Can an AI foundation model predict crypto prices?
Not the direction, in our test. Kronos, an open foundation model pre-trained on 12 billion candlesticks, called the sign of bitcoin’s and ether’s next 1, 4 and 24 hours correctly 49.8% to 51.6% of the time on data after its training cut-off (July 2024 to October 2026). The best configuration looked good before costs but did not stand out from chance, and the hourly versions lost money after fees.
The claim
Kronos treats candlesticks as a language: a tokenizer turns each bar’s open, high, low, close and volume into tokens, and a decoder is pre-trained on more than 12 billion bars from 45 exchanges. The authors report large zero-shot gains on benchmarks — a 93% higher rank correlation in price forecasting than the leading time-series foundation model — and publish the weights, which has made the model a popular choice for crypto forecasting.
What we tested
We ran the public Kronos-small and Kronos-mini models zero-shot, with the authors’ own code, on hourly BTC and ETH perpetual futures from July 2024 to October 2026 — after the end of the pre-training data, so the model had never seen these prices. Every hour it forecast the next 1 and 4 hours from the previous 400 bars, and once a day the next 24 hours, from several sampled paths. A strategy held the sign of each forecast, and twelve configurations (two models, three horizons, two coins) were judged against the same positions shifted in time.
Result
Hit rates ranged from 49.8% to 51.6%, and rank correlations with the outcome from about zero to +0.07. The best configuration, the smaller model on bitcoin’s next four hours, reached a gross Sharpe ratio of +1.31, but as the best of twelve tries it is not distinguishable from luck. The hourly and four-hour versions trade 5 to 20 times a day, and at about 0.055% per trade they lose: net Sharpe ratios of −1.5 to −8.3. The daily forecasts were small and indistinguishable from noise. An independent check on daily and hourly bitcoin with the largest public model found no correlation between forecasts and outcomes either.
Kronos direction forecasts, July 2024 – October 2026 (after the training cut-off)
| Measure | Result |
|---|---|
| Hit rate, 12 configurations | 49.8% – 51.6% |
| Rank correlation with the outcome | ≈ 0 to +0.07 |
| Best gross Sharpe ratio (BTC, 4 hours) | +1.31, within chance |
| Net Sharpe ratio, 1-hour and 4-hour versions | −1.5 to −8.3 |
What it means
Pre-training on billions of bars teaches a model what candlesticks look like, which is not the same as knowing where the next one will close. For direction, a foundation model is one more forecaster that has to beat a coin flip out of sample and after costs, and on crypto it has not done so yet. This note covers the direction of returns only.
What would change the verdict
Fine-tuning on crypto, larger model sizes and using the forecasts as one input among others were not tested here. Forecasts logged before the outcomes that beat a coin flip after costs would change the verdict.
This note re-tests a claim made by the cited authors on our own crypto market data; the summary of their method is ours. Tested · published · last reviewed . Research, not investment advice.