[Paper Review] Test-Time Adaptation for Non-stationary Time Series: From Synthetic Regime Shifts to Financial Markets

Yurui Wu, Qingying Deng+2 more|arXiv (Cornell University)|Jan 20, 2026

Stock Market Forecasting Methods0 citations

TL;DR

The paper studies a lightweight test-time adaptation framework for causal time-series forecasting and direction classification, freezing the backbone and updating only normalization parameters, with fallbacks to keep updates stable across regime shifts including financial markets.

ABSTRACT

Time series encountered in practice are rarely stationary. When the data distribution changes, a forecasting model trained on past observations can lose accuracy. We study a small-footprint test-time adaptation (TTA) framework for causal timeseries forecasting and direction classification. The backbone is frozen, and only normalization affine parameters are updated using recent unlabeled windows. For classification we minimize entropy and enforce temporal consistency; for regression we minimize prediction variance across weak time-preserving augmentations and optionally distill from an EMA teacher. A quadratic drift penalty and an uncertainty triggered fallback keep updates stable. We evaluate this framework in two stages: synthetic regime shifts on ETT benchmarks, and daily equity and FX series (SPY, QQQ, EUR/USD) across pandemic, high-inflation, and recovery regimes. On synthetic gradual drift, normalization-based TTA improves forecasting error, while in financial markets a simple batch-normalization statistics update is a robust default and more aggressive norm-only adaptation can even hurt. Our results provide practical guidance for deploying TTA on non-stationary time series.

Motivation & Objective

• Understand when small-footprint test-time adaptation (TTA) helps or hurts non-stationary time series.
• Unify adaptation options (no adaptation, BN statistics refresh, norm-only) under a causal framework.
• Evaluate TTA on synthetic regime shifts and real financial data across pandemic, inflation, and recovery periods.
• Provide practical guidance and regime-wise evaluation tools for deploying TTA in streaming forecasting.

Proposed method

• Freeze backbone f_theta and adapt a small set of norm-affine parameters (gamma, beta) at deployment.
• Use unsupervised objectives: entropy and consistency for classification; variance and EMA-teacher distillation for regression.
• Employ an uncertainty proxy to trigger a BN statistics refresh and skip updates when high uncertainty.
• Apply weak time-preserving augmentations to generate multiple transformed views for robust unsupervised learning.
• Incorporate a quadratic drift penalty to limit day-to-day parameter changes and improve stability.
• Ground updates with Diebold–Mariano tests and regime-wise evaluation to guide practical deployment.

Figure 1: Regime diagnostics for SPY volatility and returns.

Experimental results

Research questions

• RQ1 How does test-time adaptation using norm-only BN affine parameters perform under non-stationary regime shifts in time series?
• RQ2 Is BN-statistics refresh a safer default than norm-only updates in real-world financial data?
• RQ3 How do different TTA variants (no adaptation, bn_stats, norm_only) compare across synthetic shifts and financial regimes?
• RQ4 What are the statistical and economic implications (e.g., DM tests, Sharpe ratios) of TTA in markets like SPY, QQQ, and EUR/USD?
• RQ5 What practical guidelines emerge for deploying TTA in streaming time-series forecasting?

Key findings

• Norm-only adaptation helps under gradual mean/variance drift in synthetic tests by correcting low-order moment changes.
• BN statistics refresh is a robust default on noisy financial data, often outperforming the frozen baseline across SPY, QQQ, and EUR/USD in direction accuracy and risk-adjusted metrics.
• Aggressive norm-only updates can hurt performance in real markets due to overfitting to short windows.
• Diebold–Mariano tests show bn_stats significantly better than no_tta on SPY and QQQ, and bn_stats also outperforms no_tta on EUR/USD.
• Economic backtests show bn_stats achieving higher Sharpe ratios than no_tta or norm-only on SPY and QQQ; norm-only often underperforms.
• Overall guidance: start with bn_stats, add norm-only updates only when uncertainty diagnostics support them, and evaluate regime-wise.

Figure 2: Rolling forecast metrics on ETTh1 under gradual drift.