Using Kalman-Filtered Radial Basis Function Networks to Forecast Changes in the ISEQ Index

A Kalman-Filtered Feature-space approach is taken to forecast changes in the ISEQ (Irish Stock Exchange Equity Overall) Index using the previous five days' lagged returns solely as inputs. The resulting model is tantamount to a time-varying (adaptive) technical trading rule, one which achieves an out-of-sample Sharpe ('reward-to-variability') Ratio far superior to the 'buy-and-hold' strategy and its popular 'crossing moving-average' counterparts. The approach is contrasted to Recurrent Neural Network models and with other previous attempts to combine Kalman-Filtering concepts with (more traditional) Multi-layer Perceptron Network models. The new method proposed is found to be simple to implement, and, based on preliminary results presented here, might be expected to perform well for this type of problem.