Predictable variation and profitable trading of US equities: A trading simulation using neural networks

A switching rule conditioned on out-of-sample one-step-ahead predictions of returns is used to establish investment positions in either stocks or Treasury bills. The economic significance of any discernible patterns of predictability is assessed by incorporating transaction costs in the simulated trading strategies. We find that artificial neural network models produce switching signals that could have been exploited by investors in an out-of-sample context to achieve superior cumulative and risk-adjusted returns when compared to either regression or a simple buy-and-hold strategy in the market indices. The robustness of these results across a large number of stock market indices is encouraging.