Market vs Reality [Football Season]

← All reports
2026 Season · Week 4

Week 4 Report: A Better Week for the Market Forecasts

The Week 4 market forecasts covered 16 games. The market’s favorite won 9 games and lost 7.

The average Brier score was 0.2172. Lower Brier scores are better, because they reward probabilities that are both accurate and well calibrated.

The 50/50 baseline Brier score is 0.2500. This week’s result was 0.0328 lower than that baseline, so the market performed better than simply assigning every team a 50% chance.

Best and Biggest Miss

The best prediction was TEN @ BAL. Baltimore won 24–18, and the market gave BAL an 0.855 probability. The Brier score was 0.0210.

The biggest miss was NE @ BUF. New England won 29–26, despite Buffalo receiving a 0.735 probability. That game’s Brier score was 0.5476.

There were several other strong predictions. The market correctly picked MIN over MIA with MIN at 0.825, SEA over LAC at 0.755, and KC over LV at 0.685.

The market also missed several favorites, including PIT against CLE, CIN against JAX, HOU against DAL, ARI against NYG, DET against CAR, and NO against ATL. The favorite’s overall record was 9–7.

Quarter Kelly: Still in Training

The fictional Quarter Kelly model is listed as TRAINING_ONLY for Week 4. Weeks 3 and 4 are training-only weeks: no bets were intended yet, so the model placed no bets and did not change its bankroll.

The eligibility diagnostics also show zero games examined, zero missing 24-hour snapshots, zero below-threshold games, zero games passing the edge threshold, zero zero-sized wagers, and zero training observations for this week. Because the status is training-only, these games were not rejected for failing a betting threshold.

Week 5 is the first test week. From Week 5 onward, the model uses only earlier completed weeks. It never uses later market prices or outcomes from the game being evaluated.

The fixed walk-forward model takes each prior game’s 24-hour snapshots and groups the two team probabilities into ten-percentage-point bins. For each bin, it sums the historical outcome-minus-probability residuals, divides that sum by 40 plus the count of observations, and adds the resulting shrunk adjustment to the current game’s 24-hour midpoint probability.

A team must have at least a five-percentage-point edge to qualify. The model then uses quarter Kelly, with wagers capped at 10% of the bankroll. A game missing a 24-hour snapshot is skipped.

This is a small-sample, fictional walk-forward evaluation. Midpoint prices may not be executable, and fees are excluded. No real money is wagered, and the evaluation does not promise profitability.

Bet the House Benchmark

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

For Week 4, the benchmark reported:

Notable winning wagers included IND over WAS, with a reported profit of $3.84, GB over TB at $5.41, and SF over DEN at $5.19. The benchmark’s losing wagers included BUF against NE, CIN against JAX, and PIT against CLE, each reported at -$7.62. DET against CAR was reported at -$7.69.

The benchmark is fictional. No real money is wagered, it is not betting advice, and results are reported whether they are good or bad. The rules are not changed retroactively.

What We Learned This Week

The market had a solid Week 4 by its Brier score, finishing at 0.2172, or 0.0328 lower than the 0.2500 50/50 baseline. It correctly identified several strong favorites, but it also suffered some costly misses, including Buffalo, Cincinnati, and Detroit.

The Quarter Kelly system remains in its training period, so it has no Week 4 betting results to evaluate. The Bet the House benchmark, meanwhile, had a 9–7 record but finished the week down $21.33, ending at $100.66.

← Week 3 All Reports