Model Transparency • Kos Edge Analytics

NFL — held-out 2025 season (never used to tune the model)

Blend weights were selected on 2024 games only, then locked and applied unchanged to 285 games in 2025. This is the honest test: performance on games the tuning process never saw.

Spread error (MAE, points)

8.353model

9.670consensus close

13.6% better than market

Total error (MAE, points)

9.461model

10.423consensus close

9.2% better than market

Win probability calibration (Brier)

0.1636model

0.2112consensus close

22.5% better than market

Statistical significance (spread, paired bootstrap, 5,000 resamples)

Model beats the market by 1.32 points on average, 95% CI [-2.03, -0.62] — statistically significant, not noise.

Full 13-season sample (2013–2025, 3,562 games)

Unblended heuristic simulation only (no supervised ML, no live market blend) -- shows the base model already had a slight edge before the ML layer was added.

Spread error (MAE, points)

9.622model

9.918consensus close

3.0% better than market

Total error (MAE, points)

10.283model

10.505consensus close

2.1% better than market

Methodology & honest caveats

Supervised gradient-boosted model trained on 2013-2023 NFL games, blended with a Monte Carlo matchup simulation and live market lines. Blend weights were grid-searched on the 2024 season only, then applied unchanged to the 2025 season, which the tuning process never saw.

This is a backtest on completed historical games using nflverse's free closing lines (spread_line/total_line), not a live prospective track record yet -- live CLV tracking requires a working Odds API key and several weeks of accumulated open-to-close snapshots, which is in progress.
The validated blend weights only apply once a season's rolling features contain real, week-varying data. Before games are played (e.g. the 2026 season as of this report), the model automatically falls back to more conservative, lower-trust blend weights -- see detect_real_rolling_features in nfl_supervised_retrain.py.

Report generated 2026-07-17 · model nfl-v1.5-matchup-sim · source: scripts/nfl/historical_market_backtest.py, scripts/nfl/tune_blend_weights.py, scripts/nfl/full_pipeline_holdout_backtest.py

Other sports & live tracking

Live closing-line-value tracking (open vs. close) lives on CLV Tracker. It is a live sample as the season plays out — not a substitute for the held-out check above.

Product map

Core contract

How KosEdge model, handicap, and product surfaces work — in one place.

Model: Research fair — a pure sim / strength snapshot. It does not drive PLAY or LEAN on its own.
KEI: Final handicap. Model plus late information and reprice — desk factors such as injury, rest, weather, confirmation.
Edge / Tag: KEI versus the best available market only — never pure Model versus market for PLAY or LEAN.
Estimates: Numbers are research estimates, not instructions to bet and not a profitability promise.

What we show vs what we don't

We show

Honest empty, preseason, markets-only, and engine-warming states — no fake rows to fill a board.
Thin sample and low-depth runs labeled as such.
Fantasy: Model rank is projection order, not recommended pick order. ADP-aware advice lives on Builder and Mock.

We don't

We do not invent KEI, copy books into KEI cells, or imply CLV from a short live sample.
We do not sell locks, pick-service instructions, or a sure thing.

Surface guide

What each board is — and is not. Jump:

Desk status (PRESEASON / data stale)

PRESEASON / production no-go: PLAY stake tags and survivor locks stay research-only until readiness is go.
data stale: boards may use the last owned snapshot; do not treat PLAY tags as live stakes until freshness recovers.
Freshness and readiness probes still gate PLAY tags. Policy lives here — not on product boards.
CLV Tracker and Performance remain the live / held-out accountability surfaces.

Model Health & Governance

NFL — held-out 2025 season (never used to tune the model)

Full 13-season sample (2013–2025, 3,562 games)

Methodology & honest caveats

Other sports & live tracking

Core contract

What we show vs what we don't

Surface guide

Desk status (PRESEASON / data stale)

Edge Board

KEI Lines

Weekly slate / matchups

Survivor

Fantasy

Pick’em

Game Boxes / Season Model

Power Ratings

Camp Desk / Injuries & News

Insights / Doctrine

Props

Model Health & Governance

NFL — held-out 2025 season (never used to tune the model)

Full 13-season sample (2013–2025, 3,562 games)

Methodology & honest caveats

Other sports & live tracking

Core contract

What we show vs what we don't

Surface guide

Desk status (PRESEASON / data stale)

Edge Board

KEI Lines

Weekly slate / matchups

Survivor

Fantasy

Pick’em

Game Boxes / Season Model

Power Ratings

Camp Desk / Injuries & News

Insights / Doctrine

Props