Reproduce “Augmented Bilinear Network for Incremental Multi-Stock Time-Series Classification” on OpenResearch

About this paper

Mostafa Shabani, Dat Thanh Tran, Juho Kanniainen, Alexandros Iosifidis

Researchers from Aarhus and Tampere Universities developed an Augmented Bilinear Network (aTABL) and its CNN counterpart (aCNN) for incremental multi-stock time-series classification, which adapts pre-trained models to new financial data efficiently without requiring access to original training data. The approach consistently outperformed traditional fine-tuning, achieving higher F1-scores while drastically reducing storage requirements for multi-stock portfolios and maintaining zero inference overhead.

Key insights

• Efficient and effective model adaptation in dynamic environments can be achieved by learning small, complementary low-rank connections instead of retraining or extensively fine-tuning full models.
• Fixing pre-trained model parameters effectively preserves prior knowledge and prevents catastrophic forgetting when adapting to new data without access to the original training set.
• The low-rank augmentation strategy is architecture-agnostic, successfully applicable to diverse deep learning models like Temporal Attention-augmented Bilinear Layers and Convolutional Neural Networks.

Summary by alphaXiv.