Skip to content
artifact-backedgame SHAP available2025-10-24 to 2026-06-13

Model Insight Console

A master lens for what moves the saved NBA model artifacts: global feature pressure, calibration honesty, confidence behavior, ablation lift, and market benchmarking.

01

Elo diff is the largest global SHAP driver in the held-out GBM explanation.

02

Logistic currently has the strongest log-loss profile among saved model artifacts.

03

The market benchmark is still better on matched closing-line log loss, so edge claims stay gated.

04

Game-level SHAP detail requires data/processed/shap_game_level.parquet in this worktree.

Logistic accuracy

66.8%

1,214 walk-forward rows

Log loss

0.605

Lower is better for calibrated probability quality

Brier score

0.210

Squared probability error across held-out games

Calibration gap

3.8%

Weighted mean bin gap between expected and observed

3D feature influence terrain

This surface combines report-backed SHAP rank with the selected model's confidence-band reliability. It is a visualization of saved artifacts, not a neural-network loss surface.

Logistic
Drag to orbitScroll to zoom
Feature rankReliability bandInfluence height

Selected ridge

Elo diff

Rank 1 · 0.2357 mean abs SHAP

Team-strength prior is the biggest global anchor before recent-form signals enter.

Why this model leans

Global SHAP magnitude from the GBM report. Directional per-game SHAP is shown as unavailable when the Parquet detail is not present.

SHAP

Calibration truth table

Predicted home-win probability compared with observed home wins by probability band.

3.8% avg gap

Confidence buckets

Volume bars show where the model usually lives; the line shows hit rate inside each confidence band.

1,214 games

Feature group mass

global
Matchup37.3%
Elo strength33.9%
Availability16.4%
Rolling form12.4%

Decision anatomy status

Per-game explanation drilldown is gated by artifact availability.

Game-level SHAP detail unavailable

Add data/processed/shap_game_level.parquet to unlock matchup-specific positive and negative contributors without guessing in the frontend.

Ablation ladder

Feature-stage runs stay separated from the dashboard; this view only reads their saved outputs.

walk-forward

Rolling base

Team rolling-form columns only.

62.7%

0.00 pts

Δ log loss 0.0000

Rest and travel

Rolling form plus rest, back-to-back, travel, and venue context.

63.3%

+0.66 pts

Δ log loss +0.0011

Matchup differentials

Rolling form plus team-vs-team differential features.

64.3%

+1.57 pts

Δ log loss -0.0066

No availability

Full feature set with availability columns removed.

63.7%

+0.99 pts

Δ log loss -0.0081

Full GBM

Current artifacted GBM with rolling, matchup, Elo, and availability features.

64.3%

+1.65 pts

Δ log loss -0.0098

Market reality check

Closing-line comparison keeps the dashboard honest about what the model has and has not beaten.

2025-26

Matched odds

410

33.8% of saved predictions

Model log loss

0.598

Vegas 0.491 on matched games

Agreement

77.3%

Pick agreement with closing-line favorite

Edge thresholdGamesAccuracyROI
3%34568.1%-38.3%
5%30067.3%-40.9%
7%24264.9%-46.7%
10%17963.7%-55.6%

Artifact health

Missing artifacts are surfaced here instead of being guessed or recomputed in the browser.

7 checks

Evaluation report

reports/evaluation_report.md

available

Source for global SHAP table and model narrative.

Updated Jun 30, 6:37 AM · 2,210 bytes

Summary metrics

reports/summary_metrics.csv

available

Artifact-backed model comparison metrics.

Updated Jun 30, 6:37 AM · 230 bytes

GBM backtest

reports/backtest_gbm.csv

available

Walk-forward rows used for calibration and confidence bins.

Updated Jun 30, 6:34 AM · 97,329 bytes

Ablation manifest

reports/ablations/latest/manifest.json

available

Feature-stage experiment provenance.

Updated May 7, 7:45 PM · 1,818 bytes

Market comparison

reports/market_comparison.json

available

Model-versus-closing-line benchmark.

Updated Jun 30, 6:37 AM · 1,502 bytes

Global SHAP parquet

data/processed/shap_global.parquet

available

Optional full global SHAP artifact; markdown fallback is used when absent.

Updated Jun 30, 6:37 AM · 4,256 bytes

Game-level SHAP parquet

data/processed/shap_game_level.parquet

available

Required for per-game directional explanations.

Updated Jun 30, 6:37 AM · 55,265 bytes

Pipeline boundary

The dashboard reads prepared reports and backtests; it does not train models or invent predictions.

Explainability boundary

Global SHAP shows magnitude. Directional per-game reasons require the Parquet explanation artifact.

Evaluation boundary

All headline model quality numbers come from walk-forward rows and saved comparison artifacts.