Pretick reference
What can AI tell us about the next price?
Research, model backgrounds and benchmark results in one place. Another data point for understanding what numerical AI can do.
Financial forecasting studies
These notes summarize the authors' reported findings. They are not Pretick-run benchmarks or a ranking of our hosted models. Preprints may change before peer review. Reviewed September 28, 2026.
June 2026 · preprintWhat was tested: Five US equities, return forecasting, rolling evaluations and random-walk comparisons; includes Chronos-2 and TimesFM 2.5.
Reported finding: Pretrained models won many comparisons against the tested neural models, but improvements over the random-walk baseline were small and statistically significant only in a few cases.
How to read it: Compare the simple baseline before interpreting model rankings as an investment advantage.
August 2026 · preprintWhat was tested: A large financial time-series benchmark covering many public foundation models and multiple families of evaluation metrics.
Reported finding: The authors examine how conventional prediction metrics relate to financial usefulness; a good forecasting score does not automatically imply a useful financial decision.
How to read it: Match the metric to the question: next-value error, direction and trading performance are different tests.
November 2025 · preprintWhat was tested: Daily excess returns across markets; compares zero-shot use, fine-tuning and financial-domain pretraining.
Reported finding: The study reports weak off-the-shelf zero-shot and fine-tuned performance in its setup, with stronger results from finance-specific pretraining.
How to read it: Training domain matters. This is evidence for studying domain adaptation, not a verdict on every model or price series.
General forecasting benchmarks
GIFT-Eval and FEV provide public tools and datasets for comparing forecasting models across tasks. General benchmark performance provides context; it is not a financial-return backtest.
Questions to ask of a benchmark
- Was the task predicting prices, returns, volatility or direction?
- Did the evaluation preserve time order and exclude training overlap?
- Did the model beat a last-value or other simple baseline?
- Were the exact checkpoint, horizon and context comparable?
- If trading results are reported, were costs and out-of-sample periods included?
Our versioned model cards connect these questions to each hosted model's training background and supported capabilities. The library will grow as we review additional evidence.
Apply these tools with the price forecasting API guide
pretick