Thanks for open-sourcing Kronos. I tested Kronos-mini on split-adjusted, regular-session 30-minute US OHLCV from eight liquid symbols, using a 100-candle lookback and 24-candle forecast horizon.
I ran frozen and fine-tuned walk-forward tests across four fully unseen periods. Fine-tuning reliably improved validation token loss, but did not reliably improve trading outcomes: direction accuracy ranged from 45.0% to 60.4%, and in three of four periods its absolute forecast error was worse than a naïve “last price” baseline.
I also tested a strictly causal calibration layer and a cost-aware long-only policy. They produced occasional promising months, but no stable net edge across the unseen periods.
My main question: do you recommend a task-aligned fine-tuning objective for direction/return/ranking, rather than token prediction loss alone? A calibrated confidence or direct alpha head would be especially useful for intraday users. I’m happy to share the walk-forward setup and detailed results if helpful.
Thanks for open-sourcing Kronos. I tested Kronos-mini on split-adjusted, regular-session 30-minute US OHLCV from eight liquid symbols, using a 100-candle lookback and 24-candle forecast horizon.
I ran frozen and fine-tuned walk-forward tests across four fully unseen periods. Fine-tuning reliably improved validation token loss, but did not reliably improve trading outcomes: direction accuracy ranged from 45.0% to 60.4%, and in three of four periods its absolute forecast error was worse than a naïve “last price” baseline.
I also tested a strictly causal calibration layer and a cost-aware long-only policy. They produced occasional promising months, but no stable net edge across the unseen periods.
My main question: do you recommend a task-aligned fine-tuning objective for direction/return/ranking, rather than token prediction loss alone? A calibrated confidence or direct alpha head would be especially useful for intraday users. I’m happy to share the walk-forward setup and detailed results if helpful.