RefineBridge: Generative Bridge Models Improve Financial Forecasting by Foundation Models

TL;DR

RefineBridge introduces a generative refinement module based on Schrödinger Bridge to enhance transformer-based financial forecasting models, significantly improving their accuracy across various benchmarks.

Contribution

The paper presents a novel refinement module, RefineBridge, that leverages Schrödinger Bridge generative framework to improve TSFM predictions in financial time series forecasting.

Findings

RefineBridge consistently improves TSFM performance across multiple benchmarks.

The method effectively handles non-stationarity and noise in financial data.

Significant accuracy gains over baseline TSFMs are demonstrated.

Abstract

Financial time series forecasting is particularly challenging for transformer-based time series foundation models (TSFMs) due to non-stationarity, heavy-tailed distributions, and high-frequency noise present in data. Low-rank adaptation (LoRA) has become a popular parameter-efficient method for adapting pre-trained TSFMs to downstream data domains. However, it still underperforms in financial data, as it preserves the network architecture and training objective of TSFMs rather than complementing the foundation model. To further enhance TSFMs, we propose a novel refinement module, RefineBridge, built upon a tractable Schr\"odinger Bridge (SB) generative framework. Given the forecasts of TSFM as generative prior and the observed ground truths as targets, RefineBridge learns context-conditioned stochastic transport maps to improve TSFM predictions, iteratively approaching the ground-truth target from even a low-quality prior. Simulations on multiple financial benchmarks demonstrate that RefineBridge consistently improves the performance of state-of-the-art TSFMs across different prediction horizons.