Dual Adaptation of Time-Series Foundation Models for Financial Forecasting

Recent progress in time-series foundation models has expanded forecasting capabilities across domains. However, their application to finance remains constrained by data scarcity, volatility, and overfitting. We present Dual Adaptation, a lightweight adaptation of the TimesFM foundation model for financial forecasting, featuring a dual-module design: a Generalizer Adapter that learns broad temporal patterns across assets and an Identity Signature module that captures asset-specific signals, forming a lightweight layer tuned on top of a frozen foundation model. The method is evaluated in both in-domain and zero-shot settings, showing improved forecasting performance compared to the frozen model and common tuning baselines. To enhance generalization, the Identity Signature is removed during inference, allowing the Generalizer Adapter to apply the shared knowledge it has learned to unseen assets. This design improves both stability and cross-asset generalization, offering a practical solution for adapting large models to noisy, low-resource financial forecasting tasks.