The Kronos Field Kit — I Tested the AI That 34,000 Developers Starred. Here Is the Full Scoreboard
The companion to the video. Kronos is a free, open-source foundation model for financial candlesticks — built at Tsinghua University, accepted at AAAI 2026, 34,000+ GitHub stars — and one of its predictions called Bitcoin's next five days almost perfectly on my machine. So I ran the same honest test 20 times, on 5 markets (BTC, ETH, Apple, Nvidia, S&P 500), on data from months AFTER the model was trained, and kept every result. INSIDE: (1) THE FULL 20-WINDOW SCOREBOARD — every window with Kronos's call, the actual move, the direction verdict, and Kronos's path error next to a naive flat-line baseline: 10/20 direction calls (exactly a coin flip), only 2/20 windows beat the flat line (average path error 7.9% vs the flat line's 3.4%), trading every call summed to −24.9% while buy-and-hold summed to +48.2% — and the bigger Kronos-base checkpoint agreed on direction in all 20 windows, so 'use the larger model' is not the fix. (2) THE EXACT 5-RULE TEST PROTOCOL — cut the data at a date, test only on data from after training, choose windows blind on a fixed grid, always run a naive baseline, keep every result. (3) RUN IT YOURSELF IN 10 MINUTES — the pip install and the 4 lines of Python (KronosTokenizer + Kronos from_pretrained, KronosPredictor, predictor.predict) that reproduce the whole scoreboard on a laptop, no API keys, no cost. (4) WHAT KRONOS IS ACTUALLY GOOD FOR — fine-tuning on your own market, volatility and risk estimation instead of direction, synthetic candles for stress-testing, and one weighted signal inside a bigger pipeline. (5) THE 10-MINUTE AUDIT CHECKLIST for ANY AI trading tool — out-of-sample cut, naive baseline, N≥20 blind windows, every result kept, the relative-vs-absolute metric check, and the one-sentence audit: 'show me all the windows.' HONESTY: this is an educational experiment, not financial advice — the scoreboard itself is the argument against trading on any of it. Independent field kit from Hyperautomation Labs — not affiliated with the Kronos authors or Tsinghua University.
Free. No spam. Unsubscribe anytime.