Evolutionary Money Management

This paper evolves trading strategies using genetic programming on high-frequency tick data of the USDEUR exchange rate covering the calendar year 2006. This paper proposes a novel quad tree structure for trading system design. The architecture consists of four trees each solving a separate task, but mutually dependent for overall performance. Specifically, the functions of the trees are related to initiating (“entry”) and terminating (“exit”) long and short positions. Thus, evaluation is contingent on the current market position. Using this architecture the paper investigates the effects of money management by comparing an entry-entry grammar with an entry-exit grammar. The former uses the same information set across all trees, while the latter has additional information about the current profit, drawdown and duration of a trade. Thus, money management is evolved endogenously instead of being superimposed as an additional layer. Strategies are evolved with and without transaction costs, and under three different kinds of utility; risk neutrality, risk aversion and loss aversion. Under frictionless trading the strategies exploit the significant mean reversion effect in the underlying returns. This is a dominating strategy for all utility functions. Under market frictions and loss aversion the strategies spend more time in a neutral position and the characteristics of the strategy returns change. Based on fitness measures it is not found that there is significant difference between the entry-entry and entry-exit strategies and from this perspective money management does not add value. Moreover, the out-ofsample performance seem to deteriorate over time – a finding that is consistent with the adaptive markets hypothesis. Significant profits are only generated in a single out-of-sample period and it can therefore be concluded that the high-frequency foreign exchange markets are highly efficient. JEL classifications: C0, C45, C53, C63, G1, G11, F31