Decision technologies for trading predictability in financial markets

Towers, Neville (2000) Decision technologies for trading predictability in financial markets. Doctoral thesis, University of London: London Business School.

Abstract

The traditional view amongst financial economists is that well-functioning financial markets are unpredictable and hence provide no means for systematic excess profit. In the last decade, however, this view has been consistently challenged by empirical studies showing low but significant levels of predictable behaviour in asset returns, and explicit evidence that statistical forecasting models can have a measure of significant predictive power. If these apparent "systematic regularities" exist then it invites the development of methodologies to exploit predictability by generating economically significant trading strategies. This thesis develops a methodology to optimise trading strategies for arbitrary forecasting models which possess some degree of predictive ability. It exploits recent advances in decision theory to account for both expected returns and trading costs within the decision making process and thus reflect the reality of typical trading environments. In this context, we develop a framework to jointly optimise the trading performance for a pair of decision and forecasting models. The methodology consists of two stages. First, given an arbitrary forecasting model, we develop methods to approximate the optimal trading strategy dependent on the trading conditions. In particular, we develop trading strategies using parameterised decision rules and an enhanced reinforcement learning algorithm. Secondly, given a trading policy, we examine the multi-objective optimisation of the forecasting model. We describe a meta-parameter approach in which the forecasting model is optimised with respect to a number of different statistical characteristics which affect trading performance. We investigate three specific characteristics, namely forecast horizon, prediction smoothness and predictive correlation. The two stages are then combined to perform a joint optimisation over both forecasting and decision models. We empirically evaluate optimisation procedures using controlled simulations and in the application of statistical arbitrage trading. Our results demonstrate that joint optimisation can significantly improve performance in the presence of trading costs.