Prediction Engine V2

Prediction Engine V2

V2 is the established prediction engine. It estimates positive cutting times from prior results, then chooses handicap marks for the complete field. Version 2.0.1 preserves the 2.0.0 calculations.

What affects a prediction?

The model uses stable competitor identity, event, dated cutting times, diameter, species and the packaged timber properties. Gender is an active category, with a separate missing category. Personal history is combined with population estimates, so sparse history does not become an overconfident personal average.

Only valid results before prediction_as_of are admitted. Same-day, future, undated and invalid rows are excluded. Earlier results contribute recency, bounded trend and cross-event information. See recency.

Quality, moisture, division, heat, lane, weather, equipment, fatigue and same-tournament weighting do not change V2 numbers. Accepted compatibility fields are disclosed as ignored rather than presented as measured effects. See wood and diameter.

What comes back?

Normal core predictions contain a predicted median time, a calibrated central 90% interval, a separate performance standard deviation and the calculated mark. Manual overrides have interval=None. A degraded panel fallback supplies a broad prior interval with calibration_state="broad_prior", not the core's calibration guarantee. Versions, cutoff, warnings and degraded state show which case was used.

The optional residual correction is inactive in the 2.0.0 release. LLM output cannot supply numeric predictions. Legacy keys explains the five-key compatibility view.

How are marks chosen?

The optimizer compares whole-field sheets using 2,048 fixed simulation samples and at most eight coordinate passes. It prioritizes equal model-implied winning chances, then expected finish spread, departure from rounded time-gap marks, and a stable input-order tie-break.

Marks are bounded integers from 3 to 183, preserve median ordering and include a Mark 3. If the search fails, V2 returns a bounded rounded-gap sheet and exposes the fallback. Simulation explains the different uncertainty measures.

Published accuracy evidence

The frozen 128-row temporal test recorded MAE of 16.1301 seconds against 20.5172 for the strict incumbent; RMSE was 33.6904 against 44.4791. Central 90% interval coverage was 94.53%. These figures apply to that workbook and split, not every future event.

In the source checkout, python train_model.py verifies the published report and artifact without reopening the locked test. Do not rerun --open-locked-test for that release. Current preview experiments use separate evidence.

Integration

Public calculation is stateless. Authenticated ledger and shadow routes are separate facilities for stored receipts and settlements. They are not automatically used by STRATHEX's direct calculation. See REST API and the full V2 specification.