Statistical Arbitrage with Genetic Programming

Abstract

SummaryThis chapter employs genetic programming to discover statistical arbitrage strategies for a portfolio of banking stocks within the Euro Stoxx index. Binary decision rules are evolved using two different representations. The first is the classical single tree approach, where one decision tree for buy and sell orders is developed. The second version uses a dual tree structure where two decision trees are generated and the evaluation is contingent on the current market position. Hence, buy and sell rules are coevolved for long and short positions. Both single and dual trees are capable of discovering significant statistical arbitrage strategies, even in the presence of realistic market impact. This implies the existence of market inefficiencies within the chosen universe. However, the performance of the successful strategies deteriorate over time and the inefficiencies have disappeared in the second half of the out-of-sample period. The advantage of the dual trees, however, becomes apparent when transaction costs are increased and a clear asymmetric response between the two methodologies emerges. Naturally, increased costs have a negative impact on performance, but the dual trees are much more robust and can adapt to the changed environment, whereas the single trees cannot.