Methodology

V10 is a logistic regression model trained on 7,315MLB regular-season and playoff games from 2023 through 2026. Its inputs are each team's cumulative run differential, bullpen SIERA and WAR, a park-factor adjustment, and flags for the postseason and late-series games. Logistic regression was chosen deliberately: it produces well-calibrated probabilities, is interpretable, and resists overfitting on a dataset this size. More complex models — neural nets, gradient boosting — showed no statistically significant improvement on held-out data.

The model outputs a win probability for the home team. That probability is compared against the market-implied probability derived from the current moneyline to compute edge. When V10 sees meaningfully more probability than the market, there is a signal worth tracking. When the model and market agree, it is a pass.

The regression's only inputs are team-level: park-adjusted cumulative run differential for both clubs, bullpen SIERA and WAR, and flags for whether the game is in the postseason or late in a series.

Starting pitcher quality is shown on every card and weighted by innings-pitched sample size — starters under 30 innings are flagged as thin, and a TBD starter gets zero weight — but it feeds the tier and confidence rating, not the win-probability model itself. A late scratch triggers an automatic re-grade of that game's pick.

Line movement from open to current is read the same way: a confirmation signal, not a model input. Movement toward the model's side raises confidence; movement against it is treated as a warning and can downgrade a pick to a lean or a pass.

V10 achieves 55.8% honest out-of-sample accuracy — measured on 2026 games it never saw in training, not back-tested. The naive benchmark is 52.6% (picking the home team every game), so the model beats that baseline by 3.2 points.

That baseline is not the market, and the honest distinction matters. Against the market itself, V10 performs at roughly the level of the closing line: when it disagrees with Vegas it is right about as often as it is wrong, and our published closing-line-value data shows it does not consistently beat the close. This is expected — betting markets are highly efficient, and matching them is already hard.

So linerup claims no profitable edge over sportsbooks. The value is calibration and transparency: a well-built model whose every pick is published before first pitch and tracked in the open, wins and misses alike. The record shows exactly how it performs against reality — we let it speak for itself.

V7 is a logistic regression model trained on 6,385 NBA games. It achieves 68.6% honest out-of-sample accuracy against a 55.7% baseline — a 12.9 percentage-point edge over what picking the home team every game would yield.

The model outputs a win probability for the home team. Edge is computed by comparing this against the market-implied probability from the moneyline. When V7 finds meaningful edge, it produces a pick. When the model and market agree, it passes.

The signature innovation in V7 is structured injury detection. The model tracks top-8 rotation players by minutes and flags missing players before each game. A team missing one or two key rotation players has a measurably different win probability than the market often reflects — especially in the playoffs when public attention inflates the perceived strength of depleted rosters.

Additional inputs include point differential trends, pace adjustments, line movement, sharp signal, and home court advantage. Spread movement from open to current is used as a proxy for sharp money flow, same as in the MLB model.

V7's 68.6% OOS accuracy on 6,385 games represents a 12.9pp edge over the 55.7% baseline (home team win rate over the training period) — a large edge by modeling standards, driven primarily by the injury detection feature. As with V10, that edge is over the naive baseline, not the market: linerup makes no claim of beating sportsbook pricing. The final 2025-26 record is shown openly.