ABSTRACT
This chapter investigates the trading performance of a novel class of recurrent neural networks: historically consistent neural networks (HCNNs). HCNNs use a large state space to model several observable time series at once. The state transition equation is st +1 = tanh(Wst ), where the training process optimizes the weights of matrix W. The present example uses liquid securities to model a world market. HCNNs allow for easy multi-step forecasts due to the simple state space formulation. This means that a trading strategy has additional information available which goes beyond a simple percentage change or directional information of more traditional models. Additionally, an ensemble of HCNN provides a distributional forecast.
The chapter presents several trading strategies derived from HCNN-specific features. While strategies do not always beat the benchmarks, it is remarkable that performance remains stable over asset classes. The additional information contained in other series therefore helps the forecasting task. Using HCNNs several general types of trading strategies become possible that would be difficult to realize with more traditional approaches. First, the model delivers forecasts for several assets at once. HCNN are helpful in diversifying a portfolio. Second, the multi-step forecasts can be used to either identify a trend over the next, say, 20 days, or it can help in timing entries and exits at the probable low and high price of an asset. Third, the HCNN ensemble distribution allows gauging the confidence of the forecast. The distribution can also help in identifying forecasts where paths split. In this case it is better to refrain from trading.
