Forecasting Complex Systems with Shared Layer Perceptrons

Abstract

We present a recurrent neural network topology, the Shared Layer Percep-tron, which allows robust forecasts of complex systems. This is achieved by several means. First, the forecasts are multivariate, i. e., all observables are forecasted at once. We avoid overfitting the network to a specific observable. The output at time step t, serves as input for the forecast at time step t+1. In this way, multi step forecasts are easily achieved. Second, training several networks allows us to get not only a point forecast, but a distribution of future realizations. Third, we acknowledge that the dynamic system we want to forecast is not isolated in the world. Rather, there may be a multitude of other variables not included in our analysis which may influence the dynamics. To accommodate this, the observable states are augmented by hidden states. The hidden states allow the system to develop its own internal dynamics and harden it against external shocks. Relatedly, the hidden states allow to build up a memory. Our example includes 25 financial time series, representing a market, i. e., stock indices, interest rates, currency rates, and commodities, all from different regions of the world. We use the Shared Layer Perceptron to produce forecasts up to 20 steps into the future and present three applications: transaction decision support with market timing, value at risk, and a simple trading strategy.