Paper: arXiv 2412.09880

Authors: Xinghong Fu, Masanori Hirano, Kentaro Imajo

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.

Complexity vs Empirical Score

  • Math Complexity: 4.5/10
  • Empirical Rigor: 6.0/10 (capped from 8.5: transaction costs not modelled; see penalty flags)
  • Quadrant: Street Traders — empirical and applied, lower math complexity (effective empirical rigor capped from 8.5)

Why this score: The paper primarily relies on applying an existing foundation model with minor modifications (log transform, masking) rather than introducing dense theoretical mathematics, but demonstrates high empirical rigor through extensive dataset curation (100M points), backtesting across multiple markets, and published code/model weights for reproducibility.

Research Flowchart

  flowchart TD
  A["Research Goal: Can large time series models predict financial markets?"] --> B["Methodology: Continual Pre-training"]
  B --> C["Data: 100M time points<br>(Hourly & Daily prices)"]
  C --> D["Model: TimesFM<br>(Fine-tuned on Financial Data)"]
  D --> E["Evaluation: Mock Trading"]
  E --> F["Outcome: Outperformed Benchmarks<br>(Returns, Sharpe Ratio, Max Drawdown)"]
  D --> G["Comparison: Baseline Model<br>(Direct Application)"]
  G --> H["Result: Unsatisfactory Accuracy<br>(Irregular price data)"]
  F & H --> I["Conclusion: Financial Fine-tuning Essential for Price Prediction"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes continual pre‑training (fine‑tuning) of the TimesFM foundation model on financial price data to improve price prediction performance.

Source sentence
propose to fine-tune TimeFM on financial data for the task of price prediction.

Assumes 4

  • Training and validation data are drawn only from periods before 2023 to avoid look‑ahead bias. data
    Source sentence
    To avoid look-ahead bias, data from year 2023 onwards is reserved for testing.
  • Mock trading results are evaluated with zero transaction costs. execution
    Source sentence
    Realized PnL using fine-tuned TimesFM traded on S&P500 stocks using the basic strategy, assuming no trading costs.
  • Random chance benchmark is calculated as the accuracy of a model that guesses up‑ or down‑moves with the empirical up‑move frequency. statistical
    Source sentence
    For example, if 53% of the price changes in the test set is up, the random model guesses up 53% of the time, and down 47% of the time.
  • The loss function is modified (log‑transformed) but MSE still decreases, implying the model learns underlying price dynamics. model
    Source sentence
    We verified that MSE loss also certainly decreases.

Data

Backtest sample: 2023 – ? · S&P500 stocks · equities · daily

  • Topix500 stocks · Yahoo Finance market-prices public
    Source sentence
    Topix500 stocks Daily 3513 2,248,320
  • S&P500 stocks · Yahoo Finance market-prices public
    Source sentence
    S&P500 stocks Daily 3173 2,030,720
  • Cryptocurrencies (hourly) · Binance crypto-onchain public
    Source sentence
    Cryptocurrencies Hourly 79,153 50,657,920
  • Stock prices (hourly) · Yahoo Finance market-prices public
    Source sentence
    Stock prices Hourly 31,756 20,323,840

Recognised sources: Binance Yahoo Finance

Costs

Not modeled A trading or execution result is reported without costs.

Source sentence
Realized PnL using fine-tuned TimesFM traded on S&P500 stocks using the basic strategy, assuming no trading costs.

Finds 5

Reported performance (fine-tuned TimesFM (market neutral)): Sharpe 1.68 · annual return 3.6% · max drawdown 0.1%

  • Fine‑tuned TimesFM achieves higher price‑prediction accuracy than the original pre‑trained TimesFM across all horizons. predictive-accuracy
    Source sentence
    Fine‑tuned TimesFM outperforms original TimesFM on all prediction horizons, showing how fine‑tuning on financial data significantly improves performance.
  • Fine‑tuned TimesFM outperforms a random‑chance benchmark on every horizon, indicating statistically significant performance. predictive-accuracy
    Source sentence
    Fine‑tuned TimesFM outperforms random chance on all prediction horizon, hinting at statistically significant performance.
  • Macro F1‑score results mirror accuracy findings: fine‑tuned TimesFM consistently exceeds random chance and the baseline model. predictive-accuracy
    Source sentence
    the fine‑tuned TimesFM consistently outperforms random chance and the baseline model.
  • In mock trading on S&P500 daily data, the basic strategy yields up to 10% return for horizon 2, but is highly volatile. performance
    Source sentence
    While we are able to observe maximum returns of up to 10% (using horizon length of 2), this basic strategy is highly volatile
  • Using a market‑neutral strategy, horizon 128 achieves annual returns of 3.6% with an annual Sharpe ratio of 1.68. performance
    Source sentence
    H = 128 achieving annual returns of 3.6% and an annual sharpe of 1.68.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 30 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 3 papers in the archive (3 by independent authors): Time Series Foundation Models for Multivariate Financial … (discusses its findings)

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (38)

  1. Attention Is All You Need (2017)
  2. Time series analysis: forecasting and control (1970)
  3. General exponential smoothing and the equivalent ARMA process (1984)
  4. A New Approach to Linear Filtering and Prediction Problems (1960)
  5. Generalized autoregressive conditional heteroskedasticity (1986) — link
  6. Hidden Markov Models for Stock Market Prediction (2023) — link
  7. Stock Market Trend Analysis Using Hidden Markov Model and Long Short Term Memory (2021) — link
  8. Learning internal representations by error propagation (1986) — link
and 30 more
  1. Long Short-Term Memory (1997) — link
  2. Language Models are Few-Shot Learners (2020) — link
  3. N-BEATS: Neural basis expansion analysis for interpretable time series forecasting (2020)
  4. A time series is worth 64 words: Long-term forecasting with transformers (2022)
  5. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding (2018)
  6. Language Models are Few-Shot Learners (2020) — link
  7. Large Language Models Are Zero-Shot Time Series Forecasters (2023)
  8. LLM Processes: Numerical Predictive Distributions Conditioned on Natural Language (2024) — link
  9. Time-LLM: Time Series Forecasting by Reprogramming Large Language Models (2024) — link
  10. Large Language Models for Time Series: A Survey (2024) — link
  11. TEST: Text Prototype Aligned Embedding to Activate LLM’s Ability for Time Series (2024) — link
  12. Are Language Models Actually Useful for Time Series Forecasting? (2024) — link
  13. A decoder-only foundation model for time-series forecasting (2024) — link
  14. TimeGPT-1 (2024) — link
  15. Darts: User-Friendly Modern Machine Learning for Time Series (2022) — link
  16. Monash Time Series Forecasting Archive (2021) — link
  17. Informer: Beyond efficient transformer for long sequence time-series forecasting (2021)
  18. LLaMA: Open and Efficient Foundation Language Models (2023) — link
  19. Mixtral of Experts (2024) — link
  20. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale (2021)
  21. Scalable Diffusion Models with Transformers (2023) — link
  22. ViViT: A Video Vision Transformer (2021) — link
  23. On Model Evaluation under Non-constant Class Imbalance (2020) — link
  24. Macro F1 and Macro F1 (2021) — link
  25. FinGPT: Open-Source Financial Large Language Models — → 2306.06031
  26. Fine-Tuning can Distort Pretrained Features and Underperform Out-of-Distribution (2022) — link
  27. Understanding intermediate layers using linear classifier probes (2018) — link
  28. How Transferable are Video Representations Based on Synthetic Data? (2022) — link
  29. Synthetic Data from Diffusion Models Improves ImageNet Classification (2023) — link
  30. LoRA: Low-Rank Adaptation of Large Language Models (2021) — link

1 of 38 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.