Explicit model memorisation to fight forgetting in time-series prediction

The catastrophic forgetting of the previously learned patterns during continuous life-long re-training of the models is a well-known machine learning problem. We propose an explicit model memorization algorithm for a given input data pattern. Future predictions are made by the fittest model in the accumulated library for the given new input data. Selection of the fittest model is performed by the continuously retrained 'librarian' artificial neural network (ANN). Computational experiments are executed on 17 years of data for NASDAQ, DOW, NIKKEI, and DAX composite indexes for various ANN architectures and activation functions. Results demonstrated statistically significant improvement in the accuracy metrics of the explicit model memorization compared to the immediately preceding data training.