Week 5 Report: A Better-Than-Baseline Week, With Some Big Surprises

Week 5 included 15 completed games. One was a tie: NYG @ WAS finished 27-27. Ties are excluded from binary Brier scoring and from the favorite win/loss record, leaving 14 games analyzed.

The prediction set had a 9-5 favorite win/loss record. The tied game is not counted in either category. Historical binary calibration training also excludes ties.

Overall prediction accuracy

The average Brier score was 0.2236. Lower Brier scores are better. The 50/50 baseline was 0.2500, so this week's average was lower by 0.0264.

The strongest prediction was PHI @ JAX. The model gave JAX a 0.765 probability, and JAX won 19-17. That prediction had a Brier score of 0.0552.

The biggest miss was TB @ DAL. Dallas was given a 0.805 probability, but Tampa Bay won 24-16. That produced a Brier score of 0.6480.

Other notable results included correct predictions of Cincinnati over Miami, Houston over Tennessee, Detroit over Arizona, and Denver over the Los Angeles Chargers. The model missed Chicago over Green Bay, San Francisco over Seattle, Baltimore over Atlanta, and Buffalo over the Los Angeles Rams.

The full game-level Brier scores were:

• TB @ DAL: 0.6480
• PHI @ JAX: 0.0552
• CIN @ MIA: 0.0625
• LV @ NE: 0.1406
• MIN @ NO: 0.1980
• CLE @ NYJ: 0.1892
• IND @ PIT: 0.1640
• HOU @ TEN: 0.0702
• CHI @ GB: 0.2652
• DEN @ LAC: 0.1332
• DET @ ARI: 0.0930
• SF @ SEA: 0.3422
• BAL @ ATL: 0.3906
• BUF @ LA: 0.3782

Fictional Walk-Forward Quarter Kelly

Quarter Kelly was in WALK_FORWARD_HISTORICAL status this week. It began with a bankroll of $160.00 and ended with $153.44.

The model placed 4 bets, winning 1 and losing 3. The total amount wagered was $15.01. Weekly profit/loss was -$6.56, for a reported return of -4.1%. There were no tie results among these bets.

Weeks 3 and 4 were training-only weeks. No bets were intended during those weeks, and the bankroll remained at $160.00. Week 5 was the first test week; this is not a retrospective simulation for Weeks 3 and 4.

From Week 5 onward, the fixed walk-forward model uses only earlier completed weeks. It never uses later market prices or outcomes from the game being evaluated. For each prior game's 24-hour snapshots, it groups the two team probabilities into ten-percentage-point bins, calculates the sum of historical outcome-minus-probability residuals in each bin, divides that amount by 40 plus the observation count, and adds the resulting shrunk adjustment to the current game's 24-hour midpoint probability.

A team's edge must be at least five percentage points to qualify. The strategy then uses quarter Kelly with a 10% bankroll cap. Missing 24-hour snapshots cause a game to be skipped.

This week, the eligibility diagnostics found:

• 15 games examined
• 0 missing 24-hour snapshots
• 11 games below the edge threshold
• 4 games that passed the threshold
• 0 zero-sized wagers

The four wagers were on GB, CLE, IND, and NO. The GB wager was the only winner. The model used 62 training observations from the earlier weeks.

This is a small-sample, fictional, walk-forward evaluation. Midpoint prices may not be executable, and fees are excluded. The results do not establish predictive skill or profitability, and no real money was wagered.

Fictional Bet the House benchmark

Bet the House is a deliberately naive fictional benchmark. The experiment began with $10 per game in Week 3. At each stage, the entire carried-forward weekly bankroll is divided equally across all eligible market favorites, using the final permanent snapshot midpoint as the assumed purchase price. Fees are not included.

This week's benchmark started with $100.66. It placed 15 hypothetical wagers, with 9 wins, 5 losses, and 1 tie. The amount wagered was $100.66. Weekly profit/loss was -$2.17, leaving an ending bankroll of $98.49 and a reported return of -2.16%.

The benchmark's notable winning wagers included:

• MIN over NO: +$5.38
• CLE @ NYJ, on NYJ: +$5.17
• IND @ PIT, on PIT: +$4.57
• LV @ NE, on NE: +$4.03

Its notable losing wagers included:

• CHI @ GB, on CHI: -$6.71
• SF @ SEA, on SEA: -$6.71
• BUF @ LA, on LA: -$6.71
• TB @ DAL, on DAL: -$6.71
• BAL @ ATL, on ATL: -$6.72

The tie wager was handled separately. In NYG @ WAS, the contract settled at $0.50 because the game ended 27-27. The fictional benchmark wagered $6.71 on WAS at an assumed price of 0.625, producing a partial loss of -$1.34 under the stated tie-wager rule. It is not counted as either a benchmark win or loss.

Bet the House is fictional, no real money is wagered, and it is not betting advice. Results are reported whether good or bad, and the rules are not changed retroactively.

What We Learned This Week

The prediction set finished better than the 50/50 Brier baseline, with an average score of 0.2236 versus 0.2500. The result included one excellent prediction on JAX and one very costly miss on Dallas.

The favorite record was solid at 9-5, but the tie between the Giants and Washington required separate treatment. It was excluded from binary scoring and favorite results, while the fictional benchmark's tie wager produced a partial loss.

Quarter Kelly made its first test-week wagers and finished down $6.56, while Bet the House finished down $2.17. Both remain fictional evaluations with limited data, midpoint-price assumptions, and no real money involved.