NBA prediction performance review

Expanded validation and source review

Open the 65-rule validation catalog. 0 local hard failures; 0 failures across all source profiles; 62 blocked checks. The full catalog requires review. Metrics affected by provider contradictions are descriptive and source-disputed. No source values were repaired.

What the results say

Decision: continue local testing; evidence is unchanged. Against its own baseline, KPI player points error changed by +0.00 points, while team score error changed by +0.00. Lower is better. Better player averages do not establish better team predictions.

The KPI forecast selected 3 correct winners from 5 picks across 6 finished games; 1 game(s) had tied predicted scores and no winner pick. This is a small diagnostic sample, not a validated win rate.

Who is doing better?

For player points, KPI beats production on 0 of 131 matched player-games, loses on 0, and ties on 131. Against its own baseline, KPI wins 0, loses 0, and ties 131.

Average error answers how large the mistakes are. Head-to-head counts answer how often a forecast is closer. A system can win more individual comparisons and still make larger mistakes elsewhere. There is no single overall champion across unrelated objectives.

Best system for team predictions

Best system for each player statistic

Each row compares identical confirmed player-games. Outright wins count a forecast only when it is strictly closer than both alternatives. Shared-best rows are counted once in the last column, not awarded as outright wins.

Which forecast is best for each player?

All 131 matched player-games are listed, sorted from greatest KPI points improvement over production to greatest deterioration. Each leader cell names the forecast closest to the actual result for that statistic. Shared leaders are shown explicitly. No combined points/rebounds/assists score is invented.

A player-game win is diagnostic evidence from this snapshot, not a reliable personal model ranking from one appearance. Missing forecasts and projected DNPs remain visible in the coverage section and are not silently scored as zero.

Performance scorecard

Errors are measured in points unless otherwise stated. Lower error is better. Winner accuracy is a separate measure and cannot substitute for score or player accuracy.

Average team score error

Each bar is the average absolute error per team across finished games. The baseline and KPI form the controlled local comparison; production is a separately captured reference.

All games at a glance

Every score is labeled by team. Error change compares Local + KPI against Local baseline: negative means improvement, positive means deterioration. Live scores are snapshots and can change.

Player accuracy by statistic

These metrics use only confirmed players present in all three forecasts and the final boxscore. Coverage exclusions are disclosed below. Rows count player-games, not unique people.

MAE: average absolute error. Bias: average prediction minus actual; positive means too high. RMSE: gives larger misses more weight. Units follow each statistic (minutes for MIN, counts for the others).

Download all matched player statistics (CSV) · Download game comparisons (CSV)

Largest KPI points misses

Sorted by absolute points error, not by the size of the KPI change. Expand each game below for every matched player.

The two components sum to signed points error: (predicted minutes − actual minutes) × predicted points/minute, plus actual minutes × (predicted rate − actual rate). They identify where to investigate; they do not establish the underlying cause.

Coverage and participation

Counts are player-game exceptions across finished games. Missing forecast coverage can make matched-only accuracy look better than full roster performance. A missing boxscore row is unverified, not automatically a DNP. Explicit DNP projections are reported separately from played-player metrics.

What to investigate before the next games

1. Team totals and aggregation. Player points accuracy and team accuracy move differently. Inspect team score reconciliation and the effect of bounded per-player changes on team totals. This is a hypothesis to test, not an established defect.
2. Minutes and availability. Review the largest minute misses and explicit DNP projections in each game. Check pregame lineup coverage and accumulated minute adjustments before changing scoring rates.
3. Scoring rate. Where minutes are close but points are far apart, inspect shot volume, efficiency, opponent adjustments and historical sample size. Preserve the existing error evidence.
4. Luna review stays in shadow. Give Luna traceable failure evidence and request bounded hypotheses. No automatic parameter changes. The report does not measure Luna as an active prediction variant.
5. Two prospective test days. Freeze one candidate version and its acceptance criteria before tipoff; save all production and local forecasts with timestamps; settle final games; compare matched-player accuracy, coverage, team totals and winner abstentions. A substantive candidate change starts a new test window. Two days test operation and early behavior; they do not prove durable accuracy.

Current disposition: keep the candidate local. No promotion decision is justified by this single-slate comparison alone.

Sources, timing and trust checks

Original pregame evidence

These are the frozen pregame forecasts for this policy version. Their capture times and paired baseline identities are preserved.

This report preserves the production snapshot captured for this run. Production can refresh later; generation and retrieval times are listed separately for every game.

To reproduce: settle results with nba_local.py cycle, then run tools/build_detailed_comparison.py against the same run. This renderer only reads evidence and writes report files; it does not modify prediction artifacts.

Game-by-game review

Open the detail rows for player comparisons, participation exceptions and timing evidence.