Model cards¶
Every model Tycheon can route to gets a card, and every forecast references the card of
the model that produced it (ForecastDistribution.model_card). No card, no routing: a
contract test fails if a forecaster's card is missing or lacks a required section.
The cards¶
| Forecaster id | What it is | Joint sample paths | Card |
|---|---|---|---|
kronos-mini |
Kronos, 4.1M parameters, 2048-bar context | yes | Kronos mini |
kronos-small |
Kronos, 24.7M parameters, 512-bar context | yes | Kronos small |
kronos-base |
Kronos, 102.3M parameters, 512-bar context | yes | Kronos base |
timesfm-2.5-200m |
TimesFM 2.5, 200M parameters | no, marginal quantiles | TimesFM |
chronos-2 |
Chronos-2, 120M parameters | no, marginal quantiles | Chronos-2 |
random-walk |
Driftless log-price random walk | yes | Random walk |
drift |
Random walk with estimated drift | yes | Drift |
seasonal-naive |
Seasonal random walk | yes | Seasonal naive |
arima |
ARIMA on log returns | yes | ARIMA |
garch |
GARCH(1,1) on log returns | yes | GARCH |
residual(<base id>) |
A base forecaster plus a LightGBM correction from covariates | as the base | Residual-corrected |
regime-ensemble |
Members weighted by recent per-regime performance | pool mode: yes | Regime ensemble |
Kronos-large (499.2M parameters) is listed upstream as not publicly available, so it is
not supported and has no card.
What a card states¶
Required headings, enforced by tests/test_forecaster_contract.py:
| Section | Contents |
|---|---|
| Source | Where the model and its code come from, the version or commit pinned, who to cite. |
| License | The licence of the weights and of the code, and what that means for redistribution. |
| Training data | What the publisher says it was trained on. Where the publisher says nothing, so do we. |
| Limitations | Known ways it misleads, including what Tycheon observed. |
| Calibration | Whether, how, and on what window its uncertainty has been checked. |
Cards also say what it is for, how Tycheon runs it, and how it compares with the baselines.
Rules the cards follow¶
- Measured, with the window. A number on a card says what was measured, on what data, and on what machine. A single observation is called a single observation.
- Publisher claims are attributed. Training-data and licence statements are quoted from the publisher's own card and repository, with the date they were read. Tycheon does not restate them as its own findings.
- No skill claims. Each card's baseline-comparison section reports the published walk-forward result against the random walk with Diebold-Mariano tests, or says the model was not in the run. Nothing here asserts that any model beats the random walk.
- Every forecast is uncalibrated as built. Conformal calibration (
ConformalCalibrator) can wrap any forecaster, and a forecast is only labelledcalibratedwhen a holdout supports it.
For research and risk analytics. Not investment advice.