Week Three — Foundation model vs Brownian motion. Kronos on five-minute BTC.
AIThis post was created with the assistance of artificial intelligence (AI).

TL;DR

Before you orderOffer from Amazon

Get emergency and outdoor gear delivered free with Prime

  • Fast, free delivery on millions of items
  • Prime Video, Amazon Music and more included
  • Member-only deals all year
Start your free Prime trial Free trial for eligible customers · Cancel anytime
As an affiliate, we earn on qualifying purchases.

After two weeks of testing, a foundation model called Kronos was compared to a traditional Brownian motion approach for five-minute Bitcoin forecasts. The results show Brownian motion remains more reliable in out-of-sample testing, raising questions about the practical advantage of learned models in this context.

Recent testing indicates that the Kronos foundation model does not outperform traditional Brownian motion in predicting five-minute Bitcoin price movements in out-of-sample data, challenging assumptions about the superiority of learned models in this domain.

Researchers compared the performance of Kronos, an open-source foundation model trained on global exchange data, against a geometric Brownian motion baseline using historical trade data from Polybot across 497 trades. The evaluation focused on probabilistic accuracy, using metrics such as Brier score and log-loss, as well as hypothetical profit outcomes if each model’s predictions were used for trading decisions.

The results showed that Brownian motion slightly outperformed Kronos on the full sample, with lower Brier scores and log-loss values. Specifically, Brownian achieved a Brier score of 0.193 compared to Kronos’s 0.213, and a log-loss of 0.567 versus 1.080. On the out-of-sample test set of 249 trades, the difference was statistically insignificant, with Kronos’s Brier score at 0.189 and Brownian’s at 0.188, indicating comparable predictive accuracy.

Why It Matters

This finding suggests that, despite advances in machine learning, traditional stochastic models like Brownian motion still hold practical predictive value for short-term crypto trading. The lack of out-of-sample improvement from Kronos raises questions about the real-world advantage of complex foundation models over simpler, well-understood models in this specific financial context.

For traders and quantitative researchers, this underscores the importance of rigorous out-of-sample testing before deploying learned models in live environments, especially in volatile markets like cryptocurrencies.

As an affiliate, we earn on qualifying purchases.

Background

Over the past two weeks, the author’s trading bot, Polybot, which is developed privately and is not publicly available, tested multiple strategies against Polymarket’s five-minute Up/Down markets. The initial findings indicated that most “edges” were mechanical artifacts rather than genuine predictive signals, prompting investigation into whether modern models could do better. Kronos, a highly regarded foundation model trained on extensive global exchange data, was chosen for this purpose. Prior to this, traditional Brownian motion models have been a mainstay in financial modeling, based on assumptions from the early 20th century about market randomness.

“Despite the sophistication of Kronos, it does not outperform the traditional Brownian baseline in out-of-sample tests for short-term BTC predictions.”

— Thorsten Meyer

“Kronos is designed as a research tool, not a trading system, and its performance should be evaluated accordingly.”

— Research team behind Kronos

What Remains Unclear

It remains unclear whether future model improvements, different training data, or alternative configurations could enable foundation models like Kronos to surpass traditional stochastic models in short-term crypto prediction. The current tests are limited to one version of Kronos and specific market conditions.

What’s Next

Further research may explore different model sizes, training approaches, or hybrid strategies combining traditional and learned models. Additionally, live testing of Kronos in real trading environments is needed to assess practical performance and robustness under market stress.

Key Questions

Why did Kronos not outperform Brownian motion in these tests?

The results suggest that, at least in this context, complex learned models may not capture additional predictive signals beyond what traditional stochastic assumptions provide, especially when tested out-of-sample.

Could Kronos perform better with different data or training methods?

Yes, it is possible that alternative training data, larger models, or different configurations could improve performance, but current evidence shows no clear advantage over Brownian motion in this setting.

Does this mean traditional models are obsolete?

No. In short-term prediction for volatile assets like Bitcoin, traditional models like Brownian motion remain competitive and sometimes superior in out-of-sample scenarios.

What are the implications for crypto traders?

Traders should prioritize rigorous out-of-sample testing and consider the limitations of complex models, especially when deploying strategies in volatile markets.

Source: Thorsten Meyer AI

You May Also Like

Five Levers, Many Hands

Thorsten Meyer AI opens a 12-part series mapping five policy tools governments are using as AI reshapes work.

Why Remote Travel Preparedness Is Getting More Equipment-Driven

Diving into remote travel, discover how advanced gear is transforming safety and independence in isolated adventures.

Mega Millions Jackpot $800 Million

The Mega Millions jackpot has grown to $800 million, making it one of the largest in U.S. lottery history. No winner has claimed the prize yet.

Next-Gen Navigation: Quantum Compass and Other Tech on the Horizon

Discover how next-generation navigation tools like quantum compasses are revolutionizing wayfinding—continue reading to explore the future of seamless movement.