Investment strategies based on supervised learning

The most common neurocomputational approaches to support trading decisions are based on price returns forecasting through supervised neural networks, followed by a decision (or prescriptive) model. Alternative approaches have been proposed, including reinforcement learning and neurodynamic programming, in which a unified system is directly optimised with respect to some trading performance measure. The first paradigm may lead to significantly suboptimal investment strategies, while in the latter the learning process can be very difficult to accomplish successfully and efficiently. In this paper, we seek to demonstrate that, while preserving computational efficiency, it is possible to improve the financial performance of the forecast-based approach through a better optimization of the trading module, and also by considering more appropriate neural forecasting models. In particular, we propose more adequate ways of designing the training patterns from nonstationary price data; new trading rules based on different forecast horizons; and, the use of adaptation rules able to cope with transaction costs. These ideas are then tested and compared to some of the alternatives proposed in the literature, under different criteria, for several price time series, as well as with artificial data generated according to different stochastic models.