Methodology
Most services flash their wins and bury their losses. We do the opposite. Here is exactly how Blitzen turns raw market data into models — and the honest record of what they do and don't beat.
The pipeline
A model is only as honest as the path that built it. Ours is one straight line — and the same feature code runs at training time and at serve time, so what we test is what you get.
~40 sportsbooks via The Odds API + ESPN / nflverse stats, every line since it opened.
Team form, pace, EPA, injuries, weather, and market-structure features — built once, identically for train and serve.
Gradient-boosted trees per market, walk-forward by season — trained only on the past.
Score on seasons the model never saw. No leakage, no cherry-picking, bootstrap confidence intervals.
Only models that clear the closing-line-value bar are allowed to call themselves production.
The survivors run live in the terminal — every prediction stamped with its honest track record.
~40 sportsbooks via The Odds API + ESPN / nflverse stats, every line since it opened.
Team form, pace, EPA, injuries, weather, and market-structure features — built once, identically for train and serve.
Gradient-boosted trees per market, walk-forward by season — trained only on the past.
Score on seasons the model never saw. No leakage, no cherry-picking, bootstrap confidence intervals.
Only models that clear the closing-line-value bar are allowed to call themselves production.
The survivors run live in the terminal — every prediction stamped with its honest track record.
The analysts
We don't run one black box that does everything. Each market gets a specialist, trained on the features that actually move it — and each carries its own honest status, never a borrowed one.
Moneyline
Picks winners from time-bucketed line snapshots, bookmaker disagreement, and team-quality gaps. Calibrated and running live on the NBA and NFL; the college-football model is served as research, not a proven live signal.
Spread
Reads the spread from team form, pace, shooting, and market structure. The spread market is sharp at our current feature set — so Xavier stays in tune-up until new signal earns its way back.
Totals (O/U)
Projects the over/under from offensive pace and defensive efficiency. Walk-forward shows a neutral edge today — so we say so, rather than dress it up.
The architectures we run
Today's workhorse — gradient-boosted decision trees, one per market.
A sequence model that reads how a line moves toward the close, bucket by bucket.
The compound model — validated parts assembled into one interpretable pipeline.
A house-and-community meta-model that ranks the whole slate on a unit-safe score.
The scoreboard
Win rate lies — you can win a coin-flip streak and still be lighting money on fire. The honest scoreboard is closing-line value: did our bet get a better number than the market's sharpest closing price?
Beat the close consistently and you are genuinely ahead of the market. So that is the bar we hold every model to — measured on games it never trained on, with confidence intervals, not vibes.
And the honest headline today: no model beats the close after the vig yet. The signal is real on moneylines; it just doesn't clear the juice on its own. We tell you that on every model — because a tool you can't trust is worse than no tool.
Every bar is measured, never claimed. A model earns the neon proven-edge badge only when it clears 0.52 with positive average CLV on data it never trained on. Today, nothing has — and we say so on every model.
In development
One model rarely clears the bar alone. But we have validated parts, not hopeful ones — and the compound model assembles them in evidence order, each doing the one job the data says it can. This is the build we are running now.
Nominates the bet and its direction — the only validated, leakage-free CLV engine.
Confirms or vetoes the direction. The strongest mover, but it never originates or sizes.
Sets a fractional-Kelly stake at the entry price. Calibrated, so it sizes — never gates.
Walk-forward beat-the-close, pinned to Stage 1. Production-eligibility is judged only on the originating engine's closing-line value — the sizing model's numbers can never sneak a model past the gate.
The goal isn't a magic box. It's to compound a few small, proven edges into something that honestly beats the close — and to ship it the same way we ship everything: gated, measured, and labeled with exactly what it has earned.
Every model, its features, and its real track record live in the terminal. Free to explore. Backtest any of it against seasons of real odds.