Daily learning journal

• On October 3, 2026, the model evaluated 151 samples across 21 games with an overall accuracy of 74.83%. The shipped probabilities achieved a Brier score of 0.2020 compared to the baseline raw score of 0.2011, where lower is better and 0.25 represents random chance. The model adopted calibration calibrators for three batter_contact market variants (league:mlb, baseball, and global *), each improving hold-out Brier scores from 0.2198 to 0.2027 across 2,946 samples. Expected Calibration Error was 0.1184 for the model versus 0.0905 for raw predictions, indicating the model's probability bins deviated more from observed frequencies than uncalibrated outputs.
• Graded 151 markets from yesterday. Shipped probabilities scored Brier 0.2020 (raw 0.2011; a coin flip is 0.25).
• Adopted a calibration calibrator for league:mlb:batter_contact (hold-out Brier 0.2198 → 0.2027, 2946 samples).
• Adopted a calibration calibrator for baseball:batter_contact (hold-out Brier 0.2198 → 0.2027, 2946 samples).
• Adopted a calibration calibrator for *:batter_contact (hold-out Brier 0.2198 → 0.2027, 2946 samples).

Too few finished games to grade; coefficients were left unchanged.

• The model evaluated 217 markets across 15 games on 2026-09-22, achieving a Brier score of 0.1949 compared to 0.2025 for raw output and 0.2500 for random guessing. Calibration error (ECE) was 0.0430 for the model versus 0.0860 for raw predictions, indicating improved reliability. MLB calibration adjustments reduced Brier by 0.0119 across 173 samples, while CS2 calibration lagged raw output by 0.0090 across 44 samples and remains conservative pending additional data. The refit process analyzed 12,043 historical samples in 64,751 milliseconds with no coefficient updates applied.
• Graded 217 markets from yesterday. Shipped probabilities scored Brier 0.1949 (raw 0.2025; a coin flip is 0.25).
• MLB: learned calibration cut Brier by 0.0119 versus raw output (173 samples).
• CS2: calibration trailed raw output by 0.0090; coefficients stay conservative until more samples land.

• Graded 1773 markets from yesterday. Shipped probabilities scored Brier 0.2033 (raw 0.2059; a coin flip is 0.25).
• MLB: learned calibration cut Brier by 0.0037 versus raw output (1508 samples).
• batter_rbi_run markets ran 3.1%p low versus outcomes (400 samples, 35% actual). The refit shifts them up.
• batter_bb markets ran 3.3%p high versus outcomes (200 samples, 25% actual). The refit shifts them down.
• J League: calibration trailed raw output by 0.0058; coefficients stay conservative until more samples land.
• team_totals markets ran 9.1%p high versus outcomes (47 samples, 38% actual). The refit shifts them down.
• CS2: calibration trailed raw output by 0.0067; coefficients stay conservative until more samples land.
• epl, jleague crossed the sample threshold and joined the learning set automatically.

Too few finished games to grade; coefficients were left unchanged.

Too few finished games to grade; coefficients were left unchanged.

• seriea dried up on recent samples and left the learning set.

• bundesliga, lol dried up on recent samples and left the learning set.

Too few finished games to grade; coefficients were left unchanged.

Too few finished games to grade; coefficients were left unchanged.

• laliga, ligue1, leaguescup dried up on recent samples and left the learning set.

Too few finished games to grade; coefficients were left unchanged.

• Graded 524 markets from yesterday. Shipped probabilities scored Brier 0.1857 (raw 0.1990; a coin flip is 0.25).
• Adopted a calibration calibrator for esports:team_totals (hold-out Brier 0.2784 → 0.2422, 435 samples).
• Adopted a calibration calibrator for league:lol:ml (hold-out Brier 0.2177 → 0.1535, 182 samples).
• Adopted a beta calibrator for league:cs2:team_totals (hold-out Brier 0.2734 → 0.2431, 344 samples).
• MLB: learned calibration cut Brier by 0.0172 versus raw output (415 samples).
• batter_contact markets ran 13.8%p high versus outcomes (110 samples, 25% actual). The refit shifts them down.
• batter_rbi_run markets ran 12.1%p high versus outcomes (110 samples, 19% actual). The refit shifts them down.
• batter_power markets ran 8.9%p high versus outcomes (110 samples, 13% actual). The refit shifts them down.
• 8 stacking meta-models passed validation and were swapped in.

• league:mlb:batter_rbi_run calibration trailed raw output on live games (Brier 0.2470 vs 0.2409, 326 samples). Pulled and quarantined.
• Graded 1444 markets from yesterday. Shipped probabilities scored Brier 0.2102 (raw 0.2138; a coin flip is 0.25).
• Adopted a calibration calibrator for esports:team_totals (hold-out Brier 0.2847 → 0.2409, 461 samples).
• Adopted a calibration calibrator for league:lol:ml (hold-out Brier 0.2167 → 0.1428, 236 samples).
• Adopted a calibration calibrator for league:cs2:team_totals (hold-out Brier 0.2752 → 0.2442, 343 samples).
• MLB: learned calibration cut Brier by 0.0035 versus raw output (1231 samples).
• batter_rbi_run markets ran 6.1%p low versus outcomes (326 samples, 38% actual). The refit shifts them up.
• batter_contact markets ran 5.4%p high versus outcomes (326 samples, 37% actual). The refit shifts them down.
• CS2: learned calibration cut Brier by 0.0097 versus raw output (96 samples).
• team_totals markets ran 13.2%p low versus outcomes (54 samples, 57% actual). The refit shifts them up.
• 12 stacking meta-models passed validation and were swapped in.
• soccer_goals markets ran 3.8%p low versus outcomes (90 samples, 57% actual). The refit shifts them up.

• esports:ml calibration trailed raw output on live games (Brier 0.2417 vs 0.2304, 48 samples). Pulled and quarantined.
• Graded 1334 markets from yesterday. Shipped probabilities scored Brier 0.2002 (raw 0.2078; a coin flip is 0.25).
• Adopted a calibration calibrator for league:lol:ml (hold-out Brier 0.2167 → 0.1445, 288 samples).
• Adopted a calibration calibrator for esports:team_totals (hold-out Brier 0.2802 → 0.2426, 486 samples).
• MLB: learned calibration cut Brier by 0.0087 versus raw output (1238 samples).
• Adopted a calibration calibrator for league:cs2:team_totals (hold-out Brier 0.2738 → 0.2447, 342 samples).
• batter_contact markets ran 5.2%p high versus outcomes (328 samples, 35% actual). The refit shifts them down.
• CS2: calibration trailed raw output by 0.0061; coefficients stay conservative until more samples land.
• 8 stacking meta-models passed validation and were swapped in.

• baseball:batter_bb calibration trailed raw output on live games (Brier 0.2135 vs 0.2065, 119 samples). Pulled and quarantined.
• league:mlb:batter_power calibration trailed raw output on live games (Brier 0.1604 vs 0.1552, 238 samples). Pulled and quarantined.
• Graded 1038 markets from yesterday. Shipped probabilities scored Brier 0.2043 (raw 0.2051; a coin flip is 0.25).
• Adopted a calibration calibrator for esports:team_totals (hold-out Brier 0.2912 → 0.2443, 501 samples).
• Adopted a beta calibrator for league:lol:ml (hold-out Brier 0.2178 → 0.1459, 320 samples).
• Adopted a calibration calibrator for league:lol:team_totals (hold-out Brier 0.3130 → 0.2301, 160 samples).
• CS2: learned calibration cut Brier by 0.0106 versus raw output (96 samples).
• batter_rbi_run markets ran 3.3%p low versus outcomes (238 samples, 32% actual). The refit shifts them up.
• team_totals markets ran 12.3%p low versus outcomes (47 samples, 55% actual). The refit shifts them up.
• 10 stacking meta-models passed validation and were swapped in.
• totals markets ran 7.3%p high versus outcomes (47 samples, 38% actual). The refit shifts them down.
• soccer_goals markets ran 9.0%p high versus outcomes (30 samples, 40% actual). The refit shifts them down.

• league:cs2:ml calibration trailed raw output on live games (Brier 0.2782 vs 0.2529, 48 samples). Pulled and quarantined.
• Graded 1458 markets from yesterday. Shipped probabilities scored Brier 0.1993 (raw 0.2020; a coin flip is 0.25).
• Adopted a calibration calibrator for league:lol:ml (hold-out Brier 0.2191 → 0.1403, 344 samples).
• Adopted a calibration calibrator for esports:team_totals (hold-out Brier 0.2900 → 0.2445, 489 samples).
• Adopted a calibration calibrator for league:lol:team_totals (hold-out Brier 0.3078 → 0.2327, 172 samples).
• MLB: learned calibration cut Brier by 0.0037 versus raw output (1224 samples).
• soccer_goals markets ran 12.8%p low versus outcomes (100 samples, 63% actual). The refit shifts them up.
• batter_power markets ran 3.1%p low versus outcomes (324 samples, 24% actual). The refit shifts them up.
• batter_bb markets ran 4.8%p high versus outcomes (162 samples, 24% actual). The refit shifts them down.
• CS2: calibration trailed raw output by 0.0092; coefficients stay conservative until more samples land.
• 9 stacking meta-models passed validation and were swapped in.
• totals markets ran 5.7%p high versus outcomes (52 samples, 40% actual). The refit shifts them down.
• team_totals markets ran 3.4%p low versus outcomes (52 samples, 46% actual). The refit shifts them up.
• wnba dried up on recent samples and left the learning set.

• *:soccer_goals calibration trailed raw output on live games (Brier 0.2279 vs 0.2145, 230 samples). Pulled and quarantined.
• *:batter_power calibration trailed raw output on live games (Brier 0.1673 vs 0.1670, 326 samples). Pulled and quarantined.
• Graded 1648 markets from yesterday. Shipped probabilities scored Brier 0.2099 (raw 0.2108; a coin flip is 0.25).
• Adopted a calibration calibrator for league:lol:ml (hold-out Brier 0.2174 → 0.1302, 358 samples).
• Adopted a calibration calibrator for esports:team_totals (hold-out Brier 0.2916 → 0.2441, 504 samples).
• Adopted a calibration calibrator for league:lol:team_totals (hold-out Brier 0.3063 → 0.2392, 179 samples).
• MLB: learned calibration cut Brier by 0.0037 versus raw output (1231 samples).
• soccer_goals markets ran 10.4%p low versus outcomes (230 samples, 60% actual). The refit shifts them up.
• batter_contact markets ran 5.4%p high versus outcomes (326 samples, 36% actual). The refit shifts them down.
• MLS: calibration trailed raw output by 0.0157; coefficients stay conservative until more samples land.
• team_totals markets ran 11.7%p low versus outcomes (58 samples, 55% actual). The refit shifts them up.
• 8 stacking meta-models passed validation and were swapped in.
• Bundesliga: calibration trailed raw output by 0.0028; coefficients stay conservative until more samples land.
• ligue1 crossed the sample threshold and joined the learning set automatically.

• The model graded 939 samples from 30 games on September 5, 2026, achieving a Brier score of 0.2012 versus 0.2051 for raw outputs. Two market families—baseball batter_power and soccer soccer_goals—were quarantined after calibration underperformed raw output in live evaluation. MLB showed the strongest calibration improvement (0.0044 Brier reduction across 835 samples), while 14 stacking meta-models and 3 new calibration calibrators were promoted after passing validation. Bundesliga and Serie A were added to the learning set after crossing sample thresholds.
• soccer:soccer_goals calibration trailed raw output on live games (Brier 0.2104 vs 0.1938, 40 samples). Pulled and quarantined.
• baseball:batter_power calibration trailed raw output on live games (Brier 0.1781 vs 0.1723, 229 samples). Pulled and quarantined.
• Graded 939 markets from yesterday. Shipped probabilities scored Brier 0.2012 (raw 0.2051; a coin flip is 0.25).
• Adopted a calibration calibrator for league:lol:ml (hold-out Brier 0.2187 → 0.1307, 370 samples).
• Adopted a calibration calibrator for esports:team_totals (hold-out Brier 0.2884 → 0.2454, 513 samples).
• Adopted a calibration calibrator for league:lol:team_totals (hold-out Brier 0.3298 → 0.2388, 185 samples).
• MLB: learned calibration cut Brier by 0.0044 versus raw output (835 samples).
• batter_power markets ran 5.3%p low versus outcomes (229 samples, 27% actual). The refit shifts them up.
• batter_contact markets ran 4.7%p high versus outcomes (219 samples, 39% actual). The refit shifts them down.
• 14 stacking meta-models passed validation and were swapped in.
• soccer_goals markets ran 5.7%p low versus outcomes (60 samples, 50% actual). The refit shifts them up.
• team_totals markets ran 7.6%p low versus outcomes (25 samples, 56% actual). The refit shifts them up.
• bundesliga, seriea crossed the sample threshold and joined the learning set automatically.

• league:mlb:batter_bb calibration trailed raw output on live games (Brier 0.2018 vs 0.1990, 96 samples). Pulled and quarantined.
• Graded 799 markets from yesterday. Shipped probabilities scored Brier 0.1898 (raw 0.2017; a coin flip is 0.25).
• Adopted a beta calibrator for league:lol:ml (hold-out Brier 0.2190 → 0.1310, 372 samples).
• Adopted a calibration calibrator for esports:team_totals (hold-out Brier 0.2903 → 0.2463, 509 samples).
• Adopted a calibration calibrator for league:lol:team_totals (hold-out Brier 0.3235 → 0.2372, 186 samples).
• MLB: learned calibration cut Brier by 0.0112 versus raw output (726 samples).
• batter_contact markets ran 8.1%p high versus outcomes (192 samples, 32% actual). The refit shifts them down.
• batter_rbi_run markets ran 5.3%p high versus outcomes (192 samples, 26% actual). The refit shifts them down.
• soccer_goals markets ran 13.1%p high versus outcomes (50 samples, 36% actual). The refit shifts them down.
• 12 stacking meta-models passed validation and were swapped in.

• Graded 1384 markets from yesterday. Shipped probabilities scored Brier 0.2077 (raw 0.2084; a coin flip is 0.25).
• Adopted a calibration calibrator for league:lol:ml (hold-out Brier 0.2223 → 0.1395, 418 samples).
• Adopted a calibration calibrator for league:lol:team_totals (hold-out Brier 0.3246 → 0.2470, 209 samples).
• Adopted a calibration calibrator for league:cs2:team_totals (hold-out Brier 0.2914 → 0.2465, 325 samples).
• MLB: learned calibration cut Brier by 0.0031 versus raw output (1245 samples).
• CS2: calibration trailed raw output by 0.0270; coefficients stay conservative until more samples land.
• batter_rbi_run markets ran 4.4%p low versus outcomes (330 samples, 34% actual). The refit shifts them up.
• batter_bb markets ran 4.4%p high versus outcomes (165 samples, 26% actual). The refit shifts them down.
• 11 stacking meta-models passed validation and were swapped in.
• soccer_goals markets ran 12.5%p low versus outcomes (30 samples, 63% actual). The refit shifts them up.
• team_totals markets ran 5.9%p low versus outcomes (55 samples, 51% actual). The refit shifts them up.
• totals markets ran 3.4%p low versus outcomes (55 samples, 47% actual). The refit shifts them up.