Quantitative Research

A projection system for every skill player and team in the league, rebuilt week by week from only what was knowable before kickoff.

• ▶Market-anchored game, engine-owned players: each game's points come from the closing spread and total (no model beat the close out of sample); the engine owns how those points become attempts, carries, targets and yards — opponent-adjusted efficiency rates with Bayesian shrinkage, recency-weighted usage shares blended toward prior-season shares early in the year.
• ▶Every input earned its place: forecast wind (−1% passing yards per mph above 10, held out on 2024–26), yards-before-contact for backs, a fitted sack-rate blend, a tight-end tilt against stingy defenses, per-snap efficiency reversion for receivers, validated rookie usage priors — and a longer list that did not (below).
• ▶Roster-aware: injured weeks are excused rather than counted as zero usage; a traded player leaves one club's pool and carries his own share to the other from his first week there; a benched QB is read off the live depth chart, never by hand.
• ▶Replayable: any week since 2016 rebuilds from the data available entering that week, which is what makes every study below a real out-of-sample test.

Probability is a claim; a ledger is the audit. The system publishes its own calibration and has never been allowed to rewrite a graded result.

• ▶Empirical distributions, not assumptions: prop probabilities come from the measured distribution of actual-to-projected ratios, bucketed by projection size (a lognormal fit was tried first and failed calibration badly); small counts — passing TDs, interceptions — are Poisson, verified within three points at every rung.
• ▶A ledger that keeps the misses: every priced rung of every played game is stored with its probability, the market's price and what happened. A hit rate needs the losers; a hit list is just a filter.
• ▶The market as the reference: a 1,518-leg study showed the books better calibrated than the engine in every market, so the engine was anchored to them — and from October 2026 it is scored, unanchored, against the Kalshi exchange on the same legs, with Brier scores published by market as the weeks grade.
• ▶Streaming models with records: kicker and defense streaming rankings held out over 2021–25 (top-five picks 8.8 vs a 7.9 field at kicker, 9.0 vs 6.3 at defense).

An end-to-end integration with a CFTC-regulated exchange: market data in, model prices against it, orders out.

• ▶Twenty series mapped to the engine's markets — every player ladder, anytime touchdowns, and the spread, total and winner for the full game, first half and first quarter — synced hourly and frozen at kickoff as each market's close.
• ▶Period lines priced from history: full-game lines off a 14,000-perspective residual pool around the closing line; half and quarter lines off a kernel over 7,000 games (a shifted pool blurred the lumps a short period lives on — a fifth of first quarters end tied), validated walk-forward 2019–25.
• ▶Combo pricing by RFQ: signed (Ed25519) requests, maker quotes read in under a second, per-leg contribution quoted by sub-combo, idempotent order records written before any accept so nothing is ever bought twice, fills reconciled against the exchange.
• ▶A board built for the exchange: every rung priced Yes and No on the engine's range bar, a scan for mispriced rungs (floors, inverted ladders, wide gaps, with size and liquidity guards), the biggest price moves over any window, and a slip that reprices itself on every change.

Language models as analysts, held to the same standard as the engine: structured inputs, enforced outputs, graded calls, lessons that compound.

• ▶A dossier, not a prompt: each plan is written from a structured JSON dossier — projections, engine inputs with their league baselines, season context reduced from play-by-play, both injury reports with snap shares, beat reporting, weather, roster moves, and the plans' own record this season by call type.
• ▶Doctrine from the studies: the system prompt carries what the research found (volume is sticky, efficiency reverts; opponent strength matters only after six games; the market is the reference) so the writer reasons inside the evidence rather than around it.
• ▶Write once, grade always: a plan is never rewritten after kickoff. Every call is graded by code, the post-game report reviews the plan it was written against, and the lessons are carried into the next plan on that team — a closed loop between prose and results.
• ▶Schema-enforced output with numbers attached by code, never by the model; an interception call must name a volume mechanism; a departed starting lineman the engine never tracked is named in the notes so the writer weighs what the numbers cannot.

A second sport built on the same doctrine, to prove the method travels: walk-forward, out of sample, documented rejections.

• ▶From the puck up: an expected-goals model on league play-by-play since 2010, team efficiency with opponent adjustment, goalies and buildings, strength states, and skater on-ice ratings with projected lineups.
• ▶Every market from one walk: a ten-second score simulation prices moneylines, puck lines, totals and props off the same game state, with negative-binomial spreads and a pulled-goalie mixture for the last minutes.
• ▶Honest about the last step: style matchups (pace, shot profile) added nothing in the residuals — recorded as rejected, with the one finding worth keeping (own pace slightly over-credited).

The same discipline applied to intraday futures: labeled ground truth, causal evaluation, and outcomes measured in dollars.

• ▶Ground truth by hand: trading sessions labeled into regime segments under written conventions, an evaluation set built to teach and to score.
• ▶Causal replay: the Morpheus adjudicator replays sessions bar by bar with no look-ahead, scored against the labels (0.81–0.84 occupancy against a 0.62 mechanical baseline, ~90% of an oracle ceiling) — see the Projects page.
• ▶Execution harness: a playbook simulator walks the full method over any timeline, so a classification is measured in P&L; a break-even stop study quantified the trade (about 0.1R a trade for half the drawdown).
• ▶Live layer: real-time regime tracking on one-minute bars with a quarantine on the feed's recent window, bridged to NinjaTrader.

The plumbing that makes the research reproducible: 7,000 games, 117,000 player-games, a quarter million graded legs, all rebuilt from stamped, versioned feeds.

• ▶Feeds: nflverse play-by-play, box scores, snaps, rosters and depth charts; Sleeper live injuries, depth charts and news; PFR advanced stats; FTN charting; Kalshi and the Odds API; a weather forecast feed; the NHL league API.
• ▶Incremental by design: every sync writes only the rows that differ and stamps the feed; priced weeks and the engine context are cached on a version built from those stamps, so a correction upstream reprices and a quiet pull costs nothing.
• ▶Feature flags for the paid side: site-wide switches silence any scheduled pull while leaving manual pulls to the operator, and hide the pages built on a feed that is off.
• ▶Tested: roughly 650 tests across the applications; the NFL engine and exchange flow alone carry 237.