PB-TABL: Task Incremental Learning Strategy via Applying Piggyback Architecture on Temporal Attention-Augmented Bilinear Networks for Financial Time-Series Classification

Abstract

As financial markets continue to grow and generate vast volumes of data, there is a growing need for models capable of being updated incrementally. In this paper, we present PB-TABL, a novel method for financial time-series forecasting that integrates the piggyback (PB) architecture with Temporal Attention-Augmented Bilinear Networks (TABL). The proposed PB-TABL architecture addresses the challenges of task-incremental learning, including catastrophic forgetting, concept drift, and computational efficiency. By adapting pre-trained models through binary masks, PB-TABL enables efficient reuse of models for new tasks, significantly reducing training time and computational costs. We demonstrate the effectiveness of this approach through extensive experiments on large-scale Limit Order Book (LOB) data, where PB-TABL outperforms baseline models in terms of accuracy, F1 score, and overall computational efficiency. Our contributions include formulating the problem of adapting pre-trained neural networks to new financial data, introducing the PB-TABL architecture, and showing its advantages in handling time-series data with temporal dependencies while mitigating the risks of catastrophic forgetting and model degradation.