← All data & metrics Experimental · Frozen source
NFL game forecasts & backtests 11,978 data points · 1,562 fields
Full JSON Download all CSV ↓ Metric formulas & meanings Signal or noise · historical research
What the backtests support These verdicts summarize the frozen 2024–2025 evaluation. They do not promote experimental forecasts into current recommendations. A single correct pick is not proof of a repeatable pattern.
One game
No single-game verdict 80% margin range: ±18.0 points; observed coverage 82.3%
A single upset or correct pick cannot tell us whether a forecast has a lasting edge. Judge frozen predictions over many games.
The range is calibrated from historical errors; it does not capture every roster or injury change.
Across games
Measured pooled edge over Elo Model − Elo Brier -0.0078; 95% block interval -0.0140 to -0.0018 (n=569)
Lower Brier is better. The paired interval stays below zero across the pooled held-out games.
This is a historical model comparison, not proof of an advantage on the next matchup.
Across seasons
Consistency still mixed 1 of 2 seasons have a model-favoring 95% interval versus Elo
2024: -0.0139 to 0.0042 · 2025: -0.0186 to -0.0024
Only the evaluated seasons are represented. More unseen seasons and model-change controls are needed before calling this durable.
Player to next season
New signal not established Candidate MAE 3.103 vs age plus production 3.098 (3009 player-seasons)
The candidate does not beat the strongest existing baseline on this point-error measure.
Inspected historical seasons and excluded rookie or thin-history cohorts limit any claim about next-season skill.
Historical market check
Market ahead on paired subset Model − market Brier 0.0107; 95% block interval 0.0040 to 0.0185
Positive values favor the historical market on games with matched odds.
Historical odds lack archived quote times. This is not evidence of a real-time or tradable edge.
Evidence metadata Frozen artifact d5b0aa30726c · Snapshot Sep 5, 2026, 1:00 AM UTC
Read the compact evidence API
Players, teams & metrics All data points Field dictionary This-page CSV uses field path, type and value columns, including source context. Full dataset downloads are above.
11,978 matches · 351–400 shown. Every field remains available in the full download. Filters are saved in this page’s address.
Recorded data points Field path Value & type $["backtest"]["seasons"]["2024"]["baselines"]["home_rate"]["logLoss"] 0.6892763451490785 number $["backtest"]["seasons"]["2024"]["baselines"]["home_rate"]["n"] 285 number $["backtest"]["seasons"]["2024"]["games"] 285 number $["backtest"]["seasons"]["2024"]["margin"]["intervalCoverage"] 0.8175438596491228 number $["backtest"]["seasons"]["2024"]["margin"]["intervalHalfWidth"] 18.033613498390388 number $["backtest"]["seasons"]["2024"]["margin"]["mae"] 10.138896497235342 number $["backtest"]["seasons"]["2024"]["margin"]["nominalCoverage"] 0.8 number $["backtest"]["seasons"]["2024"]["margin"]["rmse"] 13.098426278932624 number $["backtest"]["seasons"]["2024"]["pairedBrierDifferenceVsElo"]["blocks"] 22 number $["backtest"]["seasons"]["2024"]["pairedBrierDifferenceVsElo"]["estimate"] -0.005196840228314786 number $["backtest"]["seasons"]["2024"]["pairedBrierDifferenceVsElo"]["lower95"] -0.013910539690707008 number $["backtest"]["seasons"]["2024"]["pairedBrierDifferenceVsElo"]["lowerIsBetter"] true boolean $["backtest"]["seasons"]["2024"]["pairedBrierDifferenceVsElo"]["replicates"] 2000 number $["backtest"]["seasons"]["2024"]["pairedBrierDifferenceVsElo"]["unit"] season-week string $["backtest"]["seasons"]["2024"]["pairedBrierDifferenceVsElo"]["upper95"] 0.0042385314513881875 number $["backtest"]["seasons"]["2024"]["probability"]["accuracy"] 0.6736842105263158 number $["backtest"]["seasons"]["2024"]["probability"]["brier"] 0.21403841522881903 number $["backtest"]["seasons"]["2024"]["probability"]["calibration"][0]["lower"] 0.1 number $["backtest"]["seasons"]["2024"]["probability"]["calibration"][0]["n"] 3 number $["backtest"]["seasons"]["2024"]["probability"]["calibration"][0]["observed"] 0.3333333333333333 number $["backtest"]["seasons"]["2024"]["probability"]["calibration"][0]["predicted"] 0.16070561349825305 number $["backtest"]["seasons"]["2024"]["probability"]["calibration"][0]["upper"] 0.2 number $["backtest"]["seasons"]["2024"]["probability"]["calibration"][1]["lower"] 0.2 number $["backtest"]["seasons"]["2024"]["probability"]["calibration"][1]["n"] 8 number $["backtest"]["seasons"]["2024"]["probability"]["calibration"][1]["observed"] 0.25 number $["backtest"]["seasons"]["2024"]["probability"]["calibration"][1]["predicted"] 0.25743165065753787 number $["backtest"]["seasons"]["2024"]["probability"]["calibration"][1]["upper"] 0.3 number $["backtest"]["seasons"]["2024"]["probability"]["calibration"][2]["lower"] 0.3 number $["backtest"]["seasons"]["2024"]["probability"]["calibration"][2]["n"] 39 number $["backtest"]["seasons"]["2024"]["probability"]["calibration"][2]["observed"] 0.3333333333333333 number $["backtest"]["seasons"]["2024"]["probability"]["calibration"][2]["predicted"] 0.3575573440720387 number $["backtest"]["seasons"]["2024"]["probability"]["calibration"][2]["upper"] 0.4 number $["backtest"]["seasons"]["2024"]["probability"]["calibration"][3]["lower"] 0.4 number $["backtest"]["seasons"]["2024"]["probability"]["calibration"][3]["n"] 56 number $["backtest"]["seasons"]["2024"]["probability"]["calibration"][3]["observed"] 0.3392857142857143 number $["backtest"]["seasons"]["2024"]["probability"]["calibration"][3]["predicted"] 0.447682025861564 number $["backtest"]["seasons"]["2024"]["probability"]["calibration"][3]["upper"] 0.5 number $["backtest"]["seasons"]["2024"]["probability"]["calibration"][4]["lower"] 0.5 number $["backtest"]["seasons"]["2024"]["probability"]["calibration"][4]["n"] 56 number $["backtest"]["seasons"]["2024"]["probability"]["calibration"][4]["observed"] 0.5178571428571429 number $["backtest"]["seasons"]["2024"]["probability"]["calibration"][4]["predicted"] 0.5498449055190057 number $["backtest"]["seasons"]["2024"]["probability"]["calibration"][4]["upper"] 0.6 number $["backtest"]["seasons"]["2024"]["probability"]["calibration"][5]["lower"] 0.6 number $["backtest"]["seasons"]["2024"]["probability"]["calibration"][5]["n"] 57 number $["backtest"]["seasons"]["2024"]["probability"]["calibration"][5]["observed"] 0.7017543859649122 number $["backtest"]["seasons"]["2024"]["probability"]["calibration"][5]["predicted"] 0.6415866237864815 number $["backtest"]["seasons"]["2024"]["probability"]["calibration"][5]["upper"] 0.7 number $["backtest"]["seasons"]["2024"]["probability"]["calibration"][6]["lower"] 0.7 number $["backtest"]["seasons"]["2024"]["probability"]["calibration"][6]["n"] 45 number $["backtest"]["seasons"]["2024"]["probability"]["calibration"][6]["observed"] 0.7333333333333333 number
← Previous Page 8 of 240 Next → Dataset details Source date: Sep 5, 2026, 1:00 AM UTC
272 scheduled games, 32 team outlooks, probability calibration, score errors and player-production experiments. Frozen preseason research; does not change public rankings.
Original capture date; opening or downloading does not refresh the source. Per-record timestamps take precedence. Definitions may describe unavailable inputs.
How to read and verify this data JSON bracket paths; [] means every array element. CSV uses concrete zero-based indices. Types are observed, not a promise that future fields cannot change. null is unavailable; empty array/object means an empty collection. Zero is preserved as a measured or supplied numeric value. Source-native units and scoring. Consult metric-definitions for formulas; ambiguous units remain unspecified, never inferred.
sha256 of UTF-8 JSON.stringify(data), excluding this envelope. Source receipt hashes remain inside data. CSV prefixes formula-looking text with an apostrophe; numeric signs are preserved. JSON keeps original strings.
ecbde32e2751aef21c0bf12e3dfcd6ef10021f583afb03d6db2f3bce002f9316
Published observations, approved source-labeled aggregates, research and definitions are provided within each dataset's stated scope. Read each record's source, period, scoring and availability. Research visibility does not promote a model into public rankings. Missing means unavailable, never zero. No account records or provider write access.