🔬 HOW THE SLUDGE LINE IS BUILT — AND HOW WELL IT ACTUALLY DOES
Our line is a results-only engine — power ratings built purely from final scores, no market odds. For every sport we post, we hold out recent seasons the model never saw and grade it once. Here are the report cards — pick a sport.
Held out the 2023 & 2024 NFL seasons — games the model never saw — and graded once. Unedited.
10.3
Avg margin error (points)
mkt close ≈ 10.5
65.6%
Winners picked
mkt ≈ 66% · home-pick 55.6%
0.2213
Brier score
coin flip 0.25, lower better
570
Held-out games
never trained on
Calibration — when we say a team wins X%, do they?
We said
Games
Predicted
Observed
0.50-0.55
129
53%
53.5%
0.55-0.60
117
57%
64.1%
0.60-0.65
110
63%
64.5%
0.65-0.70
78
68%
74.4%
0.70-0.75
73
73%
71.2%
0.75-0.80
32
78%
75.0%
0.80-1.00
31
90%
80.6%
Trained on 2019, 2020, 2021, 2022 · holdout 2023 & 2024 graded exactly once · deployed engine parameters (not re-tuned); graded once on the holdout. Regenerated for worker v6.14+: season regression ON and total_shrink 0.35, matching what the engine actually runs for this sport.. Preseason excluded. Params: k_factor=25, carryover=0.6667, base_rating=1505, hfa_elo=40, divisor=28.7, sigma=12.15, ewma_span=12, league_avg_pts=22, season_regression=true, total_shrink=0.35. Generated 2026-08-10T00:21:20.697528Z
Held out the 2023 & 2024 NBA seasons — games the model never saw — and graded once. Unedited.
11.2
Avg margin error (points)
mkt close ≈ 11.5
64.9%
Winners picked
mkt ≈ 68% · home-pick 54.6%
0.214
Brier score
coin flip 0.25, lower better
2640
Held-out games
never trained on
Calibration — when we say a team wins X%, do they?
We said
Games
Predicted
Observed
0.50-0.55
505
53%
52.3%
0.55-0.60
482
57%
53.9%
0.60-0.65
445
63%
64.3%
0.65-0.70
380
68%
63.4%
0.70-0.75
309
73%
72.2%
0.75-0.80
235
78%
80.4%
0.80-1.00
284
90%
88.0%
Trained on 2019, 2020, 2021, 2022 · holdout 2023 & 2024 graded exactly once · deployed engine parameters (not re-tuned); graded once on the holdout. Worker v6.11 (2026-08-08): season regression is now a per-sport setting and this report was generated with it OFF — matching what the engine actually runs for this sport. Prior reports were all generated with regression ON while the worker ran it OFF for every sport, because regressSeason() was never called.. Preseason excluded. Params: k_factor=15, carryover=0.75, base_rating=1505, hfa_elo=80, divisor=32.2, sigma=12.65, ewma_span=15, league_avg_pts=112, season_regression=false. Generated 2026-08-08T05:08:34.236253Z
Held out the 2023 & 2024 NHL seasons — games the model never saw — and graded once. Unedited.
2.15
Avg margin error (goals)
mkt close ≈ 1.9
59.5%
Winners picked
mkt ≈ 59% · home-pick 55.1%
0.236
Brier score
coin flip 0.25, lower better
2793
Held-out games
never trained on
NHL isn't bet on a point spread, so judge it on winners + calibration (margins here are in goals, not points). Hockey is the hardest major sport to predict — matching the market on winners with clean calibration is the real win.
Calibration — when we say a team wins X%, do they?
We said
Games
Predicted
Observed
0.50-0.55
849
53%
52.5%
0.55-0.60
768
57%
56.6%
0.60-0.65
531
63%
61.6%
0.65-0.70
373
68%
66.8%
0.70-0.75
169
73%
69.8%
0.75-0.80
85
78%
82.4%
0.80-1.00
18
90%
88.9%
Trained on 2019, 2020, 2021, 2022 · holdout 2023 & 2024 graded exactly once · deployed engine parameters (not re-tuned); graded once on the holdout. Worker v6.11 (2026-08-08): season regression is now a per-sport setting and this report was generated with it ON — matching what the engine actually runs for this sport. Prior reports were all generated with regression ON while the worker ran it OFF for every sport, because regressSeason() was never called.. Preseason excluded. Params: k_factor=8, carryover=0.75, base_rating=1505, hfa_elo=35, divisor=135, sigma=2.1, ewma_span=20, league_avg_pts=3, season_regression=true. Generated 2026-08-08T05:08:34.290106Z
Held out the 2025 MLB seasons — games the model never saw — and graded once. Unedited.
3.5
Avg margin error (runs)
mkt close ≈ 3
56.7%
Winners picked
mkt ≈ 59% · home-pick 54.3%
0.244
Brier score
coin flip 0.25, lower better
2472
Held-out games
never trained on
MLB isn't bet on a point spread, so judge it on winners + calibration (margins here are in runs, not points). Baseball is high-variance — a steady edge over the baseline is the realistic target.
Calibration — when we say a team wins X%, do they?
We said
Games
Predicted
Observed
0.50-0.55
1744
53%
54.0%
0.55-0.60
584
57%
59.6%
0.60-0.65
140
63%
77.1%
0.65-0.70
4
68%
100.0%
Trained on 2019, 2020, 2021, 2022, 2023, 2024 · holdout 2025 graded exactly once · deployed engine parameters (not re-tuned); graded once on the holdout. Worker v6.11 (2026-08-08): season regression is now a per-sport setting and this report was generated with it OFF — matching what the engine actually runs for this sport. Prior reports were all generated with regression ON while the worker ran it OFF for every sport, because regressSeason() was never called.. Preseason excluded. Params: k_factor=5, carryover=0.6667, base_rating=1505, hfa_elo=25, divisor=170, sigma=3.6, ewma_span=30, league_avg_pts=4.5, season_regression=false. Generated 2026-08-08T05:08:34.401204Z
Beyond the ratings — the live adjustments
The base number is pure power ratings from final scores. On top of that, three results-grounded adjustments fine-tune each game — and every one shows in the line's breakdown, so you can see exactly what moved it.
⚾ Starting pitchers (MLB). When the probable starters post, we shade the number by their ERA — regressed 30% toward league average so a tiny sample doesn't swing it, weighted to about 62% of the game, and clamped so no single arm moves the line more than 1.5 runs. An ace against a struggling spot-starter is a real edge the raw ratings miss.
🥅 Goalies (NHL). ESPN posts the probable starter but no pre-game save percentage, so we learn each goalie's SV% from their own game results (an exponential moving average) and adjust for it, clamped to stay sane. A backup getting the net is worth real goals.
🏥 Injuries. Power ratings can't see that a star is out until results slowly catch up. So a named player can be logged as out with a run/point value and an end date, and the model shades that team on every game until they're back — labeled "injury (auto)" right in the breakdown. We keep it to genuine number-movers (a superstar bat, not a backup). Pitchers aren't logged here; the starter adjustment already prices whoever is actually throwing.
How a pick is graded
There is no wiggle room, and that is the point.
Every pick is logged before the game starts, at the price available in that moment — timestamped, never backdated. After the game it settles W / L / push against the result, and units are booked at the exact price it was logged at. Nothing is deleted, edited, or quietly dropped — winners and losers both stay on the Receipts page permanently.
Beat-the-close is the number that actually matters. The closing line is the sharpest price a game ever has — by then all the money and information are in. So for every pick we also check whether our number beat where the market closed. Win or lose on the night, consistently getting a better number than the close is the one thing variance can't fake. It is the real fingerprint of an edge, and we publish it on every pick.
What this is — and what it isn't
We claim: a transparent, results-only number that is market-competitive in accuracy, adjusted for the things that genuinely move a game, and graded in public on every single pick.
We don't claim: to beat the closing line consistently, to "lock" anything, or to win at some impossible clip. Nobody does that. Any account posting "14-2 in the playoffs" with no timestamped, graded record behind it is showing you a screenshot, not a result — hidden losses, deleted picks, and cherry-picked streaks are the oldest trick in the business.
Our deal is simpler: here is the number, here is how it is built, here is every pick we have made with the date and price, and here is how it did against the close. Check the math yourself. That is the whole pitch.