Riskless arbitrage is when there is a guaranteed positive payoff without the commitment of any capital. Statistical arbitrage is when the positive payoff is not guaranteed, but can be expressed as a probability. Statistical arbitrage also refers to the use of statistics to identify possible arbitrage opportunities.
The biggest risk associated with statistical arbitrage is “model risk,” where the model of the world is not consistent with reality.
Pairs trading is a fundamental type of statistical arbitrage. The question is, “What security is mispriced relative to the other security?” Once the mispricing is identified, then it is a matter of implementing the trade: buy the relatively underpriced security and short the relatively over-priced security. By doing this, one can create a “market neutral” portfolio where the profit is determined by the relative mispricing, not by the general level of the security prices.
Cointegration analysis illuminates the long-run relationship between security prices. An error correction model shows how the deviations from the long-run relationship correct over time. This is perhaps one of the more interesting areas of statistical arbitrage research since it can be used to estimate the length of time a long or short position needs to be maintained to yield a profit.
It is possible to create a completely general model of statistical arbitrage through stochastic, dynamic optimization. This involves creating a model of the price processes, subject to the constraint that the initial cost of the portfolio is zero and that the portfolio is self-financing.