Improve Deep Learning Performance in Stock Time-series Prediction with Importance Weighting and Transfer Learning
Open Access DepositedThe financial stock market is recognized as one of the most challenging domains for time-series forecasting, because the stock data always presents high volatility, non-stationarity nature, and complex temporal dependencies. Deep learning has emerged as a promising solution, thanks to its advantages in large-scale data processing and automatic feature selection. However, deep learning models are limited in explicitly capturing known data pattern, such as recency bias and temporal dependencies. This study examines two advanced techniques, Importance Weighting and Transfer Learning, to enhance the performance of deep learning models in stock market prediction. For Importance Weighting, we introduce a novel application of the Cumulative Importance Weighting (CIW) method and compare its effectiveness to the widely used Attention mechanism. Surprisingly, despite its conceptual simplicity, CIW consistently outperforms Attention in forecasting stock time-series. We also experiment with various ways to combine CIW and Attention, and find that one fusion approach significantly boosts accuracy beyond CIW alone, demonstrating the strength of a dual learning framework. For Transfer Learning, our experiments reveal that while training a model from scratch can occasionally match performance, the pretrain-and-unfreeze approach leads to far more stable and reliable fine-tuning, especially when the data is noisy or perturbed. This study further proposes a novel method using Wasserstein distance to trace distribution evolution in latent feature space. It finds that the effectiveness of Transfer Learning stems from the internal mechanism of pretrained models, which transforms inputs into generalizable internal representations in a consistent manner across all hidden layers. This finding provides a unifying explanation for the established theoretical perspectives such as feature reuse and domain alignment.
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.