Moneyball with a Research Program

Most sports-modeling efforts build a single predictor and expect it to be profitable. This program refuses that framing. It asks, for each of sixteen sports traded on Kalshi and Polymarket, a prior question: how learnable is this outcome at all? A single MLB game sits closer to a coin flip than a single NBA game, and no amount of feature engineering repeals that. So the first deliverable for every sport is not a model but a measurement of its irreducible noise floor and its ceiling.

The stated edge, if it exists, comes from one of three places: better priors (ratings that misprice less than a narrative-driven crowd), faster or richer information (injuries, lineups, weather, line movement, in-game state), or structural and behavioral inefficiencies (thin liquidity, favorite bias, cross-venue mispricing). Crucially, the program is explicit about what it is not doing — it is not copying whale flow or scalping manipulated short-horizon crypto markets, both of which it had tried before and judged fragile and gameable.

The engineering premise is that heterogeneous sports data — play-by-play, box scores, 1v1 match logs, race results — can be reduced to two tables with identical columns for every sport. GAMES holds one row per contest: the two sides, scores, winner, season, point-in-time ratings, and the de-vigged market closing probability that serves as the benchmark. MARKETS holds one row per price observation, joinable back to GAMES for Closing-Line-Value analysis.

The payoff of this discipline is that the market mathematics — de-vigging, expected value, Kelly sizing, log-loss, Brier, expected calibration error, and CLV — is written exactly once and works across every sport, as is a generic leak-free, point-in-time Elo that gives each sport a baseline to beat. Schema is enforced rather than assumed, and each sport has a runnable ingestion pipeline that materializes a compact parquet backbone and registers it in a self-documenting manifest.

Above the backbone sits not a single tabular model but a stack. The team view runs leak-free Elo into a calibrated gradient-boosted tree, validated walk-forward. A player view composes per-player ratings, minutes-weighted, into team strength — the documentation notes NBA beats team-only on all metrics and that MLB starting-pitcher information adds signal. In-game, trained state-to-win-probability models replace a weaker diffusion approximation and are checked against real minute-by-minute replay.

The philosophy is consistent throughout: the deliverable is a calibrated probability rather than a pick, so the system optimizes probabilistic loss and treats a near-diagonal reliability curve as a gate a model must pass before it is allowed to trade. The market is treated as a strong teammate rather than an adversary — a per-sport blend weights model against market rather than pretending the model can simply overrule a liquid close.

The headline is delivered without softening: the calibrated models clear every naive baseline, yet none of them out-scores the liquid close, and that residual is the market being efficient. The soccer backbone makes the point concretely and reproducibly. On 14,000 matches across six European leagues, the de-vigged closing line is near-perfectly calibrated — an actual home-win rate of 0.43 against a mean implied probability of 0.428 — and a naive leak-free Elo does not beat it, posting a Brier of 0.241 against the market's 0.212.

The program is equally candid about its own near-misses. Two eye-catching AUCs, golf at 0.982 and tennis at 0.876, turned out to be leakage; under audit they fall to leak-free values of 0.572 and 0.723. The general rule the project extracts is that 90%-plus accuracy is real only on a confident subset or late within a game, and never pre-game across all games.

Before any model is written, every sport receives a six-panel exploratory analysis that reduces the backbone to an entity table — one row per team or player with Elo, win rate, form, margins, and home/away splits — and probes its structure with PCA and t-SNE, correlation heatmaps, margin and Elo violins, rating drift, and, where odds exist, a de-vigged closing-line density.

The visual grammar is deliberately simple. A sport with a smooth win-rate gradient and a wide Elo spread — golf, tennis — shows an elite ridge that a rating model can exploit. A sport that collapses to a single tight blob with no gradient — MLB, UFC, both near the coin-flip AUC floor — announces that its ceiling is low by nature. Establishing that ceiling first is itself a finding, because it tells the program not to mistake variance for edge in a high-noise sport.

Having shown that the world is priced efficiently, the program relocates the search. A forecasting edge must be right about the world; a market-structure edge only needs to be right about the plumbing — which desk is quoting, which side the crowd piles onto, which venue moves first, how quickly injury news propagates, how a resolution source settles a contract. The historical-close data that repeatedly killed forecasting edges structurally cannot contain order flow, cross-venue quotes, or in-play tape, so it could never disprove the structural families.

This frames roughly a hundred candidate signal families — designed by adversarially collaborating quant agents and red-teamed — as live hypotheses rather than truths. The caveats are stated as sharply as the thesis: diversification reduces variance, not edge, and is useless if each edge is near zero on a liquid venue; rating-keyed families are one view measured many ways, so eight nominal signals at ρ̄≈0.5 collapse to N_eff≈1.8; and a hundred families is a data-snooping machine that demands false-discovery control and forward-CLV as the un-foolable arbiter. The honest expected outcome is a modest market-making book plus a few un-provable niches — and capital allocation to any forecasting family stays at zero until it clears live CLV.