1Name the stylised facts of financial returns and check each one on twenty years of real daily exchange rates and interest rates.
2Measure fat tails with excess kurtosis, the Jarque-Bera test and counts of extreme days (M-C08: Skewness and fat tails).
3Show that returns have almost no autocorrelation while their squares and absolute values have a lot: volatility clusters.
4Test the leverage effect honestly: strong in the textbook equity case, weak in FX, with a sign that depends on how the pair is quoted.
Why it matters
On 24 Jun 2016, the day after the UK's EU referendum, the European Central Bank's fixing put the pound at $1.3704, down from $1.4869 the day before: a fall of 8.16%. Measured in GBP/USD's own standard deviation over twenty years, that is a 13.6-standard-deviation day. A normal model of returns expects one like it once in 1.5 × 10⁴¹ trading days. Every risk model in this level starts from what real returns look like, so we start by looking.
Real dataTwenty years of daily exchange rates and interest rates, 2006 to 2025
Exchange ratesreal: the European Central Bank's euro reference rates, one fixing a day, set by a concertation between central banks that normally takes place around 14:10 CET, 5,119 TARGET business days from 2 Jan 2006 to 31 Dec 2025. GBP/USD and the other sterling and dollar pairs are ratios of the same day's euro rates (triangulated), not traded closes. A fixing is one snapshot a day, so moves that happen and reverse in between are missed: the sterling flash crash of 7 October 2016 in Asian hours, and sterling's record intraday low on 26 September 2022, show only in part.
Interest ratesreal: SONIA and Bank Rate from the Bank of England (5,053 UK business days), and the US 10-year Treasury rate (5,003 US business days). The Treasury series is the 10-year zero-coupon rate that the RiskAtoms Risk Engine bootstraps from the Treasury's daily par yield curve, not the published par yield. Returns are 100 × ln(Pt/Pt−1) in percent; rate changes are in basis points; each runs from one published value to the next, and nothing is filled in.
All series come via the RiskAtoms Risk Engine's stored history, downloaded 11 Nov 2026. Sources: European Central Bank; Bank of England; U.S. Department of the Treasury.
One series is not market data: the UK equity index in the trading desk's history from M3.2: Market risk, 750 days simulated with a seed for the bank's books illustrative. It is used once, to show what a leverage effect looks like, and says nothing about real equity markets.
Researchers who study price changes across markets keep finding the same handful of regularities, the stylised factsStylised fact: a statistical property found in the returns of many assets, markets and periods, stated loosely enough to hold across them all.. Rama Cont's survey (2001) lists eleven; the five that matter most for risk models are fat tails, little autocorrelation in returns, volatility clustering, the leverage effect and aggregational Gaussianity. L2.5.02: Return distributions measured skewness and kurtosis on an illustrative series; here every number is from real fixings and rates. L3.2.03: Equity risk promised models of equity volatility and fat tails: this lesson measures the tails, L6.1.04: GARCH models models the volatility, and equity in VaR models follows in M6.2: Market risk modelling.
1. Fat tails, in every series
Each row below is a daily series from 2006 to 2025. The standard deviation is the n − 1 sample one; annualised vol uses √252. Every series has excess kurtosisExcess kurtosis: the fourth central moment over the variance squared, minus 3. Zero for a normal; positive means fatter tails and a sharper peak. well above zero, and every Jarque-BeraJarque-Bera statistic: n/6 × (skewness² + excess kurtosis²/4). Under normality it is roughly chi-squared with 2 degrees of freedom. statistic is far above its 5% critical value of 5.99: not one of them is normal.
Daily returns (FX, %) and changes (rates, bp), 2006–2025. A normal expects 0.32 days beyond 4 sd in 5,118. Derived from the real data above.
Series
Days
Daily sd
Ann. vol
Skew
Excess kurt.
Jarque-Bera
Beyond 4 sd
GBP/USD
5,118
0.602%
9.55%
−0.69
10.55
24,149
22
EUR/USD
5,118
0.559%
8.87%
−0.08
4.31
3,973
16
EUR/GBP
5,118
0.490%
7.78%
0.47
6.62
9,540
27
GBP/JPY
5,118
0.797%
12.65%
−1.06
13.69
40,911
27
GBP/CHF
5,118
0.634%
10.06%
−2.40
63.31
859,596
26
US 10-year (bp)
5,002
5.79bp
−0.14
2.93
1,803
11
SONIA (bp)
5,052
5.66bp
−5.16
176
6.5 × 10⁶
70
GBP/USD has 22 days beyond 4 standard deviations, where a normal expects 0.32. GBP/CHF's excess kurtosis of 63.3 comes largely from two days: 15 Jan 2015, when the Swiss National Bank dropped its minimum rate of CHF 1.20 per euro (−14.50%), and 6 Sep 2011, when it introduced it (+7.77%). SONIA's 176 is a different animal: an overnight rate that follows Bank Rate, with 367 unchanged days, most of the rest within 1bp, and big steps on decision days (and, before 2009, money-market spikes, many at month- and quarter-ends). A rate pinned to the policy rate is not a free market price, and its “returns” are mostly near-zeros and jumps.
The ten largest GBP/USD days, with what drove them:
GBP/USD daily log returns at the ECB fixing, in percent; the z uses the full-sample sd. Sources for the events are in the Origins.
Day
Return
In sd
Normal: one day in
What happened
24 Jun 2016
−8.16%
−13.6
1.5 × 10⁴¹
The morning after the UK's EU referendum: sterling touched about $1.32 overnight, a level last seen in 1985
20 Jan 2009
−4.81%
−8.0
7.4 × 10¹⁴
Fears for UK banks after the government's latest rescue package; sterling hit $1.3863 during the day
22 Oct 2008
−4.05%
−6.7
5.7 × 10¹⁰
Global financial crisis
19 Mar 2009
3.89%
6.5
9.8 × 10⁹
The day after the Federal Reserve said it would buy up to $300 billion of Treasuries (18 Mar)
27 Jun 2016
−3.85%
−6.4
6.1 × 10⁹
The first Monday after the referendum
24 Oct 2008
−3.49%
−5.8
1.5 × 10⁸
Global financial crisis
7 Jan 2009
3.41%
5.7
7.2 × 10⁷
Global financial crisis
11 Nov 2022
3.22%
5.4
1.2 × 10⁷
US inflation came in below forecasts; the figures were released on 10 Nov, after that day's fixing
9 Mar 2009
−3.11%
−5.2
4.4 × 10⁶
Global financial crisis
11 Oct 2019
3.09%
5.1
3.5 × 10⁶
The UK and Irish prime ministers said they saw a pathway to a possible Brexit deal (10 Oct)
Six of the ten are falls in sterling, and the biggest by far is a fall, which is why GBP/USD's skewness is negative (−0.69). Six come from the global financial crisis of 2008–09 and three from the referendum and the Brexit talks. A normal fitted to the whole sample treats each as a once-in-many-lifetimes event; the data had ten in twenty years.
2. Returns barely remember yesterday
The autocorrelationAutocorrelation: the correlation of a series with its own value a given number of days earlier (the lag). of GBP/USD returns at lag 1 is 0.037, just outside the 95% band of ±0.0274 for white noise (M-D09: Time series and autocorrelation), and the Ljung-BoxLjung-Box test: adds up the squared autocorrelations at lags 1 to h into one statistic, compared with a chi-squared distribution with h degrees of freedom. statistic over 10 lags is 38.3 (p = 0.000034). Statistically real, economically nil: a lag-1 correlation of 0.037 could predict 0.14% of the next day's variance. L6.1.03: Time series: stationarity and autocorrelation fits it as an AR(1).
3. Volatility clusters
Square the returns, or take their absolute values, and the picture changes. The lag-1 autocorrelation of squared returns is 0.216 and of absolute returns 0.206; all 20 lags of the squares and all 20 of the absolute values sit above the band, and their Ljung-Box statistics are 640 and 1,359. The size of a move is predictable even when its direction is not: volatility clusteringVolatility clustering: large moves tend to be followed by large moves of either sign, and small by small, so turbulent and calm spells alternate..
Left bar at each lag: absolute returns. Right bar: returns. The shaded strip is the ±1.96/√n band.
Year by year the same series swings between calm and storm. GBP/USD's annualised vol was 16.7% in 2009 and 5.3% in 2014, against 9.5% over the whole sample. Tomorrow's risk looks more like today's than like the twenty-year average: that is what the GARCH models of L6.1.04: GARCH models capture.
4. Asymmetry, and a leverage effect that depends on the quote
In equities, volatility tends to rise after prices fall: the leverage effectLeverage effect: volatility tends to rise after prices fall. Named for one explanation in equities (a falling share price raises the firm's debt-to-equity ratio).. Two measures, for each series: the correlation of today's return with tomorrow's squared return, and the average volatility over the 20 days after a fall of more than 2 sd divided by the same after a rise of more than 2 sd.
Real FX returns 2006–2025; the desk index is 750 seeded days from the bank's books, not market data.
Series
corr(rₜ, r²ₜ₊₁)
Vol after falls ÷ after rises
Big falls / rises
GBP/USD
−0.055
1.10
132 / 123
EUR/USD
+0.059
0.96
145 / 133
EUR/GBP
+0.061
0.94
117 / 136
GBP/JPY
−0.048
1.10
134 / 114
GBP/CHF
−0.023
1.05
116 / 85
Desk's UK index illustrative
−0.115
1.21
19 / 14
The FX numbers are small, and their signs flip with the quote. GBP/USD gives −0.055: falls in the pound are followed by a little more volatility. EUR/GBP gives +0.061, and that says the same thing, because EUR/GBP rises when sterling falls. EUR/USD gives +0.059: a rising euro, which is a falling dollar, came with more volatility. An exchange rate has no natural “down”: whatever asymmetry there is belongs to whichever currency is the riskier one in the pair at the time. The desk's seeded index shows a larger effect (−0.115, ratio 1.21), but it is illustrative and only 750 days long; it shows what the measure looks like, not what equities do. L6.1.05: Asymmetric and multivariate volatility fits asymmetric GARCH models and finds the FX effect mild.
Read the asymmetry
5. Aggregational Gaussianity, slowly
Add daily returns into weekly, monthly or quarterly ones and the distribution should look more normal: the aggregational GaussianityAggregational Gaussianity: returns over longer horizons (weeks, months) look closer to normal than daily returns. that L2.5.02: Return distributions mentioned. On GBP/USD it happens, but slowly:
Non-overlapping sums of daily GBP/USD log returns, 2006–2025. Derived.
Horizon
Periods
sd (%)
Skew
Excess kurtosis
1 day
5,118
0.60
−0.69
10.55
5 days
1,023
1.36
−1.24
10.67
10 days
511
1.91
−1.48
8.83
21 days
243
2.66
−0.92
4.09
63 days
81
4.67
−0.46
2.64
Weekly returns are as fat-tailed as daily ones (10.7), because the referendum week holds two of the biggest days. Monthly kurtosis falls to 4.1 and quarterly to 2.6, but with only 81 quarters that last figure is itself very uncertain. The sd grows roughly with √(days), as L3.2.13: Scaling and aggregating risk assumed, though clustering bends that too.
Sort itWhich stylised fact is it?
Match each observation to the fact it shows.
GBP/USD had 22 days beyond 4 sd in twenty years.
The autocorrelation of squared returns is 0.22 at lag 1 and still positive at lag 20.
After a big fall, volatility over the next month is higher than after a big rise.
Quarterly returns have excess kurtosis 2.6, daily ones 10.6.
2009 was 3.1 times as volatile as 2014.
What these data can and cannot say
A fixing is one snapshot a day, so intraday extremes are missed and the tails here are, if anything, too thin. The crosses are triangulated, not traded. Every statistic above treats the sample as one population; with clustering, the tail of a calm year and of a crisis year are different things. That is why the next lessons fit distributions (L6.1.02: Fitting distributions by maximum likelihood) and then let the volatility move (L6.1.04: GARCH models).
The formal version
Returns, moments, autocorrelation, asymmetry
Fixings Pt on published days t=0,…,n; returns rt in percent (for a rate yt in percent, changes Δyt=100(yt−yt−1) in bp); sample mean rˉ; sample sd s with divisor n − 1:
rt=100ln(Pt/Pt−1),mk=n1∑t(rt−rˉ)k,S=m23/2m3,K=m22m4−3JB=6n(S2+41K2)≈χ22 under normality,#{t:∣rt−rˉ∣>4s} against 2nΦ(−4)ρ^k(x)=∑t=1n(xt−xˉ)2∑t=k+1n(xt−xˉ)(xt−k−xˉ),x∈{rt,rt2,∣rt∣},Q(h)=n(n+2)∑k=1hn−kρ^k2≈χh2leverage: corr(rt,rt+12),L=mean{σ^t(20):rt>2s}mean{σ^t(20):rt<−2s},σ^t(20)=201∑j=120rt+j2
Sources: skewness and kurtosis as in M-C08: Skewness and fat tails (the “biased” moment ratios that Jarque and Bera use); Jarque and Bera (1980, 1987) for JB; the ACF and Ljung and Box (1978) as in M-D09: Time series and autocorrelation; the facts themselves from Cont (2001) and McNeil, Frey and Embrechts (2015, ch. 3). The JB and Ljung-Box χ² laws assume independent, identically distributed data, which these returns are not (fact 3), so read them as measures of distance, not exact p-values.
Try it
Start with GBP/USD in its real order: the autocorrelation of squared returns starts at 0.216 and the stormiest 252-day stretch is 3.1 times as volatile as the calmest. Now shuffle the days. Watch the kurtosis and the standard deviation stay put while the clustering disappears (after the first shuffle the lag-1 autocorrelation of squares is −0.008). Switch to the US 10-year: rate changes cluster too.
Fat tails or clustering? Shuffle the days and see which survives
Shuffling keeps every one of the 5,118 daily moves and only changes their order.
Series
Order of the days
Autocorrelation of
Excess kurtosis10.55
Standard deviation (%)0.6016
ACF of squares, lag 10.216
Ljung-Box Q(10), squares640.4
p-value< 10⁻¹⁶
ACF of moves, lag 10.037
Stormiest / calmest 252 days3.10×
Band ±0.0274: bars inside it are what white noise would give 19 times in 20.
In the real order the squares remember: big days bunch together and some 252-day stretches are far stormier than others. Shuffle, and the memory vanishes while the excess kurtosis and the standard deviation stay exactly the same. Fat tails are a property of the distribution; clustering is a property of the order.
Worked example
The referendum day, and how far from normal GBP/USD is
From two fixings to a test statistic
Step 1 of 5
1
The return
The ECB fixing put GBP/USD at 1.48691 on 23 June 2016 and 1.37040 on 24 June (dollars per pound, from the euro rates of the day).
r=100×ln(1.486911.37040)=−8.160%
Regulatory lens
The Basel market risk standard answers fat tails and clustering in two lines. A bank using internal models “must use a 97.5th percentile, one-tailed confidence level” for expected shortfall (MAR33.3), an average over the tail rather than one quantile (L2.5.08: Expected shortfall), and “the ES measure must be calibrated to a period of stress” (MAR33.5), with the search for the most stressful 12 months going back to and including 2007 (MAR33.7). A model fitted to calm years alone would miss both the tail and the cluster. M4.3: Market risk capital (FRTB) and CVA covers these rules; M6.2: Market risk modelling builds the VaR and ES engines on the facts measured here. (Basel Framework MAR33, version effective 1 January 2023.)
Spark Atoms
✦Spark AtomCotton, 1963+
Benoit Mandelbrot's “The variation of certain speculative prices” (Journal of Business, October 1963) studied cotton prices back to 1816 and found tails “so extraordinarily long” that the sample variance behaved erratically. In the same paper he noted that “large changes tend to be followed by large changes, of either sign, and small changes tend to be followed by small changes”, a remark he credited to Hendrik Houthakker. Two of the five facts on this page, more than forty years before these data begin.
✦Spark AtomFischer Black's newsletter, 1976+
In the issue of 17 May 1976 of his newsletter Fischer Black on Options, Black wrote: “When the price goes down, the volatility goes up, and when the price goes up, the volatility goes down.” He added that he did not know which way the causation ran, and that common factors might drive both.
✦Spark Atom9% in minutes, 2.7% at the fixing+
In early Asian trading on 7 October 2016 sterling fell by around 9% against the dollar and quickly recovered much of it (BIS Markets Committee, January 2017). The ECB fixing that afternoon shows −2.68%. Daily data cannot see an event that happens and reverses between two fixings.
Key takeaways
Every daily series from 2006 to 2025 is fat-tailed: GBP/USD has excess kurtosis 10.6, 22 days beyond 4 sd (a normal expects 0.32) and a Jarque-Bera statistic of 24,149.
Returns barely remember yesterday (lag-1 autocorrelation 0.037), but their squares and absolute values do (0.216 and 0.206, all 20 lags above the band): volatility clusters.
Shuffling the days keeps the tails and destroys the clustering: tails belong to the distribution, clustering to the order.
The leverage effect is an equity story. In FX it is weak (GBP/USD −0.055) and its sign follows the quote; the desk's seeded index only illustrates the measure.
A normal distribution puts 0.0063% of days more than 4 standard deviations from the mean. How many such days should 5,118 normal daily returns contain? (Two decimals.)
days
Question 2 · objective 3
You shuffle twenty years of daily GBP/USD returns into a random order. Which of these is unchanged?
Question 3 · objective 4
GBP/USD returns have skewness −0.69; EUR/GBP returns have skewness +0.47. What is the best reading?
Question 4 · objective 2
1,000 daily returns have skewness −0.5 and excess kurtosis 4. What is the Jarque-Bera statistic? (One decimal.)
Question 5 · objective 3
GBP/USD's lag-1 autocorrelation of returns is 0.037, outside the 95% band of ±0.027. What follows?
Glossary
Stylised fact
Stylised fact: a statistical property found in the returns of many assets, markets and periods, stated loosely enough to hold across them all.
Fat tails
Fat tails: more very large moves, in either direction, than a normal distribution with the same standard deviation would give.
Excess kurtosis
Excess kurtosis: the fourth central moment over the variance squared, minus 3. Zero for a normal; positive means fatter tails and a sharper peak.
Jarque-Bera statistic
Jarque-Bera statistic: n/6 × (skewness² + excess kurtosis²/4). Under normality it is roughly chi-squared with 2 degrees of freedom.
Autocorrelation
Autocorrelation: the correlation of a series with its own value a given number of days earlier (the lag).
Ljung-Box test
Ljung-Box test: adds up the squared autocorrelations at lags 1 to h into one statistic, compared with a chi-squared distribution with h degrees of freedom.
Volatility clustering
Volatility clustering: large moves tend to be followed by large moves of either sign, and small by small, so turbulent and calm spells alternate.
Leverage effect
Leverage effect: volatility tends to rise after prices fall. Named for one explanation in equities (a falling share price raises the firm's debt-to-equity ratio).
Aggregational Gaussianity
Aggregational Gaussianity: returns over longer horizons (weeks, months) look closer to normal than daily returns.
Fixing
Fixing: an official rate set once a day at a stated time, such as the ECB's euro reference rates. It shows the market at that moment only.
A. J. McNeil, R. Frey and P. Embrechts, Quantitative Risk Management: Concepts, Techniques and Tools, revised edition, Princeton University Press, 2015, ch. 3 (Empirical Properties of Financial Data). Intermediate.
R. Cont, “Empirical properties of asset returns: stylized facts and statistical issues”, Quantitative Finance 1(2), 2001. Advanced.
R. Cont, “Empirical properties of asset returns: stylized facts and statistical issues”, Quantitative Finance 1(2), 223–236, 2001.
B. Mandelbrot, “The variation of certain speculative prices”, Journal of Business 36(4), 394–419, 1963.
E. F. Fama, “The behavior of stock-market prices”, Journal of Business 38(1), 34–105, 1965.
F. Black, “One way to estimate volatility”, Fischer Black on Options 1(8), 17 May 1976.
C. M. Jarque and A. K. Bera, “A test for normality of observations and regression residuals”, International Statistical Review 55(2), 163–172, 1987.
G. M. Ljung and G. E. P. Box, “On a measure of lack of fit in time series models”, Biometrika 65, 297–303, 1978.
A. J. McNeil, R. Frey and P. Embrechts, Quantitative Risk Management, revised edition, Princeton University Press, 2015, ch. 3.
Basel Committee on Banking Supervision, Basel Framework, MAR33 Internal models approach: capital requirements calculation, version effective 1 January 2023.
BIS Markets Committee, “The sterling ‘flash event’ of 7 October 2016”, Markets Committee Papers No 9, January 2017.
Swiss National Bank, press releases of 6 September 2011 (minimum exchange rate of CHF 1.20 per euro) and 15 January 2015 (minimum exchange rate discontinued).
Federal Reserve, FOMC statement, 18 March 2009; CNN Money, 24 June 2016; PA via Shropshire Star, 20 January 2009; RTÉ, 11 October 2019; CNBC, 10 November 2022.
European Central Bank euro reference rates; Bank of England SONIA and Bank Rate; U.S. Department of the Treasury par yield curve; all via the RiskAtoms Risk Engine, downloaded 11 November 2026.
Watch this lesson
Video version
Prefer to watch? The same lesson as a 14-minute video will sit here, made from this page's script.