Learning to Trade from Zero-Knowledge Using Particle Swarm Optimization.

Competitive co-evolutionary particle swarm optimization (CEPSO) algorithms have been developed to train neural networks (NNs) to predict trend reversals. These approaches considered the optimization problem, i.e. training of the NNs to maximize net profit and to minimize risk, as a static optimization problem. Based on the dynamic nature of the financial stock market, this paper proposes that the training should rather be treated as a dynamic optimization problem. A new dynamic CEPSO is proposed and used to train a NN on technical market indicators to predict trade actions. In addition, this paper incorporates approaches to combat saturation of the activation functions – an aspect neglected in previous research. The dynamic CEPSO is evaluated and compared with the static CEPSO approach, a buy-and-hold strategy, and a rule-based strategy. Results show that the new CEPSO performs significantly better on a selection of South African stocks.