la-liga-score-predictor / EVALUATION_SUMMARY.md
addweb-solution's picture
Upload folder using huggingface_hub
653c38a verified
|
Raw
History Blame Contribute Delete
2.09 kB
# Evaluation Summary
## Scope
This evaluation summary is for the public `la_liga_score_predictor` release bundle.
Competition scope:
- Spanish La Liga only
Prediction scope:
- pre-match fixtures only
## What the Model Predicts
- expected home goals
- expected away goals
- final scoreline
- home / draw / away probabilities
- confidence level
- abstain / score-range signal for fragile matches
## Evaluation Snapshot
Artifact-era selection snapshot:
- exact-score rate: `0.2056`
- outcome accuracy: `0.6177`
- recent-window exact-score rate: `0.1535`
- recent-window outcome accuracy: `0.5307`
- selection method: `decode_sweep_no_retrain_recent_guardrails`
These are offline selection metrics from the model-development process.
They are useful for context, but they are not a guarantee of future live performance.
## How to Read the Numbers
- `exact-score rate`
- percentage of matches where the full predicted scoreline matched exactly
- `outcome accuracy`
- percentage of matches where the winner/draw direction was correct
- `recent-window` metrics
- the same measurements on more recent data slices
- these matter because football prediction quality can drift over time
## What This Does Not Mean
These numbers do not mean:
- the model is validated for leagues outside Spanish La Liga
- the model is validated for in-play prediction
- the model will reproduce the same scores on every new match window forever
## Practical Expectation for Developers
The most realistic use of this package is:
- a pre-match prediction input for a football product
- a confidence-aware signal, not a certainty engine
- something to pair with clear product wording around uncertainty
For most products, the most useful fields are:
- `predicted_score`
- `result_probabilities`
- `confidence_level`
- `abstain_recommended`
- `predicted_score_range` when present
## Recommendation
If you use this model publicly:
- show confidence clearly
- surface low-confidence fixtures carefully
- avoid presenting exact scores as guarantees
- keep the competition scope explicit: Spanish La Liga only