Arbitrage and relative value
Statistical arbitrage
Statistical arbitrage bets that short-term deviations between related securities will correct, or that short-horizon signals predict which stocks will do better than others. Each bet is barely better than a coin toss; spread across thousands of stocks, with the market’s direction hedged out, the portfolio is what makes money.
At a glance
At a glance
- Makes money from
- Small, predictable deviations across many stocks, netted into a market-neutral book
- Holds positions for
- Hours to a few weeks; often about a day
- Trades
- Single stocks long and short, futures and ETFs to hedge
- Biggest risk
- Crowding – many funds holding similar positions unwind together – and costs eating the edge
- Who does it
- Quantitative funds such as Renaissance Technologies, D. E. Shaw, Two Sigma and Citadel
- 1.44 – Sharpe ratio of a published stat-arb backtest (Avellaneda–Lee, 1997–2007)
- 4–10× – typical gross leverage (longs plus shorts over capital)
- 20bp – cost of trading 1% of a day’s volume (at 2% daily volatility)
- 3% – a year lost to costs (20% daily turnover at 6bp)
- t > 3 – the bar for a new signal (Harvey, Liu and Zhu, 2016)
- 4 days – of the August 2007 quant quake (losses, then a rebound)
The business
The simplest version is a pair: two stocks that usually move together – two banks, two oil producers. When the gap between them stretches far beyond normal, sell the one that has risen and buy the one that has fallen, and close the trade when the gap comes back. Pairs trading began at Morgan Stanley in the mid-1980s; one of the group’s alumni, David Shaw, founded D. E. Shaw in 1988, the same year Renaissance Technologies started its Medallion fund.
Modern stat arb does the same thing with residuals: a stock’s return after removing what the market, its sector and common style factors explain. Thousands of residuals are forecast at once, from signals such as short-term reversal, order flow, news and analysts’ revisions, and combined into one portfolio whose net exposure to the market, sectors and factors is close to zero. What is left is a bet on thousands of small relative moves – barely better than chance one at a time, reliable in aggregate.
Trading a mean-reverting spread
The returns can be extraordinary at the top. Renaissance’s Medallion fund, open only to its employees, has been reported to have averaged about 66% a year before fees and 39% after from 1988 to 2018, charging fees of 5% of assets and 44% of profits. It is kept at around $10 billion, because the signals stop working if more money trades them – capacity, not ideas, is the limit.
A trade, step by step
Worked example
A pairs trade in two banks
Bank A’s daily returns have volatility 3%, bank B’s 2%, and they are 0.9 correlated. Today A falls 4% and B falls 1%.
- Hedge ratio1.35
- What A should have done−1.35%
- Residual−2.65%
- Its usual size1.31%
- How unusual?z ≈ −2.0
- The tradeBuy $1m of A, sell $1m of Bneutral to B
- If half the gap closes+$13,250
- Costs at 5bp a side−$2,350
About $10,900 if the gap halves as expected – and a loss if A fell for a reason, such as bad news of its own, that the model cannot see. One pair is a coin toss with slightly loaded odds; a fund runs thousands so that the odds, not the toss, decide the year.
How quickly does the gap close? If it shrinks by a tenth each day, it halves in trading days. The half-life sets the holding period, and so the turnover and the costs: a signal that reverts in a day must be traded every day.
How fast a gap closes
The maths that matters
Breadth. The fundamental law of active management puts it in one line: the information ratio is roughly the skill per bet times the square root of the number of independent bets, . A signal whose correlation with outcomes is only 0.02, applied to 2,500 independent bets a year, gives – a respectable risk-adjusted return from a nearly useless-looking forecast. The catch is “independent”: stocks share factors, and bets on them are less independent than they look.
Sharpe ratios scale with the square root of time. A strategy earning 0.05 standard deviations a day has an annual Sharpe ratio of . Its t-statistic over years is : a Sharpe ratio of 1 over five years gives .
Overfitting. Try enough ideas and some will look good by chance. The best of 100 strategies with no edge at all has a t-statistic of about 2.5 on average – past the usual 1.96. Harvey, Liu and Zhu (2016) argued that, given how many factors have been tried, a new one should clear ; a Bonferroni correction for 100 tries would ask for a p-value below 0.0005.
Why a good backtest is not enough
Costs. Trading costs grow faster than the size traded. A common rule of thumb is the square-root law: the cost of trading shares is about , for daily volatility and daily volume . Trading 1% of a day’s volume in a stock with 2% daily volatility costs about , or 20 basis points. A book that turns over 20% of its value a day at 6 basis points loses basis points a year to costs alone.
Where the money comes from
A year of a stat-arb book
- Universe: a few thousand liquid stocks, sometimes across many countries.
- Risk model: strip out market, sector and style exposures so the book is neutral to them.
- Signals: mean reversion, short-term momentum, order flow, news and earnings reactions, alternative data.
- Combination and optimisation: turn forecasts into target positions, net of a model of trading costs.
- Execution: trade patiently through algorithms, because costs are often the difference between a profitable signal and a losing one.
Avellaneda and Lee (2010) published a clean example: regress each stock on statistical factors or sector ETFs, model the residual as mean-reverting, buy when it is more than 1.25 standard deviations cheap and sell when it is that rich, and close as it comes back. Over 1997–2007 the factor version had a Sharpe ratio of 1.44 after assumed costs – but most of that came before 2003, and returns fell as more money chased the same idea. Every published edge decays this way.
Risk and what goes wrong
The defining risk is crowding. Funds that use similar data and similar methods hold similar positions, and when one is forced to sell, prices move against all of them. The book is hedged against the market, not against other quants.
The August 2007 quant quake, simulated
| Day | Return at 2:1 leverage |
|---|---|
| Tuesday 7 August | −1.16% |
| Wednesday 8 August | −2.83% |
| Thursday 9 August | −2.86% |
| Friday 10 August | +5.92% |
Pairs trading begins · Mid-1980s
A group at Morgan Stanley trades pairs of stocks systematically – the start of statistical arbitrage.
The quant quake · August 2007
Many quantitative equity funds lost heavily over a few days with no news, as one or more large funds sold similar positions to raise cash. The losses mostly reversed – too late for funds that had cut their positions.
The meme-stock squeeze · January 2021
Retail buying drove up heavily shorted stocks such as GameStop. Funds short the same names, quant and fundamental, were squeezed together; the fundamental fund Melvin Capital lost 53% in the month.
- Signal decay: an edge that is published or widely found shrinks as money trades it.
- Short squeezes: a stock sold short can rise without limit, and stock borrow can be recalled.
- Model risk: a stock moves for a reason the factors do not capture – a takeover, a fraud – and the “reversion” never comes.
- Operational risk: a bug that sends wrong orders fast across thousands of stocks.
Technology and data
- Clean data on prices, corporate actions and fundamentals, recorded as it was known at the time, so backtests do not see the future.
- Research platforms that test an idea across decades and thousands of stocks in minutes, with realistic costs.
- Risk models of factor exposures, and optimisers that turn forecasts into trades under constraints.
- Execution algorithms that split each order to minimise its cost, and measure what each trade actually cost.
- Alternative data – card spending, web traffic, satellite images, text – where an edge has not yet been traded away.
Who works here
Quant researcher
Finds and tests signals, and decides whether a backtest is real: checks for look-ahead bias, overfitting and costs. Most ideas fail; the job is to fail them quickly and honestly.
Portfolio manager
Runs a book of combined signals: sizes risk, watches crowding and drawdowns, and decides when a signal has stopped working.
Quant developer
Builds the data pipelines, research platform, optimiser and trading systems that every idea depends on.
Execution researcher
Models trading costs and builds the algorithms that keep them down – often worth as much as a new signal.
In interviews for this desk
Research interviews test statistics, regression and time series, probability, and Python data work, often with a case: is this signal real, and how would you test it without fooling yourself?
A strategy makes 0.05 standard deviations a day. What is its annual Sharpe ratio?
, assuming the days are independent. Autocorrelated returns would change the scaling.
A backtest shows a Sharpe ratio of 1 over five years. Is it significant? Does it matter that it was the best of 100 you tried?
The t-statistic is – significant on its own at 5%. But the best of 100 strategies with no edge scores about 2.5 on average, so after 100 tries it is no evidence at all. Correct for the number of trials, or test on data you have not touched.
Two stocks have daily volatilities of 3% and 2%, correlated 0.9. What is the hedge ratio, and how big is the residual?
Regressing A on B: . The residual’s volatility is a day.
A spread shrinks by 10% of its size each day. What is its half-life?
, so days. The approximation is close for small daily reversion.
Your book turns over 20% of its value a day and each trade costs 6 basis points. What do costs take a year?
basis points a day, times 250 days: 300 basis points, or 3% of capital – before leverage, which multiplies it.
A signal has an information coefficient of 0.05 across 500 stocks. What information ratio would you expect, and what is wrong with the estimate?
for one independent rebalance – but the 500 bets share factors, so the effective number is much smaller, and costs come off the top.
What happened in August 2007, and what would you do differently?
Crowded long–short positions were sold fast by at least one large fund, so prices moved against every fund holding the same trades, and they recovered after the selling stopped. Watch crowding, keep leverage low enough to survive a forced-selling episode, and be careful about cutting at the bottom.
Firm guides: D. E. Shaw · Two Sigma · Citadel
Glossary
- Pairs trade
- Buying one stock and selling a similar one, betting on the gap.
- Residual
- The part of a stock’s return not explained by the market, sector and factors.
- z-score
- How many standard deviations a spread or residual is from its average.
- Cointegration
- Two prices that wander individually but whose combination keeps returning to an average.
- Half-life
- How long a gap takes, on average, to shrink to half its size.
- Information coefficient
- The correlation between a forecast and what then happened.
- Fundamental law
- Information ratio ≈ skill per bet × the square root of the number of bets.
- Turnover
- How much of the book is traded in a period.
- Market impact
- How far your own trading moves the price against you.
- Crowding
- Many funds holding the same positions, so they all suffer when one sells.
- Look-ahead bias
- A backtest using information that was not available at the time.
- Overfitting
- Tuning a model to past noise, so it looks good in a backtest and fails live.
Further reading
- Statistical Arbitrage in the US Equities Market – Marco Avellaneda and Jeong-Hyun Lee, 2010. Quantitative Finance
- What Happened to the Quants in August 2007? – Amir Khandani and Andrew Lo, 2008
- Pairs Trading: Performance of a Relative-Value Arbitrage Rule – Evan Gatev, William Goetzmann and K. Geert Rouwenhorst, 2006
- …and the Cross-Section of Expected Returns – Campbell Harvey, Yan Liu and Heqing Zhu, 2016. Review of Financial Studies
- The Man Who Solved the Market – Gregory Zuckerman, 2019. Renaissance Technologies and Jim Simons