Comparative Evaluation of Continual Learning Methods in Financial and Industrial Time-Series Data

Machine learning has seen success in financial and industrial domains, but adapting to changing data is challenging. Retraining models with new data can lead to catastrophic forgetting, especially in time-series data. While Continual learning (CL) offers a solution, its application to time-series data remains understudied compared to computer vision. Leveraging publicly available time-series datasets known for their high data volume, real-time decision demands, and potential for significant events, we comprehensively assess the performance, challenges, and limitations of CL methods. In the financial sector, we predict stock price movements during market regime shifts, while in the industrial sector, we forecast machines’ remaining useful life. We evaluate a range of CL approaches, including Naive, Replay, GDumb, EWC, GEM and LwF. Through this comparative evaluation, we discuss the challenges faced in time-series applications that are not covered by existing CL approaches, we show the domain-specific nature of problems within the time-series space and, lastly, we highlight the importance of balancing adaptation and consolidation in CL for time-series data.