About the system

FIRELINE is a basketball prediction model covering NBA and WNBA. For every game it estimates each team's strength in that exact context, converts the difference into a win probability and a projected margin, and compares both against the market to find price inefficiencies.

How it works

• Latent team power — an EMA-updated rating per team, learned from game margins over time.
• Situational adjustments — home court, back-to-backs, recent workload, rest days, travel hours, jet lag direction, and motivation flags (playoff locked / eliminated).
• Missing players — each absence is priced by production, lineup coupling and out probability; losing multiple core players compounds the penalty non-linearly.
• Probability — the adjusted power difference maps to a win probability and projected margin through calibrated logistic curves.

The model — feature inputs

What sets it apart

• Price ladders, not picks — most forecast sites say "take team X". FIRELINE publishes the whole curve: the exact odds and handicap where the edge crosses +1%, +5%, +10%… and the stake each level justifies. If your book moves, the answer is already on the page.
• Injuries priced by fit, not star power — an absence is valued by how the remaining lineup is built around that player (lineup coupling × probability of missing), not by a generic per-player rating. Two equally-rated players can cost very different amounts.
• Depth collapses are non-linear — losing one core player hurts; losing two or three compounds beyond the sum of parts, because the bench can't absorb it. Most models scale linearly; this one doesn't.
• Travel physics — jet lag is directional (eastbound and westbound carry different coefficients), travel hours are continuous rather than a yes/no flag, and "didn't have to move" is its own signal.
• Motivation flags — a team that already locked its seed, or one that's mathematically eliminated, doesn't play like a team with something to play for. The model tracks both.
• No black box — every number on the card is backed by visible intel: workload, rest, travel, form, absences. You can audit why the model likes a spot before staking on it.
• Honest validation — backtests are walk-forward only: every call was made with information available before tip-off. No refit, no lookahead, no cherry-picked windows.

What the site shows

• Moneyline & handicap ladders — the odds or line needed for +1% to +20% edge, with the Kelly stake each level would justify.
• Intel — the situational features the model detected for each side, so you can see why the number is what it is.
• Live mode — in-game win probability and margin tracking once a game tips off.

Staking

Stakes shown are pure Kelly fractions of bankroll, derived from model edge at each price. Production staking applies a fractional divisor and caps.

Why public?

FIRELINE was built and used privately over the last year, returning 4×+ on its bankroll. When legislation in my country restricted sports betting, the model lost its practical use for me — so instead of shelving it, it's published as-is: live forecasts, registered picks and the full walk-forward history, for anyone to audit or reuse.

Disclaimer — outputs are statistical estimates, not betting advice. Past performance does not guarantee future results.

Teams

Latent team power and the coupled player pool the model prices from — production is the learned per-player output, coupling how much of the team's power runs through that player, value the missing-player penalty. Players flagged missing on today's slate carry their status.

Backtest

Walk-forward simulation of the selected model over past seasons — every prediction is produced with data available before tip-off only. No lookahead, stakes sized by the same Kelly logic shown on the site.

Return — rolling 7-day

Bankroll — compounded

Bankroll — by bet

By season

click a season to filter the game log

Season	Games	Accuracy	ML hit	ML ROI	HC hit	HC ROI

Game log

date	match	result	prob mkt / fl	margin mkt / fl	ml op	hc op	pf op