Strategies · Library · Stocks
§3.18Pairs and relative value
Statistical arbitrage – optimization
Given expected returns for many stocks and a covariance model, choose weights that maximize return per unit of risk, optionally forced to net to zero dollars. It is a sizing method wrapped around some other signal.
- Asset class
- Stocks
- Held for
- Days
- Data
- ARCANE has part of the data
- Symphony
- Cannot be built as a symphony
The rule
- Signal
- Expected return per stock and a risk model of how stocks move together.
- Entry and exit
- Recompute weights each rebalance; trades follow the change in weights.
- Sizing
- Weights from the inverse covariance matrix times expected returns, scaled to total exposure and made dollar-neutral.
How it fails
- Estimated covariance is unstable
- Garbage-in expected returns
- Needs shorts and many names
Data it needs
- expected returns from a signal
- risk model
- daily prices
Prices permit a covariance estimate; a robust risk model and the expected returns are not provided.
How ARCANE uses it
Not a Call driver itself; Alpha could cite it as the way to size several views together.
Needs a matrix optimizer and short positions; the closest builder tool is inverse-volatility weighting, which drops the core idea.
Filed under the pairs and statistical arbitrage card
Find two securities whose prices moved together for a year. When the gap between them opens past two standard deviations, buy the laggard and sell the leader.
A paper plan from this card holds 5 to 126 days, stops at 1.5× the expected move and uses at most 50% of your risk budget. The card
Source: Kakushadze and Serur (2018), 151 Trading Strategies, §3.18, p. 56 of the SSRN version. Described in our words. Pairs and relative value is ARCANE’s filing, not the book’s.