Week Three — Foundation model vs Brownian motion. Kronos on five-minute BTC.

Recent testing shows that the open-source foundation model Kronos does not outperform the traditional Brownian motion model in predicting five-minute Bitcoin price movements in out-of-sample data.

Over two weeks, a research effort compared Kronos, a large-scale foundation model trained on global exchange data, with a Brownian motion baseline in predicting BTC’s short-term movements. The test involved reconstructing market conditions for 497 trades and evaluating each model’s probability forecasts against actual outcomes.

The results indicate that, on the full dataset, Brownian motion slightly outperformed Kronos based on Brier scores and log-loss metrics. Specifically, Brownian had a Brier score of 0.193 versus 0.213 for Kronos, and a log-loss of 0.567 versus 1.080, suggesting that Brownian predictions were more accurate and less overconfident overall.

In the out-of-sample test, which involved 249 trades never seen by the models during training, the performance difference was negligible—only a 0.0011 Brier score advantage for Brownian—statistically indistinguishable. This indicates that Kronos did not demonstrate a significant predictive edge over the traditional model in unseen data.

Why It Matters

This finding is important because it challenges assumptions that modern, learned models necessarily outperform classical mathematical models like Brownian motion in short-term crypto trading. It suggests that, at least in this context, the added complexity of foundation models may not translate into better predictive performance, raising questions about their practical trading utility.

Background

The testing builds on prior work where a trading bot, Polybot, used a Brownian motion model to estimate BTC price probabilities, finding limited edge in live markets. Kronos, developed by a research team and trained on millions of candlesticks from global exchanges, was introduced as a potentially superior alternative. The current testing represents a rigorous, out-of-sample evaluation designed to avoid overfitting and assess real predictive power.

Previous research indicated that most trading edges identified by the bot were mechanical artifacts that did not persist outside the sample, prompting the exploration of more sophisticated models like Kronos.

“Kronos does not outperform Brownian motion in out-of-sample predictions for five-minute BTC movements, with performance differences falling within the noise margin.”

— Thorsten Meyer (researcher)

“Kronos is designed as a research tool, not a trading system, and its performance in this context highlights the challenge of applying large models to short-term market prediction.”

— Research team behind Kronos

What Remains Unclear

It remains unclear whether different configurations, larger models, or alternative training data could yield better predictive performance. Additionally, the impact of live trading conditions, including market impact and latency, has not yet been tested.

What’s Next

Further research will explore larger Kronos models, different market conditions, and real-time deployment to assess whether foundation models can eventually surpass traditional approaches in crypto trading. Ongoing testing and refinement are expected in the coming months.

Key Questions

Why didn’t Kronos outperform Brownian motion in this test?

The current results suggest that, in this short-term prediction context, the complex learned features of Kronos did not provide a significant edge over the classical Brownian model, which is a well-understood baseline.

Can foundation models like Kronos be useful for trading at all?

While this test shows no clear advantage in short-term prediction, foundation models may still have potential in longer-term forecasts, risk management, or other trading strategies. Further research is needed.

What are the limitations of this current testing?

The evaluation was conducted offline on historical data, and real-time market dynamics, including liquidity and latency, were not simulated. Larger models or different training approaches might perform differently.

Will future tests change these results?

Possibly. Ongoing research and model improvements could lead to better performance. The current findings are specific to the tested models and data conditions.

Source: Thorsten Meyer AI