Statistical Arbitrage
Also known as: stat arb, StatArb, pairs trading, spread trading
What is it?
Statistical arbitrage trades the relationship between two or more instruments instead of the direction of any one of them: when a historically stable spread stretches, it sells the strong side, buys the weak side, and profits if the two come back together. Take two instruments that have tracked each other closely for years, normally staying within about one percent of their usual ratio. The gap opens to 2.6%.
A stat-arb system shorts the outperformer, goes long the laggard in matched size, and closes both legs when the gap returns to its average. Note what the position is not exposed to: if both instruments fall 5% together, the trade is unaffected, because the profit comes from convergence rather than from the market going anywhere. Despite the name, this is not arbitrage.
Real arbitrage has a mechanism that forces prices back together; here there is only a historical tendency, and tendencies end. A central bank diverges, a company issues debt, an index rebalances, and the relationship simply re-rates to a new level. The uncomfortable part is that the most attractive-looking signal - the widest spread you have ever seen - is also the reading most likely to mean the relationship has broken rather than stretched.
Why it matters: Statistical arbitrage pays when a stretched relationship snaps back, and loses most when the relationship has genuinely broken rather than stretched.
Spread z-score = (current spread - mean spread) / standard deviation of the spread
The position is market-neutral by construction, so the entire outcome rests on whether the historical relationship still holds.
Real-world example
A metals pair trade opened at a 2.3-sigma spread and converged within nine sessions, while the same signal at 2.5 sigma in a different regime kept widening for four months as the ratio re-rated to a new level.
How SignalBots handles it
SignalBots correlation views show how closely two instruments have actually been moving, so a pair trade is opened on a measured relationship rather than an assumed one. See /risk-warning.
Pro tip
Cap how long a convergence trade may stay open. A spread that has not reverted within the window it usually reverts in is evidence the relationship changed, not that the edge grew.
Common pitfalls
Adding size as the spread widens. Scaling into a diverging pair is indistinguishable from martingale until the relationship is proven intact.
Frequently asked questions
Is statistical arbitrage safe because it is market-neutral?
No. Market-neutral only removes exposure to the overall direction; it leaves full exposure to the relationship itself, which can widen indefinitely. Your capital is at risk and nothing forces the two legs back together.
How is it different from a hedge?
A hedge is intended to cancel risk on a position you already hold. A stat-arb pair is an active bet in its own right: you want the two legs to move apart from where they are now and back toward their historical relationship.
How many bars of history do I need to trust a spread?
Enough to have seen the relationship survive different regimes, which usually means several years rather than several months. A correlation measured only in calm conditions tells you nothing about what happens when volatility spikes.
Do I need to trade equal dollar amounts on both legs?
Usually not equal dollars but equal risk, sized by the ratio between the two instruments' volatility. Matching notional value alone leaves the more volatile leg dominating the result.
Why does correlation break down exactly when it matters?
Because correlations are measured over ordinary conditions, and stress changes the drivers. In a liquidity event participants sell whatever they can rather than whatever is expensive, which moves instruments together or apart for reasons the historical relationship never contained.
Trading involves substantial risk of loss. Historical and backtested results do not guarantee future performance. Read the full risk warning.
