Skip to content

Kronos mini

Forecaster id kronos-mini ยท tycheon.models.kronos.KronosForecaster("mini")

Source

  • Model: NeoQuasar/Kronos-mini, pinned to Hugging Face commit f4e68697d9d5aed55cef5c96aabc3376bcad9f81.
  • Tokenizer: NeoQuasar/Kronos-Tokenizer-2k, pinned to commit 26966d0035065a0cae0ebad7af8ece35bc1fb51c. Each Kronos model must be paired with its own tokenizer.
  • Code: upstream shiyu-coder/Kronos, vendored byte-identical at commit 67b630e67f6a18c9e9be918d9b4337c960db1e9a in third_party/kronos (ADR 0002). Tycheon replaces upstream's sampler so the sample paths are kept rather than averaged; a test shows the replacement matches upstream to the bit under greedy decoding.
  • Paper: Shi et al., Kronos: A Foundation Model for the Language of Financial Markets, arXiv:2508.02739. The upstream README notes acceptance at AAAI 2026. Tycheon cites the paper as the authority on the model; it has not independently re-derived its claims.

Facts below are from the upstream README, the Hugging Face cards and config.json, read 2026-10-04: 4.1M parameters; context length 2048; a decoder-only transformer (d_model 256, feed-forward 512, 4 heads, 4 layers); K-lines are quantised by the tokenizer into hierarchical discrete tokens (two 10-bit streams, s1_bits = s2_bits = 10; the tokenizer uses group size 5) before the transformer models them autoregressively.

License

MIT, for both the weights (Hugging Face metadata) and the code (upstream LICENSE, retained at third_party/kronos/LICENSE and shipped inside the Tycheon wheel). Tycheon itself is Apache-2.0; the vendored files keep their own notice.

Training data

As stated by the publisher on the Hugging Face card (read 2026-10-04): Kronos is "trained on data from over 45 global exchanges", on "a massive, multi-market corpus of over 12 billion K-line records". The card does not state the date range, the frequency mix or the split between asset classes, and does not say what the model has never seen; the paper is the place to look. Tycheon makes no claim beyond these publisher statements.

Intended use

Zero-shot forecasting of candlestick series (open, high, low, close, volume, amount). Tycheon feeds it the most recent min(len(history), lookback, 2048) bars, forecasts all six channels, returns the close as the forecast target, and keeps the other channels in ForecastDistribution.extras. Missing volume and amount are filled the way upstream does.

Limitations

  • Zero-shot and uncalibrated. The spread of the sample paths is whatever the pre-trained model believes. Nothing has checked that a 90% interval contains the outcome 90% of the time.
  • Not anchored to the last close. The forecast is decoded from tokens, and the tokenizer is lossy. On one synthetic series (SYN-GARCH, 900 bars known) the tokenizer round-trip of the last close was off by +0.75%, and the median of the first forecast step sat +1.34% from the last close, against a daily volatility of about 1.2% on that series. This is the model's behaviour, not a porting error (the port is bit-identical to upstream). Measuring and correcting it belongs to calibration.
  • Spread, one series. Across 20 sample paths the standard deviation of log(P[10] / P[0]) was 0.0309, against 0.0336 for the random walk's volatility scaled to the same horizon (ratio 0.92x). With 20 paths that standard deviation carries roughly +-16% sampling error, and it is one series.
  • Context. The model was trained for 2048-bar contexts; Tycheon never feeds it more and cuts longer histories to the most recent bars before normalising, so the statistics describe what the model reads (upstream normalises over everything it is handed). The forecast horizon must not exceed 2048 bars: upstream decodes only the last 2048 tokens, so a longer horizon would silently return the wrong shape, and Tycheon raises instead.
  • Frequencies and markets are not specified. The cards do not say which bar sizes it was trained on. Calendar features (minute, hour, weekday, day, month) are read from the history's own timezone; for intraday data pass bars in the exchange's local time.
  • Upstream quirks Tycheon refuses to inherit: upstream ignores top_p when top_k > 0 (Tycheon rejects the combination), never calls .eval() (Tycheon does, because dropout left on would add noise), and averages its sample paths into one (Tycheon keeps them).
  • Cost. CPU-only, no GPU: 23 s for 20 paths over a 900-bar context and a 10-step horizon, plus 1 s to load, on the development machine (Windows 10 Pro, CPU only, PyTorch 2.14.1). Cost grows with paths, context and horizon; CUDA and Apple MPS are supported by the device switch but were not tested on any accelerator.

Calibration

Uncalibrated as built (calibration_status="uncalibrated"): the spread of the sample paths is whatever the pre-trained model believes. Calibration is available for any forecaster: wrap it with ConformalCalibrator (split or adaptive conformal) and the forecast carries a status (calibrated, stale or uncalibrated) earned on a holdout it was not fitted on. See calibration. In the small benchmark (benchmarks/results/small/0.1.0.json: Tycheon 0.1.0, three synthetic series, horizon 5, 108 pooled walk-forward origins), zero-shot Kronos-mini's nominal 90% interval contained 57.4% of outcomes, so its raw intervals were overconfident on that data. It ran with 20 sample paths, against 100 for the other models, so its tail quantiles are noisy.

How Tycheon runs it

Weights and tokenizer load with from_pretrained(..., revision=<pinned SHA>) from safetensors files (no pickle, so loading executes no code), are moved to the chosen device and put in eval mode with gradients off. Sampling exposes temperature, top_k and top_p (default T=1.0, top_k=0, top_p=0.9, as upstream) and an explicit seed driving a torch.Generator; the seed and every setting are recorded in the forecast metadata. Series are batched in chunks of max_batch_rows paths, and predict_batch accepts several series of differing lengths (upstream requires equal lengths).

Baseline comparison

Measured in the small benchmark (benchmarks/results/small/0.1.0.json: Tycheon 0.1.0, three synthetic series, horizon 5, 108 pooled walk-forward origins). Pooled MASE against the no-change forecast: 1.025. Achieved coverage of the nominal 90% interval: 57.4%. One-sided Diebold-Mariano p-values that it has lower loss than the random walk: 0.685 (squared error) and 0.756 (CRPS), so it is indistinguishable from the random walk. Its 57.4% directional accuracy and IC of 0.168 are within what noise produces with 108 origins. It ran zero-shot with 20 sample paths and 256 bars of history. Synthetic data: this shows how the machinery behaves, not whether the model forecasts real markets.

For research and risk analytics. Not investment advice.