network value, the smallest?
                        Answer

                        Near b = 3.6: the standard deviation of the log ratio is 1.45 at b = 2 and 1.29 for b between 3.25 and 3.75. With the constant set so that the ratio is 1 on average in logs, minimising this spread is least squares, so the minimum is at the least-squares slope b = 3.59 of the lecture. The choice of b also changes the reading: the last ratio, October 2025, is 6.5 with b = 2 and 3.2 with b = 3.6. The ratio is a diagnostic of the model, not a fair value.

                        The ratio drifts upwards over the sample: in this period, market value grew faster than transactions, as Lecture 1 also found with the trend regressions. The ratio is a diagnostic that depends on the model, with \(b = 2\), and on the proxy for users, transactions; it is not a fair value, and a ratio above 1 does not by itself show that Bitcoin is overpriced.

                        In brief

                        • Metcalfe value: \(\exp(\hat c + 2\ln n_t)\), with \(b\) fixed at 2 and \(\hat c\) set so that the ratio is 1 on average in logs
                        • The ratio was below 0.5 in 2015–2016 and above 4 in 2021 and 2025 (maximum 9.1, June 2025)
                        • The ratio drifts upwards: in this sample value grew faster than transactions
                        • The ratio is a diagnostic that depends on the model (\(b = 2\)) and on the proxy (transactions), not a fair value
                        BA_L2_metcalfe

                        Metcalfe's Law: the Estimates

                        Slide 36 in the PDF

                        The table estimates formula (2.8). Standard errors use the Newey–West method, which allows the errors to be correlated over time, with 12 lags in levels and 3 lags in changes. On the full sample, 132 months, the elasticity is \(\hat b = 3.59\) with a standard error of 0.43 and \(R^2 = 0.58\). The test of Metcalfe's \(b = 2\) gives \(t = (3.59 - 2)/0.43 = 3.66\), so \(b = 2\) is rejected at 5%; the estimated elasticity is larger.

                        The estimate is not stable. It is 2.71 (standard error 0.35) for 2014–2019 and 1.17 (0.48) for 2020–2025, with \(R^2 = 0.20\) in the second period. The levels regression may also be spurious: the Engle–Granger test, which asks whether the two log series share a common long-run path (cointegration), does not reject the null of no cointegration (p-value 0.57).

                        Two checks remove the common trend. With a linear time trend added to the regression, \(\hat b\) falls to 0.27 (standard error 0.47); in monthly changes, \(\Delta\ln M_t\) on \(\Delta\ln n_t\), it is 0.31 (0.15), with \(R^2 = 0.05\). Once the trend is removed, the link between transactions and value is weak. Since 2020, exchange-traded funds, institutional flows and macroeconomic news are plausible drivers of the price, but the regression does not identify them.

                        In brief

                        Samplemonths\(\hat b\)standard error\(R^2\)
                        2014–2025132\(3.59\)\(0.43\)\(0.58\)
                        2014–201962\(2.71\)\(0.35\)\(0.52\)
                        2020–202570\(1.17\)\(0.48\)\(0.20\)
                        2014–2025, with a linear time trend132\(0.27\)\(0.47\)\(0.89\)
                        2014–2025, monthly changes \(\Delta\ln M_t\) on \(\Delta\ln n_t\)131\(0.31\)\(0.15\)\(0.05\)
                        • Newey–West standard errors (12 lags in levels, 3 in changes)
                        • Levels, full sample: test of \(b = 2\), \(t = (3.59 - 2)/0.43 = 3.66\); rejected at 5%
                        • But the levels regression may be spurious: Engle–Granger test, no cointegration (p-value 0.57)
                          • with a time trend or in monthly changes, \(\hat b\) falls to about 0.3
                        • Since 2020 the link is weak (\(R^2 = 0.20\)); ETFs, institutional flows and macro news are plausible drivers, not identified here

                        Metcalfe's Law in Levels and in Changes

                        Slide 37 in the PDF

                        The chart draws the two regressions of the table. The left panel plots the log market value against the log number of transactions per day, one point per month, 132 months, with the fitted line of slope 3.59 in red and Metcalfe's slope of 2, drawn through the means, as a dashed green line. The right panel plots the monthly changes of the two logarithms, 131 months, with the fitted slope of 0.31.

                        The two clouds look very different. In levels, both series trend upward over the sample, so the points form a long, tilted cloud and the fitted slope is steep. In changes, the trends are removed and the cloud is almost round: the slope is 0.31 with a standard error of 0.15, \(R^2 = 0.05\), and the correlation of the monthly changes is 0.22. A month with more transactions than the month before is only weakly associated with a rise in market value.

                        A relation that holds in levels but almost disappears in changes is the classic warning sign of a spurious regression: two trending series look related because both grow, not because one drives the other. Together with the failed cointegration test, the chart explains why the levels estimate of 3.59 cannot be read as an elasticity of value with respect to use.

                        In brief

                        Metcalfe's Law in Levels and in Changes
                        • Left: log levels, 132 months; fitted slope 3.59, against Metcalfe's slope of 2 (dashed green)
                        • Right: monthly log changes, 131 months; slope 0.31 (SE 0.15), \(R^2 = 0.05\), correlation 0.22
                        • Both levels trend upwards, so their cloud is long and tilted; changes remove the trends, and the link almost disappears
                        • A relation that holds in levels but not in changes is the warning sign of a spurious regression
                        BA_L2_metcalfe_changes

                        Section summary

                        Bitcoin's drawdowns are deep and long: \(-83.4\%\) at the worst, with 76% of days more than 10% below the record. An LPPL model fits the 2017 bubble closely after the fact, but its critical time moves with the estimation window. Metcalfe's law gives an elasticity of 3.59 in levels, but the relation is not cointegrated and falls to about 0.3 once the trend is removed.

                        Summary

                        Volatility must be annualised with the market's own calendar: formula (2.2) with \(A = 365\) gives 64.1% a year for Bitcoin, against 19.2% for the S&P 500. Daily returns are not Normal: the Jarque–Bera statistic (2.3) exceeds 10,000 for all four assets, and crash days are far more frequent than the Normal distribution allows. Volatility clusters: for Bitcoin the Ljung–Box statistic (2.4) with 10 lags is 212.7 for squared returns, against 20.2 for returns. Since 2020 Bitcoin co-moves with equities, like a risk asset, with an average correlation of 0.39 with the S&P 500.

                        Robust variance-ratio tests do not reject a random walk at horizons of 2 to 10 days, in any sub-period; this is not proof of efficiency. The LPPL model fits the 2017 bubble with \(R^2 \approx 0.98\), but \(\hat t_c\) moves with the window, so a good in-sample fit does not date the crash. Metcalfe's law in levels gives \(\hat b = 3.59\), with a standard error (SE) of 0.43, but there is no cointegration, and in monthly changes \(\hat b = 0.31\). Lecture 4 models the two main facts: volatility clustering, with GARCH, and heavy tails, with VaR 1% and ES 2.5%.

                        In brief

                        • Annualise with the market's own calendar: \(\sigma_a = s\sqrt{365}\) gives 64.1% for Bitcoin, against 19.2% for the S&P 500
                        • Normality is rejected (Jarque–Bera \(> 10\,000\) for all four assets); crash days are far more frequent than the Normal distribution allows
                        • Volatility clusters: Ljung–Box \(Q(10) = 212.7\) for \(r_t^2\), against 20.2 for \(r_t\)
                        • Since 2020 Bitcoin co-moves with equities, like a risk asset: correlation 0.39 with the S&P 500
                        • Robust variance-ratio tests do not reject a random walk at 2–10 days, in any sub-period; this is not proof of efficiency
                        • LPPL fits the 2017 bubble with \(R^2 \approx 0.98\), but \(\hat t_c\) moves with the window: a good in-sample fit does not date the crash
                        • Metcalfe in levels: \(\hat b = 3.59\), standard error (SE) 0.43; no cointegration, and in monthly changes \(\hat b = 0.31\)
                        • Lecture 4 models these facts: volatility clustering (GARCH) and heavy tails (VaR 1%, ES 2.5%)

                        Interpretation Exercises

                        Slide 39 in the PDF

                        The five exercises use the numbers of the lecture. Each asks for an interpretation in a few sentences: what the number or test shows, what it does not show, and what follows for a decision about holding or valuing crypto assets.

                        In brief

                        1. A treasurer reports Bitcoin's annual volatility as 53.3%. What did she do wrong, and what is the right figure?
                        2. Jarque–Bera \(= 29\,620\) for Bitcoin. State \(H_0\), the conclusion, and the consequence for a VaR computed with the Normal distribution.
                        3. Bitcoin, 2014–2026: Ljung–Box \(Q(10) = 20.2\) (p 0.028) for \(r_t\), \(212.7\) for \(r_t^2\); robust \(Z^*(5) = -0.19\) (p 0.853). Is Bitcoin predictable? In what sense? What do the tests not cover?
                        4. LPPL fit on 1 January – 1 December 2017: \(R^2 = 0.98\), \(\hat t_c\) = 25 December 2017. Would you have sold on 1 December? What does a high \(R^2\) not tell you?
                        5. Metcalfe: \(\hat b = 3.59\) (SE 0.43) in levels on 2014–2025, no cointegration (p 0.57), \(\hat b = 0.31\) (SE 0.15) in monthly changes. What does this imply for valuations based on user growth?

                        Appendix: Further Material

                        The Appendix holds optional material for self-study. It measures the heaviness of the tails with the tail index, checks the stability of kurtosis, looks at weekdays, crisis episodes and safe havens, gives the details of the LPPL estimation and of the rolling estimates, and works through the variance-ratio statistic and a list of Bitcoin's large drawdowns.

                        Appendix: Further Material

                        Slide 40 in the PDF
                        Optional material for self-study.
                        • The Tail Index and the Hill Estimator
                        • Tail Indices of Crypto, Equities and Gold
                        • Tail Plot and Hill Plot
                        • Stability of Kurtosis over Time
                        • Weekday and Weekend Volatility of Bitcoin
                        • Three Crisis Episodes
                        • Bitcoin as a Hedge, Safe Haven or Risk Asset
                        • Average Returns on the Worst Equity Days
                        • Estimating the LPPL
                        • LPPL Estimates: Stability of \(\hat t_c\)
                        • A Rolling Metcalfe Elasticity
                        • Measuring Volatility Through Time
                        • Rolling Correlation
                        • The Robust Variance-Ratio Statistic
                        • Worked Example: Variance Ratio from Autocorrelations
                        • The Variance-Ratio Profile
                        • Bitcoin's Large Drawdowns
                        • Worked Example: Accelerating Oscillations

                        The Tail Index and the Hill Estimator

                        Slide 41 in the PDF

                        Far in the tail, the probability that the daily loss \(L = -r\) exceeds a level \(x\) often follows a power law, \(P(L > x) \approx c\,x^{-\alpha}\) for large \(x\), where \(c > 0\) is a constant and \(\alpha\) the tail index. A smaller \(\alpha\) means a heavier tail, and moments of order \(\alpha\) or higher are infinite: with \(\alpha < 4\) the kurtosis does not exist. The estimator of Hill (1975) uses the \(k\) largest losses: \(\hat\alpha\) is the inverse of the average log ratio of each of these losses to the threshold \(L_{(k+1)}\), the next largest loss, and a 95% confidence interval (CI) is \(\hat\alpha(1 \pm 1.96/\sqrt k)\). Here \(k\) is 2.5% of the sample. Large ratios on average mean that the extreme losses lie far above the threshold, and give a small \(\hat\alpha\), a heavy tail.

                        In brief

                        • Far in the tail, losses follow a power law:
                          \[ P(L > x) \approx c\,x^{-\alpha} \quad \text{for large } x \]
                          • \(L = -r\): the daily loss; \(c > 0\): a constant; \(\alpha\): the tail index
                          • smaller \(\alpha\) = heavier tail; moments of order \(q \ge \alpha\) are infinite: \(\alpha < 4\) means infinite kurtosis
                        • Hill (1975) estimator from the \(k\) largest losses \(L_{(1)} \ge L_{(2)} \ge \dots \ge L_{(k+1)}\), with a confidence interval (CI):
                          \[ \hat\alpha = \Bigg[\frac1k\sum_{i=1}^{k} \ln\frac{L_{(i)}}{L_{(k+1)}}\Bigg]^{-1}, \qquad \mathrm{CI}_{95\%} = \hat\alpha\Big(1 \pm \frac{1.96}{\sqrt k}\Big) \]
                          • \(L_{(k+1)}\): the threshold of the tail; here \(k = 2.5\%\) of \(n\)
                          • large ratios \(L_{(i)}/L_{(k+1)}\) on average give a small \(\hat\alpha\): a heavy tail

                        Tail Indices of Crypto, Equities and Gold

                        Slide 42 in the PDF

                        The table reports the Hill estimates of the left tail (losses) and the right tail (gains) with 95% intervals, for daily returns since 2014 (Bitcoin, S&P 500, gold) and 2016 (Ethereum) to 18 September 2026. All left-tail indices are below 4, from 2.62 for Ethereum to 3.11 for gold, so the sample kurtosis of the four-asset table estimates a quantity that may not exist, as the next slides show. Crypto is more volatile than equities, but the shape of its tails is close: the intervals of Bitcoin, Ethereum and the S&P 500 overlap. For Bitcoin, losses have a heavier tail than gains, 2.71 against 3.37.

                        In brief

                        left tail \(\hat\alpha\)95% CIright tail \(\hat\alpha\)95% CI
                        Bitcoin\(2.71\)\([2.20, 3.21]\)\(3.37\)\([2.74, 4.00]\)
                        Ethereum\(2.62\)\([2.10, 3.15]\)\(3.09\)\([2.48, 3.71]\)
                        S&P 500\(2.68\)\([2.08, 3.29]\)\(2.80\)\([2.17, 3.43]\)
                        Gold\(3.11\)\([2.42, 3.80]\)\(3.04\)\([2.36, 3.71]\)
                        • Daily returns since 2014 (Bitcoin, S&P 500, gold) and 2016 (Ethereum), to 18 September 2026
                        • Interpretation
                          • all left-tail indices are below 4: the sample kurtosis of the four-asset table is not a stable number (next slides)
                          • crypto is more volatile, but the shape of its tails is close to that of equities; the intervals overlap
                          • for Bitcoin, losses have a heavier tail than gains (2.71 against 3.37)
                        SFM_ch14_asset_class

                        Tail Plot and Hill Plot

                        Slide 43 in the PDF

                        The left panel plots the empirical probability that the standardised loss exceeds \(x\), on log-log scales, for Bitcoin and the S&P 500. A power-law tail is a straight line with slope \(-\alpha\); the Normal distribution, in red, bends down steeply. The right panel is a Hill plot: the estimate \(\hat\alpha\) against \(k\), the share of the largest losses used, with 95% bands, for 2018–2026. For Bitcoin \(\hat\alpha\) is 4.33 at \(k = 1\%\), 2.71 at 2.5%, 2.60 at 5% and 1.94 at 10%; for the S&P 500 it is 2.92, 2.81, 2.32 and 1.95. With a small \(k\) few observations enter and the bands are wide; with a large \(k\) the centre of the distribution enters and biases \(\hat\alpha\) downwards. The estimate is reported on the flat part of the plot, here \(k \approx 2.5\%\).

                        In brief

                        Tail Plot and Hill Plot
                        • Left: \(P(\text{loss} > x)\) on log-log scales; a power-law tail is a straight line with slope \(-\alpha\), the Normal (red) bends down
                        • Right: Hill estimate against \(k\) with 95% bands, 2018–2026. Bitcoin: 4.33 (\(k = 1\%\)), 2.71 (2.5%), 2.60 (5%), 1.94 (10%); S&P 500: 2.92, 2.81, 2.32, 1.95
                        • Small \(k\): few observations, wide bands; large \(k\): the centre of the distribution biases \(\hat\alpha\) down. Report \(\hat\alpha\) on the flat part, here \(k \approx 2.5\%\)
                        BA_L2_tails

                        Stability of Kurtosis over Time

                        Slide 44 in the PDF

                        The chart computes Bitcoin's skewness (left) and excess kurtosis (right) on rolling windows of 365 days. The kurtosis ranges from 1.3 to 58.0; the maximum belongs to the window ending in October 2020, which contains 12 March 2020, and the latest value is 6.4. When that single day enters and later leaves the window, the estimate jumps by a factor of ten, and the skewness jumps with it. With a tail index below 4 the sample kurtosis has no stable limit, so a single kurtosis number for crypto returns should not be read as a fixed property of the asset.

                        In brief

                        Stability of Kurtosis over Time
                        • Skewness and excess kurtosis of Bitcoin on rolling 365-day windows
                        • Kurtosis ranges from 1.3 to 58.0 (window ending October 2020, which contains 12 March 2020); today 6.4
                        • One day moves the estimate by a factor of ten: with \(\alpha < 4\) the sample kurtosis has no stable limit (previous slides)
                        BA_L2_rolling_moments

                        Weekday and Weekend Volatility of Bitcoin

                        Slide 45 in the PDF

                        The chart shows the standard deviation of Bitcoin's daily returns by the weekday of the closing day, for 2014–2019, 2020–2023 and 2024–2026. There is no weekend effect in the mean return, but the variance at weekends is only 35–66% of the weekday variance, and the gap is largest on Saturdays in the two recent periods. Crypto trades on Saturdays, but banks, macroeconomic news and institutional traders do not, so risk still follows the traditional business calendar.

                        In brief

                        Weekday and Weekend Volatility of Bitcoin
                        • Standard deviation of Bitcoin's daily returns by closing day, three periods
                        • No weekend effect in the mean; but weekend variance is only 35–66% of weekday variance
                        • Crypto trades on Saturdays, but banks, macro news and institutional traders do not: risk follows the traditional calendar
                        SFM_ch14_asset_class

                        Three Crisis Episodes

                        Slide 46 in the PDF

                        The table gives the cumulative log returns, in log-return percentage points, of the four assets over three crisis episodes; converted to simple returns with \(e^{r/100} - 1\), the COVID-19 losses were \(-33.4\%\) for Bitcoin, \(-48.1\%\) for Ethereum and \(-33.9\%\) for the S&P 500. The global COVID-19 shock hit crypto as hard as equities, while gold did not move. The two crypto-specific shocks, the collapse of TerraUSD in May 2022 and the bankruptcy of FTX in November 2022, cost Bitcoin and Ethereum at least 20% as simple returns and equities almost nothing. In these three episodes the link is asymmetric: the equity crash reached crypto, but the crypto crashes did not reach equities.

                        In brief

                        EpisodeperiodBitcoinEthereumS&P 500gold
                        COVID-1919 Feb – 23 Mar 2020\(-40.6\)\(-65.5\)\(-41.4\)\(0.3\)
                        Terra/Luna5 – 12 May 2022\(-23.0\)\(-33.8\)\(-5.4\)\(-3.0\)
                        FTX5 – 21 Nov 2022\(-29.9\)\(-38.4\)\(4.6\)\(3.4\)
                        • Cumulative log returns over each episode, in log-return percentage points
                          • as simple returns \(e^{r/100} - 1\): COVID-19 Bitcoin \(-33.4\%\), Ethereum \(-48.1\%\), S&P 500 \(-33.9\%\)
                        • Interpretation
                          • a global shock (COVID-19) hits crypto as hard as equities; gold does not move
                          • crypto-specific shocks (the collapse of the stablecoin TerraUSD; the bankruptcy of the exchange FTX) cost Bitcoin and Ethereum at least 20% (simple returns), equities hardly anything
                          • in these three episodes the link is asymmetric: the equity crash reached crypto, the crypto crashes did not reach equities

                        Bitcoin as a Hedge, Safe Haven or Risk Asset

                        Slide 47 in the PDF

                        Baur & Lucey (2010) define a hedge as an asset uncorrelated or negatively correlated with equities on average, and a safe haven as one uncorrelated or negatively correlated with equities in extreme market stress. The slide uses a simpler check: the mean return of an asset on the worst 5% of S&P 500 days, with a \(t\)-statistic. Over 2018–2026 these are 110 common days with the S&P 500 below \(-1.77\%\), on average \(-2.96\%\). On these days Bitcoin lost \(3.36\%\) on average (95% CI \(-4.72\) to \(-1.99\)) and Ethereum \(4.99\%\) (\(-6.68\) to \(-3.31\)), more than the equities themselves, while gold's mean was \(+0.01\%\) (\(-0.27\) to \(0.29\)). For gold a zero mean is not rejected; this does not show that its mean is non-negative, nor that gold is a safe haven in the sense of Baur & Lucey (2010).

                        In brief

                        • Baur & Lucey (2010): a hedge is uncorrelated or negatively correlated with equities on average; a safe haven is uncorrelated or negatively correlated with equities in extreme market stress
                        • A simpler check here: the mean return of asset \(x\) on the set \(\mathcal{S}\) of the worst 5% of S&P 500 days
                          \[ \bar x_{\mathcal{S}} = \frac{1}{n_{\mathcal{S}}}\sum_{t \in \mathcal{S}} x_t, \qquad t = \frac{\bar x_{\mathcal{S}}}{s_{\mathcal{S}}/\sqrt{n_{\mathcal{S}}}} \]
                          • 2018–2026, common days: \(n_{\mathcal{S}} = 110\) days with the S&P 500 below \(-1.77\%\), on average \(-2.96\%\)
                        • Means with 95% CI: Bitcoin \(-3.36\%\) \([-4.72, -1.99]\), Ethereum \(-4.99\%\) \([-6.68, -3.31]\), gold \(+0.01\%\) \([-0.27, 0.29]\)
                        • Reading
                          • on these days crypto falls on average more than equities
                          • gold: a zero mean is not rejected; this does not show that its mean is non-negative, nor that gold is a safe haven in the sense of Baur & Lucey (2010)

                        Average Returns on the Worst Equity Days

                        Slide 48 in the PDF

                        The chart shows the mean daily returns of the previous slide as bars, with their 95% confidence intervals, on the 110 worst S&P 500 days of 2018–2026. The S&P 500 bar is the average of the selected days themselves; Bitcoin and Ethereum fall further, with wider intervals because their returns vary more, and gold stays at zero. For a corporate treasury, the reading is that in this sample Bitcoin lost on average on the days when equities were under stress.

                        In brief

                        Average Returns on the Worst Equity Days
                        • Mean daily return on the 110 worst S&P 500 days of 2018–2026, with 95% confidence intervals
                        • For a corporate treasury: in this sample, Bitcoin lost on average on the days when equities were under stress
                        BA_L2_safe_haven

                        Estimating the LPPL

                        Slide 49 in the PDF

                        The LPPL formula (2.7) can be rewritten with \(C_1 = C\cos\phi\) and \(C_2 = C\sin\phi\), so that the log price is a linear function of \(A\), \(B\), \(C_1\) and \(C_2\) once \(t_c\), \(m\) and \(\omega\) are fixed. The estimation therefore has two steps: for given \((t_c, m, \omega)\), ordinary least squares gives the four linear parameters; a numerical search over \((t_c, m, \omega)\) then minimises the sum of squared errors. The objective has many local minima, so the search starts from many points and the model is fitted on many windows. The critical time is the least precise parameter: fits look convincing after a crash, and real-time warnings give false alarms.

                        In brief

                        • Rewrite the cosine term with \(C_1 = C\cos\phi\) and \(C_2 = C\sin\phi\):
                          \[ \ln P_t = A + B f_t + C_1 g_t + C_2 h_t + \varepsilon_t \]
                          \[ f_t = (t_c - t)^m, \quad g_t = (t_c - t)^m\cos(\omega\ln(t_c - t)), \quad h_t = (t_c - t)^m\sin(\omega\ln(t_c - t)) \]
                        • Two steps
                          1. for given \((t_c, m, \omega)\), the model is linear in \((A, B, C_1, C_2)\): ordinary least squares
                          2. search over \((t_c, m, \omega)\) to minimise the sum of squared errors \(\sum_t \varepsilon_t^2\)
                        • Caution
                          • many local minima: start the search from many points; fit on many windows
                          • \(t_c\) is the least precise parameter; fits look convincing after the crash, real-time warnings give false alarms

                        LPPL Estimates: Stability of \(\hat t_c\)

                        Slide 50 in the PDF

                        The table reports three fits that all start on 1 January 2017 and end on 15 November, 1 December and 8 December 2017. In all three \(\hat B \approx -2.8\), indicating super-exponential growth, \(|\hat C|/|\hat B|\) is below 0.09 and \(R^2 \approx 0.98\), so a good fit does not mean a reliable forecast. The estimated \(\hat t_c\) moves from 4 January 2018 to 25 December 2017 to 8 December 2017, by four weeks when the window moves by three weeks; the lead, the number of days from the window end to \(\hat t_c\), falls from 50 to 24 to 0. On 15 November \(\hat m\) and \(\hat\omega\) sit on the bounds of the search, 0.9 and 13, a sign of a poorly identified fit, and on 8 December \(\hat t_c\) equals the last observation, a degenerate solution. In practice one fits many windows and reports the share that pass the filters, the log-periodic power law singularity (LPPLS) confidence indicator.

                        In brief

                        window ends\(n\)\(\hat t_c\)lead (days)\(\hat m\)\(\hat\omega\)\(|\hat C|/|\hat B|\)\(R^2\)
                        15 Nov 20173194 Jan 201850\(0.900\)\(13.00\)\(0.089\)\(0.980\)
                        1 Dec 201733525 Dec 201724\(0.851\)\(12.49\)\(0.088\)\(0.982\)
                        8 Dec 20173428 Dec 20170\(0.741\)\(11.14\)\(0.081\)\(0.979\)
                        • All windows start on 1 January 2017; \(\hat B \approx -2.8\) in all three: super-exponential growth
                        • Interpretation
                          • \(R^2 \approx 0.98\) in every window: a good fit does not mean a reliable forecast
                          • \(\hat t_c\) moves by four weeks when the window moves by three weeks
                          • 15 November: \(\hat m\) and \(\hat\omega\) sit on the bounds of the search (0.9 and 13), a sign of a poorly identified fit
                          • 8 December: \(\hat t_c\) equals the last observation, a degenerate solution
                        • Practice: fit many windows and report the share of windows that pass the filters (the log-periodic power law singularity (LPPLS) confidence indicator)

                        A Rolling Metcalfe Elasticity

                        Slide 51 in the PDF

                        The chart estimates the Metcalfe elasticity \(\hat b\) of formula (2.8) on rolling 36-month windows, with Newey–West 95% intervals; the red dashed line is Metcalfe's \(b = 2\). The estimate peaks at 2.70 for the window ending in March 2018, turns negative in 2021–2023, when market value fell while transactions grew, and ends at 0.98 in October 2025. The interval contains 2 in only 28% of the windows, so the data show no stable network-value elasticity.

                        In brief

                        A Rolling Metcalfe Elasticity
                        • \(\hat b\) on rolling 36-month windows with Newey–West 95% intervals; red: Metcalfe's \(b = 2\)
                        • \(\hat b\) peaks at 2.70 (window ending March 2018), turns negative in 2021–2023 (value fell while transactions grew), ends at 0.98 (October 2025)
                        • The interval covers 2 in only 28% of the windows: there is no stable network-value elasticity
                        BA_L2_metcalfe_rolling

                        Measuring Volatility Through Time

                        Slide 52 in the PDF

                        The rolling volatility on day \(t\) is the standard deviation of the last \(w\) returns, multiplied by \(\sqrt{A}\), where \(A\) is 365 for crypto and 252 for equities; the window moves forward one day at a time. The choice of \(w\) is a trade-off: a short window reacts fast but is noisy, a long window is smooth but late. The standard error of the estimate is roughly \(\hat\sigma/\sqrt{2w}\), about 13% of \(\hat\sigma\) for \(w = 30\), so a monthly volatility of 50% is known only to within several percentage points. Lecture 4 replaces the window by a model, the exponentially weighted moving average (EWMA) or GARCH.

                        In brief

                        • Rolling volatility on a window of the last \(w\) days, annualised:
                          \[ \hat\sigma_{a,t} = \sqrt{A}\,\sqrt{\frac{1}{w-1}\sum_{j=0}^{w-1}\big(r_{t-j} - \bar r_t\big)^2}, \qquad \bar r_t = \frac1w\sum_{j=0}^{w-1} r_{t-j} \]
                          • \(w\): the window length (30 days for crypto, 21 trading days for equities, about one month)
                          • \(A\): returns per year (365 or 252); the window moves forward one day at a time
                        • Trade-off in the choice of \(w\)
                          • short window: reacts fast, but noisy; long window: smooth, but late
                          • the standard error of \(\hat\sigma\) is roughly \(\hat\sigma/\sqrt{2w}\): about 13% of \(\hat\sigma\) for \(w = 30\)
                        • Lecture 4 replaces the window by a model: exponentially weighted moving average (EWMA) or GARCH

                        Rolling Correlation

                        Slide 53 in the PDF

                        The rolling correlation is the Pearson correlation of Bitcoin and S&P 500 returns on the last \(w = 250\) common days, about one year of equity trading. Its standard error is approximately \((1 - \rho^2)/\sqrt w\), about 0.06 here, so changes smaller than 0.1 are noise. The two markets have different calendars. The prices are joined on common days first and the returns computed afterwards, so that Monday's Bitcoin return covers Friday to Monday, like the equity return. The closes are also not simultaneous: the crypto close is usually taken at midnight in Coordinated Universal Time (UTC), the equity close at 16:00 New York time, and weekly returns give a check on this mismatch.

                        In brief

                        • Pearson correlation of two return series \(x\) (Bitcoin) and \(y\) (S&P 500) on the last \(w\) common days:
                          \[ \hat\rho_{xy,t} = \frac{\sum_{j=0}^{w-1}(x_{t-j} - \bar x_t)(y_{t-j} - \bar y_t)}{\sqrt{\sum_{j=0}^{w-1}(x_{t-j} - \bar x_t)^2}\,\sqrt{\sum_{j=0}^{w-1}(y_{t-j} - \bar y_t)^2}} \]
                          • \(w = 250\) common days, about one year of equity trading
                          • standard error \(\approx (1 - \rho^2)/\sqrt{w} \approx 0.06\): changes below 0.1 are noise
                        • Two calendars
                          • join prices on common days first, then compute returns: Monday's Bitcoin return then covers Friday to Monday, like the equity return
                          • closes are not simultaneous (midnight UTC, Coordinated Universal Time, vs 16:00 New York): weekly returns give a check

                        The Robust Variance-Ratio Statistic

                        Slide 54 in the PDF

                        Under volatility clustering, the variance-ratio test that assumes independent and identically distributed (i.i.d.) returns rejects too often. The robust statistic \(Z^*(q)\) of Lo & MacKinlay (1988) divides \(\widehat{\mathrm{VR}}(q) - 1\) by a standard error built from \(\hat\delta_j\), an estimate of the variance of the lag-\(j\) autocorrelation that allows the volatility to change over time; it is approximately standard Normal, and the random walk is rejected at 5% when \(|Z^*(q)| > 1.96\). The variance ratio is also linked to the autocorrelations: \(\mathrm{VR}(q) \approx 1 + 2\sum_{j=1}^{q-1}(1 - j/q)\rho(j)\), so positive autocorrelations push it above 1 and negative ones below 1.

                        In brief

                        • Volatility clustering makes the independent and identically distributed (i.i.d.) test too eager to reject; Lo & MacKinlay (1988) give a robust version:
                          \[ Z^*(q) = \frac{\widehat{\mathrm{VR}}(q) - 1}{\sqrt{\hat\theta^*(q)}} \;\approx\; N(0, 1) \]
                          \[ \hat\theta^*(q) = \sum_{j=1}^{q-1}\Big[\frac{2(q-j)}{q}\Big]^2\hat\delta_j, \qquad \hat\delta_j = \frac{\sum_{t=j+1}^{n}(r_t - \bar r)^2(r_{t-j} - \bar r)^2}{\Big[\sum_{t=1}^{n}(r_t - \bar r)^2\Big]^2} \]
                          • \(\hat\delta_j\): the variance of the lag-\(j\) autocorrelation when volatility changes over time
                          • reject the random walk at 5% if \(|Z^*(q)| > 1.96\)
                        • Link with autocorrelations: \(\mathrm{VR}(q) \approx 1 + 2\sum_{j=1}^{q-1}\big(1 - \frac{j}{q}\big)\rho(j)\)
                          • positive autocorrelations push VR above 1, negative ones below 1

                        Worked Example: Variance Ratio from Autocorrelations

                        Slide 55 in the PDF

                        For \(q = 2\) the link formula reduces to \(\mathrm{VR}(2) \approx 1 + \rho(1)\). For Bitcoin since 2014, \(\hat\rho(1) = -0.022\), so \(\mathrm{VR}(2) \approx 0.978\), close to the direct estimate of 0.977. For the S&P 500 over the same period, \(\hat\rho(1) = -0.126\) gives \(\mathrm{VR}(2) \approx 0.874\), against a direct estimate of 0.875. The S&P 500 shows mild daily mean reversion, typical of an index, where the bid–ask bounce and stale prices of some members create negative autocorrelation. The robust test still does not reject, because volatility clustering makes the variance of \(\widehat{\mathrm{VR}}\) large.

                        In brief

                        • For \(q = 2\): \(\mathrm{VR}(2) \approx 1 + 2\,(1 - \tfrac12)\,\rho(1) = 1 + \rho(1)\)
                        • Bitcoin, since 2014: \(\hat\rho(1) = -0.022\), so \(\mathrm{VR}(2) \approx 0.978\); the direct estimate is 0.977
                        • S&P 500, same period: \(\hat\rho(1) = -0.126\), so \(\mathrm{VR}(2) \approx 0.874\); the direct estimate is 0.875
                        • Interpretation
                          • the S&P 500 index shows mild daily mean reversion, typical of an index (bid-ask bounce, stale prices of its members)
                          • the robust test does not reject because the variance of \(\widehat{\mathrm{VR}}\) is large when volatility clusters

                        The Variance-Ratio Profile

                        Slide 56 in the PDF

                        The chart plots \(\widehat{\mathrm{VR}}(q)\) for horizons \(q = 2\) to 30 days over 2018–2026; the shaded area is the robust 95% band for Bitcoin under the random walk, which widens with the horizon. Bitcoin's ratio is 0.989 at 5 days (\(Z^* = -0.19\)) and 1.077 at 20 days (\(Z^* = 0.66\)), and Ethereum's is 1.11 at 20 days (\(Z^* = 0.95\)), so the crypto ratios rise slowly above 1 without leaving the band. The S&P 500 ratio is 0.796 at 20 days (\(Z^* = -0.76\)): the point estimate suggests mean reversion, but it is not significant.

                        In brief

                        The Variance-Ratio Profile
                        • \(\widehat{\mathrm{VR}}(q)\) for \(q = 2, \dots, 30\) days, 2018–2026; shaded: the robust 95% band for Bitcoin
                        • Bitcoin: VR(5) = 0.989 (\(Z^* = -0.19\)), VR(20) = 1.077 (\(Z^* = 0.66\)); Ethereum VR(20) = 1.11 (\(Z^* = 0.95\))
                        • S&P 500: VR(20) = 0.796 (\(Z^* = -0.76\)): mean reversion in the point estimate, but not significant
                        BA_L2_vr_profile

                        Bitcoin's Large Drawdowns

                        Slide 57 in the PDF

                        The table lists the episodes since September 2014 in which Bitcoin fell by more than 50% from a record, with the peak and trough dates and prices in USD, the depth, the date of the next record and the days under water. There are five such falls in twelve years, so a crash of this size is a regular event rather than an exception. The falls after the peaks of December 2017 and November 2021 took 1,080 and 847 days to recover, and the latest fall, from 124,753 USD on 6 October 2025 to 58,559 USD on 30 June 2026, \(-53.1\%\), had not been recovered by 18 September 2026.

                        In brief

                        PeakUSDtroughUSDdepth (%)new recorddays under water
                        2014-09-174572015-01-14178\(-61.1\)2015-12-15454
                        2017-12-1619 4972018-12-153 237\(-83.4\)2020-11-301 080
                        2021-04-1363 5032021-07-2029 807\(-53.1\)2021-10-19189
                        2021-11-0867 5672022-11-2115 787\(-76.6\)2024-03-04847
                        2025-10-06124 7532026-06-3058 559\(-53.1\)not yet–
                        • Episodes since September 2014 deeper than 50%
                        • Interpretation
                          • five falls of more than half in twelve years: a crash of this size is a regular event
                          • the 2017 and 2021 peaks took 1 080 and 847 days to recover
                        SFM_ch14_bubbles_etf

                        Worked Example: Accelerating Oscillations

                        Slide 58 in the PDF

                        The peaks of the cosine term in (2.7) satisfy \(\omega\ln(t_c - t_k) - \phi = 2\pi k\), so the distance from one peak to \(t_c\) is a constant multiple of the distance from the previous one: \((t_c - t_{k+1})/(t_c - t_k) = e^{-2\pi/\omega}\). For the Bitcoin 2017 fit, \(\hat\omega = 12.49\) gives \(e^{-2\pi/12.49} = 0.605\): a peak 30 days before \(t_c\) is followed by peaks 18.1, 11.0 and 6.6 days before \(t_c\). The inverse ratio, \(\lambda = e^{2\pi/\omega} = 1.65\), is the "preferred scaling ratio" of the hierarchy of traders in Johansen, Ledoit & Sornette (2000). Shrinking gaps between local tops are the visual signature of an LPPL bubble.

                        In brief

                        • Peaks of the cosine term satisfy \(\omega\ln(t_c - t_k) - \phi = 2\pi k\), so consecutive distances to \(t_c\) shrink by a constant factor
                          \[ \frac{t_c - t_{k+1}}{t_c - t_k} = e^{-2\pi/\omega} \]
                        • Bitcoin 2017 fit (slide “Bitcoin 2017: an LPPL Fit”): \(\hat\omega = 12.49\), so \(e^{-2\pi/12.49} = 0.605\)
                          • a peak 30 days before \(t_c\) is followed by peaks 18.1, 11.0 and 6.6 days before \(t_c\)
                        • The ratio \(\lambda = e^{2\pi/\omega} = 1.65\) is the “preferred scaling ratio” of the hierarchy of traders in Johansen, Ledoit & Sornette (2000)
                        • Practical reading: shrinking gaps between local tops are the visual signature of an LPPL bubble

                        References