Primary Submission Category: Causal Discovery
Causality Inspired Models for Trading-off Invariance and Prediction Error in Financial Time Series Forecasting
Authors: Yutong Lu, Daniel Cunha Oliveira, Xi Lin,
Presenting Author: Xi Lin*
Time series forecasting in finance is a pivotal task with significant implications for investment strategies, risk management, and economic planning. However, it is fraught with challenges due to the inherent complexity, noise, and volatility of financial markets. Conventional forecasting models often fail to generalize when faced with regime switching and distributional shifts. In this research, we leverage the use of causal discovery and invariant prediction techniques to resolve the aforementioned obstacles in asset returns forecasting.
We introduce a novel framework which integrates causal discovery, to identify causal predictors, with forecasting models. This approach balances the trade-off between invariance to distributional changes and minimization of prediction errors. To the best of our knowledge, we are the first to conduct comparative analysis among state-of-the-art causal discovery algorithms, for example LiNGAM, DYNOTERS, Invariant Causal Prediction, etc., benchmarked against non-causal feature selection techniques, in the application of forecasting asset returns. Our empirical evaluations demonstrate the efficacy of our approach in yielding stable and accurate predictions, outperforming baseline models, particularly in tumultuous market conditions.