Construction of a General Trading Approach for Financial Markets with Artificial Neural Networks

In this work, we research several aspects of creating a general trading strategy by developing a forecasting model that uses an Artificial Neural Network (ANN) model that is based on the Convolutional Neural Network (CNN). In particular, we introduce inverted inputs and demonstrate that they reduce directional bias and reduce correlation with respect to the buy-and-hold strategy (for the underlying instrument). We empirically address issues of applying an ANN to create a trading strategy that does not use the ANN output to predict price (or its change) but provides a specific trading allocation of the underlying security for the next day of trading by using a global Sharpe-ratio-dependent cost function, instead of the often-used sum of local (or individual) squared prediction errors. The importance of the Sharpe-dependent cost function and Sharpe ratio being an appropriate measure of trading strategy is addressed and discussed. We propose a method of comparison of the trading results to random trading that employs the Sharpe-ratio distribution. We also discuss the uniqueness of the trained solution and ways to make it more independent of the initialization of the ANN's weights, either by averaging, or by the sharing of markets when pre-training the convolutional layers. The proposed method tested well in the controlled environment of artificially generated time series with different properties, extracting signal where present. It is applied to real market time series, and compared with the performance of more traditional methods, and shows promise for creating a less risky, profitable, trading strategy for a portfolio consisting of alternative investments together with the buy-and-hold of underlying securities.