Does the public checkpoint add value over simple baselines?
I separated forecast behavior from ranking and portfolio performance, then compared the public checkpoint with random, momentum and reversal strategies under the same backtest setup.
Projects/Kronos evaluation
A forecast model,
tested against the simple.
An independent evaluation of the public Kronos-small checkpoint and an inference-path optimization. Kronos itself is an upstream model by its original authors.
- My role
- Evaluation, tracing, KV cache
- Upstream
- Kronos by its original authors
- Period
- Jul – Sep 2026
- Checkpoint
- Kronos-small (public)
Bounded KV cache11.7× fasterBit-identical output · CPU, 400-bar lookback
Ranking signalp = 0.295Not significant (Newey–West, 5-path run)
Redrawn from the evaluation’s daily reports: compounded excess over CSI300 after costs. The annualized mean, +1.93%, is the figure quoted elsewhere.
Evaluation, tracing and cache optimization.
I traced the tokenizer and autoregressive prediction path, built a real-data walkthrough, evaluated the public model and implemented a bounded KV cache. On CPU, with a 400-bar lookback and 100 greedy steps, it ran 11.7 times faster with bit-identical output, for contexts within the model's valid length.
A modest result, kept in context.
In this CSI300 run, Kronos-small recorded 1.93% annualized excess return after costs; the reversal baseline recorded 1.66%. A five-path run of the same setup gave Kronos 6.01%, and its ranking signal was not statistically significant (Newey–West p = 0.295). The chart keeps the weaker random and momentum comparisons instead of showing only the favorable pair.
One public-checkpoint study.
This is not a numerical reproduction of the paper, whose best model, Kronos-large, is not public; nor is it a claim of general superiority or live performance. Its conclusions are limited to the checkpoint, data window, universe, costs and implementation used here. I did not pretrain or author the original Kronos model.