Large Scale Financial Time Series Forecasting with Multi-faceted Model

Data-driven approaches using deep neural networks have been successful in modeling complex financial time series and generating accurate predictions without requiring extensive domain knowledge. However, most of the existing models that assume independent and identically distributed (i.i.d) data may not generalize well to novel situations or distributional shifts across or inside financial scenarios. To address this challenge, we introduce an invariant learning-based regularizer with relaxed bounds that expands the range of feasible solutions and mitigates over-convergence issues in Invariant Risk Minimization (IRM). In practice, the regularizer can be incorporated into both linear and nonlinear financial time series forecasting models. Experimental results on real-world large-scale financial datasets show that our proposed method enables more robust and adaptable financial forecasting models, enhancing the overall performance and generalizability of financial forecasting on both in-distribution and out-of-distribution (OOD) samples.