AIToday
Large Language ModelsZenn AI/MLPublished: Oct 22, 2026, 01:00 JST

Kronos tested on BTC/USDT: 46.7%–50.0% directional accuracy, no better than coin flip

Send on LINE
Kronos tested on BTC/USDT: 46.7%–50.0% directional accuracy, no better than coin flip

3 Key Points

  1. What happened

    A tester ran Kronos, a candlestick-only foundation model accepted at AAAI 2026, on 400 hours of BTC/USDT hourly data and predicted 120 hours ahead, scoring 46.7%–50.0% directional accuracy.

  2. Why it matters

    On this single 120-hour window, the model showed no statistically meaningful directional signal — its hit rate did not clearly beat a coin toss — so its output alone is not a basis for investment decisions.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, in your inbox every morning.

Context & Analysis

The test used only public Binance REST API data, with no paid API keys and no GPU required, and the workflow — data preparation, tokenizer, model inference — ran without installation trouble on Python 3.12 with the README's dependency versions. Kronos-mini predicted 24 hours from 200 hours of lookback in 0.3 seconds on CPU, and Kronos-small's 400-to-120-hour run completed in about 20 seconds on CPU, suggesting deployment barriers are low for experimentation.

The gap between price-level fit and directional skill is the central finding. A MAPE of 2.33% looks close, but it reflects natural tracking of the market's range rather than a directional signal usable for investment decisions. The tester's own disclaimer applies: the output is one probabilistic sample, and results from one symbol and one period may differ elsewhere.

FAQ
What does Kronos actually predict?
It predicts future candlestick (OHLCV) values. The tester fed 400 hours of past BTC/USDT hourly data and asked for 120 hours ahead.
Did averaging multiple samples improve directional accuracy?
No. With sample_count=10, directional accuracy fell to 45.0% and the correlation between predicted and actual differences turned negative at -0.116.
How long did prediction take on CPU versus GPU?
The Kronos-small run took 19.62 seconds on CPU and 2.80 seconds on GPU (RTX 5060 Ti), about 7× faster, with nearly identical MAE and MAPE.
Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articlePHP2Shell: Calif and Wiz used AI to find critical PHP flaw