Decision tools for bettors and fantasy players.

We provide the scientific foundation for the SportsLabs pricing engine. Using football-data.co.uk odds for six European leagues over thirteen seasons (30,632 matches), we test six pre-registered hypotheses. (i) Sharp-book closing odds are close to calibrated; the power and Shin de-vig methods reduce the favourite–longshot bias left by multiplicative normalisation (calibration slope 1.013 and 1.023 vs. 1.049, the last z=3.46), but log-loss differences between methods are economically tiny (< 1.5 × 10⁻⁴ nats) and not significant. (ii) Closing prices beat pre-closing prices in every season (log-loss gain 0.0024, t=6.0). (iii) A consensus-mispricing strategy in the spirit of Kaunitz et al. (2017) is profitable only when the consensus is the sharp book: betting when the best market price exceeds de-vigged sharp-book pre-closing fair odds by ≥ 2% yields 11,937 bets, ROI +4.7% (95% cluster-bootstrap CI [1.6, 8.0]), mean closing-line value (CLV) +3.4%, whereas the same rule built on the multiplicatively de-vigged market average loses money (−2.7%). No configuration survives a Holm correction across 75 tests (minimum adjusted p=0.15), realised ROI decays after 2018 and is not significant in walk- forward out-of-sample tests, yet CLV stays positive (+2.6 to +3.3%) in every sub-period. (iv) CLV predicts realised ROI one-for-one (slope 0.91, SE 0.09) and detects an edge with roughly 300 times fewer bets than ROI. (v) Dynamic Dixon–Coles and Elo models are 0.021 nats worse than the close and add no out-of-sample information. (vi) The Poisson “at least one” transform that underpins anytime-scorer pricing is well calibrated but slightly under-predicts scoring (71.6% vs. 72.7%). We translate these results into an engine specification—sharp-consensus de-vig, EV thresholds, quarter-Kelly with exposure caps, and CLV as the primary performance metric— and develop the player-prop methodology (rate×opportunity, empirical-Bayes shrinkage, correlated same-game parlays) that the engine applies where markets are thinner.

We formulate a sports-wagering research architecture that separates three tasks: forecasting the joint distribution of game and player outcomes; detecting mispricing relative to executable odds; and allocating risk under estimation error. The design combines dynamic team strength, player opportunity, conditional efficiency, opponent interactions, latent game environments, and timestamped market information. Its principal claim is methodological: a model should be promoted only when it improves paired out-of-sample scoring and produces positive net economic results under realistic execution. Mathematical complexity is not evidence of an edge. The supplied benchmark is audited using within-market Brier skill. NHL and NFL show the largest proportional improvements over their respective baselines; the benchmark does not establish that soccer is the most informative or profitable market. We propose hierarchical count and possession models, a regularized market-residual ensemble, held-out probability calibration, robust expected-value filters, and joint-scenario portfolio optimization. Every performance target and staking constant below is a research hypothesis or illustrative control, not a promised outcome.