Consistent Superiority of Fine-tuned TimesFM over AR(1) Baseline

Determine whether the fine-tuned TimesFM—obtained via continual pre-training on financial price data with log-transformed loss and randomized masking—consistently outperforms a first-order autoregressive (AR(1)) model across major financial markets (including S&P 500 equities, TOPIX500 equities, foreign exchange, and cryptocurrencies) under the paper’s mock market-neutral trading evaluation framework, as measured by Sharpe ratio and neutral cost.

Background

The paper fine-tunes the 200M-parameter decoder-only time series foundation model TimesFM on real financial price data and evaluates it via accuracy, macro F1, and mock market-neutral trading across multiple markets. While fine-tuning improves performance over the original TimesFM and random baselines, results are mixed relative to a simple AR(1) model, especially in currencies and crypto, prompting an explicit acknowledgment of uncertainty regarding consistent superiority.

Establishing whether the fine-tuned TimesFM reliably surpasses AR(1) under the authors’ trading and evaluation setup (e.g., horizon-based market-neutral strategies, Sharpe ratio, and neutral cost) is central to justifying the added complexity of foundation models for financial forecasting. This uncertainty motivates further investigation into data balance, loss functions, and fine-tuning strategies to close the gap.