Statistical Arbitrage: Trading Many Small Edges at Once

What statistical arbitrage means

Statistical arbitrage, usually shortened to stat arb, is a family of strategies that take many simultaneous positions based on statistical relationships, each with a small expected edge, constructed so that broad market movements cancel out. The word arbitrage is loose: nothing is riskless. What is meant is that the strategy profits from statistical regularities rather than from directional views.

The canonical version generalises pairs trading. Rather than modelling one relationship between two stocks, it models the expected return of every stock given a set of common factors, then trades the residual: the part of each stock’s move that the factors do not explain, betting that this residual mean reverts.

How a residual mean reversion model works

1. 1

   Define the factor model

   Regress each stock’s returns on a set of common drivers: the market, sector or industry, and style factors such as size, value, and momentum. Principal component analysis is a common alternative that extracts factors directly from the covariance structure.

2. 2

   Compute the residual

   The residual is the stock’s return minus what the factors explain. If a stock rises 3 percent on a day its sector rose 3 percent, the residual is roughly zero and there is no signal.

3. 3

   Build a score from accumulated residuals

   Sum residuals over a short window and standardise, producing a z-score per stock. Large positive scores mean the stock has drifted above what its factor exposures justify.

4. 4

   Rank and select

   Short the highest scores, buy the lowest, typically taking the top and bottom deciles across several hundred names.

5. 5

   Optimise the portfolio

   Solve for position sizes that maximise expected residual return subject to constraints: net market exposure near zero, sector exposures bounded, position limits, turnover limits, and estimated transaction costs.

6. 6

   Rebalance on a schedule

   Daily or intraday, replacing positions as scores decay. Turnover is high, which is why cost modelling is embedded in the optimisation rather than applied afterwards.

Signals commonly used in stat arb

Individually, each of these produces an information coefficient, the correlation between prediction and outcome, of perhaps 0.02 to 0.05. That sounds negligible, and it is, for a single position. Combined across hundreds of names with careful risk control, it produces a usable return stream. This is the fundamental law of active management at work: performance depends on both the quality of a signal and the number of independent bets made with it.

Risk management in a neutral portfolio

• Factor exposure limits. Net beta near zero, bounded sector and style exposures, and limits on exposure to any single factor. Re-checked at every rebalance.
• Position limits. No single name above a small fraction of gross exposure, and limits relative to average daily volume so positions can be exited.
• Turnover and cost budgets. The optimiser must trade off expected alpha against estimated market impact, otherwise it will churn the portfolio for edges smaller than the cost of capturing them.
• Leverage discipline. Neutral portfolios have low volatility, which invites high leverage. The August 2007 quant crisis demonstrated how leverage plus crowding turns a low-volatility strategy into a violent one.
• Stress testing. Simulate a forced deleveraging scenario where correlations spike and everyone exits the same residual positions simultaneously.
• Borrow and short constraints. Short legs require availability; a strategy whose shorts cannot be borrowed is not implementable.

A realistic version for an individual

Competing with well-resourced firms on daily-rebalanced residual reversal is not realistic. A scaled-down implementation changes the parameters to those an individual can sustain.

Universe

100 to 300 liquid stocks, or a set of sector ETFs, with point-in-time membership data.

Factor model

A simple one: market plus sector dummies, estimated on 250 days of returns. Sophistication here adds less than data quality does.

Signal

Residual return over 5 days, standardised. Optionally blended with a longer-horizon momentum residual.

Holding period

5 to 10 days rather than 1 to 2, which cuts turnover and cost dramatically at a modest cost in signal decay.

Portfolio

10 to 20 longs and 10 to 20 shorts, equal risk weighted, with sector exposures capped rather than perfectly neutralised.

Rebalance

Weekly, at the open, with a no-trade band so that small score changes do not generate turnover.

Leverage

At most 1.5x gross. The temptation to lever a low-volatility strategy is exactly the mechanism that produces disasters.

Even this version requires reliable point-in-time data, the ability to short a list of names, and disciplined cost accounting. If any of those is unavailable, a long-only factor implementation is the honest alternative.

Testing stat arb without fooling yourself

• Point-in-time everything. Index membership, sector classifications, fundamentals, and shares outstanding all change and are frequently restated. Using current values is the single most common source of inflated results.
• Model market impact, not just spread. Stat arb trades many names at once; the aggregate footprint matters. Use a square-root impact model as a minimum.
• Account for borrow costs and availability on every short, historically, not just today.
• Test the portfolio, not the signal. A signal with a positive information coefficient can still produce a losing portfolio after constraints and costs.
• Deflate for the number of models tried. Stat arb research involves testing many specifications; adjust your significance threshold accordingly.
• Simulate a crowding event. Add a scenario where residual positions move against you by several standard deviations simultaneously and confirm the leverage survives it.