CLS: an adaptive learning procedure and its application to time series forecasting

The CLS procedure constructs and trains multilayered perceptrons by adding units to the network dynamically during training. The ultimate aim of the learning procedure is to construct an architecture which is sufficiently large to learn the problem, but necessarily small to generalize well. Units are added as they are needed. By showing that the newly added unit makes fewer mistakes than before, and by training the unit not to disturb the earlier dynamics, eventual convergence to zero-errors is guaranteed. The CLS network was trained with currency exchange data for the period 1988-9 on hourly updates. The first 200 trading days were used as the training set and the following three months as the test set. The network is evaluated for long-term forecasting without feedback (i.e., only the forecast prices are used for the remaining trading days) and shows accurate prediction, making at least 20% profit on the last 60 trading days of 1989.<
>