Financial Fine-tuning a Large Time Series Model

TL;DR

This work asks whether a foundation time-series model can predict financial prices and answers affirmatively by finetuning TimesFM on 100M financial time points. By applying a log-domain loss and randomized masking during continual pre-training, the authors substantially improve price-prediction accuracy and macro- F1 scores relative to the baseline, and demonstrate practical value via mock trading that achieves a Sharpe ratio up to 1.68 on the S&P500 with a market-neutral strategy. The study highlights the importance of data characteristics in finance, the benefits of targeted finetuning over vanilla models, and releases code and weights to enable reproducibility and future exploration of financial time-series foundation models.

Abstract

Large models have shown unprecedented capabilities in natural language processing, image generation, and most recently, time series forecasting. This leads us to ask the question: treating market prices as a time series, can large models be used to predict the market? In this paper, we answer this by evaluating the performance of the latest time series foundation model TimesFM on price prediction. We find that due to the irregular nature of price data, directly applying TimesFM gives unsatisfactory results and propose to fine-tune TimeFM on financial data for the task of price prediction. This is done by continual pre-training of the latest time series foundation model TimesFM on price data containing 100 million time points, spanning a range of financial instruments spanning hourly and daily granularities. The fine-tuned model demonstrates higher price prediction accuracy than the baseline model. We conduct mock trading for our model in various financial markets and show that it outperforms various benchmarks in terms of returns, sharpe ratio, max drawdown and trading cost.

Figures (7)

• Figure 1: We show that the baseline foundation time-series model TimesFM fails at the task of financial market price prediction. Fine-tuning TimesFM on financial data siginificantly improves its prediction accuracy.
• Figure 2: Training and Validation loss curves for fine tuning TimesFM following the recipe in Table \ref{['table:hyperparameters']}. Training generally asymptotes at around 70% of the original loss value.
• Figure 3: Accuracy score of fine-tuned versus original TimesFM and a chance-rate model, when evaluated on the full test set.
• Figure 4: Macro F1 score of fine-tuned versus original TimesFM, when evaluated on the full test set.
• Figure 5: Realized PnL using fine-tuned TimesFM traded on S&P500 stocks using the basic strategy, assuming no trading costs.