Time GPT and the Quest for Foundation Forecasting

Time-series foundation models (TSFMs) are trained once on very large, heterogeneous collections of sequences and then reused as strong general-purpose forecasters. They are "foundational" in the same sense as large language models: the pre-training phase learns broad regularities — seasonality, trend breaks, cross-scale patterns — that transfer to new series and even new domains with little or no task-specific data.

In time series, the two leading models are Google's TimesFM, a decoder-only Transformer pre-trained on roughly 100 billion real-world time points and released with checkpoints for fine-tuning, and TimeGPT-1, a commercial TSFM positioned for out-of-the-box forecasting and anomaly detection.

The appeal is practical: short time-to-value across thousands of series, flexible horizons and frequencies, and a clean path from zero-shot baseline to fine-tuned model when domain-specific data becomes available. The question is whether that appeal holds up in finance, where the data behaves differently from nearly everything else these models were trained on.

Why finance is different

Financial series violate many of the assumptions that make TSFMs work well on industrial, retail, or energy data.

Returns are heavy-tailed, exhibit clustered volatility, and undergo frequent regime changes. The predictable component of returns at daily horizons is tiny, while their variance is strongly time-varying. As a result, volatility models often outperform return-level forecasters when the target is risk, and methods must be tested for stability through turbulent periods.

A useful violation of intuition: lagged values typically help forecasting in domains like demand or energy, but linear autocorrelations of liquid asset returns are small or near zero — so naive "use yesterday to predict today" is weak. What persists is volatility, not mean. These stylized facts shape both modelling targets (quantiles and tail metrics rather than means) and evaluation protocols (rolling origins, stress windows).

Two finance use cases

1. Value-at-Risk

Goel, Pasricha, and Kanniainen (arXiv:2410.11773) evaluate TimesFM on S&P-100 daily returns over approximately 19 years, with more than 8.5 years out-of-sample. After fine-tuning, TimesFM delivers better coverage (actual-over-expected exceedances) than classical GARCH and Generalized Autoregressive Score (GAS) benchmarks, and achieves quantile-score performance comparable to the best econometric alternative.

The paper is careful to note that zero-shot use is not optimal for tails — adaptation matters — so the gain is not merely from scale but from targeted fitting. Critiques apply: the study covers one market and frequency, expected shortfall is not the central focus, and GAS/GARCH remain competitive baselines that risk teams already trust.

2. Operations at scale

TradeSmith reports replacing legacy tree-based models with TimeGPT to forecast 22,000+ financial series daily and 100,000+ forecasts per month, citing higher accuracy and lower latency with minimal tuning.

As impressive as that sounds, from a machine-learning operations perspective, it pays to be conservative. Experience from the M4 and M5 forecasting competitions cautions that simple or hybrid statistical methods are genuinely hard to beat, and claims for Transformer-based models should be verified against strong baselines and careful data handling before they are taken at face value.

Practical considerations

The pattern across studies is clear: pre-training helps, but most of the lift arrives after fine-tuning.That raises two practitioner concerns.

Data contamination. If a pretrained TSFM has seen your evaluation period (or close proxies) during pretraining, backtests can be subtly biased upward. The LLM literature has documented how benchmark overlap inflates scores (Deng et al., 2024; Jiang et al., 2024); the same risk exists for TSFMs unless you enforce a pretraining cut-off date and test on instruments or windows demonstrably unseen by the model.

Evaluation hygiene. Prefer rolling-origin backtests, long out-of-sample spans that include stress, and risk-first diagnostics (coverage tests for VaR/ES). To guard against research debt, apply backtest-overfitting checks — such as the probability of backtest overfitting or deflated Sharpe-style adjustments — when strategies depend on model forecasts.

These steps do not negate the value of TSFMs; they make their evaluation comparable to robust econometric baselines and keep governance tight when models are used in risk.

Where synthetic data fits

Foundation models excel at generalisation, but specialisation requires fine-tuning — and fine-tuning is only as good as the data available for it. Financial time series present a structural problem here: the events that matter most for risk (tail events, regime breaks, liquidity crises) are by definition the rarest in the historical record.

Synthetic data addresses this directly. Using high-quality synthetic time series that embed realistic market dynamics — heavy tails, volatility clustering, asymmetry, cross-asset dependencies — a TSFM like TimeGPT or TimesFM can be fine-tuned on a virtually unlimited supply of scenarios that reflect the statistical signatures of your domain, including the extreme conditions that historical data under-represents.

This is core to what we build at Ahead Innovation Labs. Our synthetic market infrastructure generates realistic, scenario-conditioned market environments using diffusion-based generative models, providing the kind of targeted training data that lifts a foundation model from a useful baseline to a tool that handles the conditions where accuracy matters most — the tails, the regime shifts, and the crises history was too short to supply enough of.

References

• Das, A., et al. (2024). A Decoder-Only Foundation Model for Time-Series Forecasting. ICML 2024. arXiv:2310.10688

• Garza, A., Challu, C., Mergenthaler-Canseco, M. (2023). TimeGPT-1. arXiv:2310.03589

• Goel, A., Pasricha, P., Kanniainen, J. (2024). Time-Series Foundation AI Model for Value-at-Risk Forecasting. arXiv:2410.11773

• Creal, D., Koopman, S.J., Lucas, A. (2013). Generalized Autoregressive Score Models with Applications. Journal of Applied Econometrics.

• Makridakis, S., et al. (2020). The M4 Competition. International Journal of Forecasting.

• Makridakis, S., et al. (2022). M5 Accuracy Competition. International Journal of Forecasting.

• Zeng, A., et al. (2023). Are Transformers Effective for Time Series Forecasting? AAAI 2023. arXiv:2205.13504

• AI-Powered Investment Forecasting (TradeSmith). Nixtla Success Story, 2025.

• Deng, C., et al. (2024). Unveiling the Spectrum of Data Contamination in Large Language Models. Findings of ACL.

• Jiang, M., et al. (2024). Investigating Data Contamination for Pre-training Language Models. arXiv:2401.06059

• Glasserman, P., et al. (2023). Assessing Look-Ahead Bias in Stock Return Predictions with Large Language Models. arXiv:2309.17322

Further reading

• For the architectures behind synthetic generation, see our article on conditioned diffusion models.

• For measuring synthetic data quality, see our article on synthetic data accuracy.

• For why history-based validation falls short, see our article on why backtesting is not enough.

This article is for informational purposes only and does not constitute investment advice.