Methodology

How we measure ourselves, the rules we follow when we bet, and what the model doesn't handle well.

What we measure

The first-order metric is closing line value (CLV) — how often we get a price better than where the market settles by game time, weighted by stake. CLV is the cleanest signal that bets are finding real edge rather than getting lucky. The closing line is the market's most informed estimate after sharps have moved it; beating it consistently is the closest thing to a skill score in this business. We report it in percentage points (pp) of implied probability — +2pp means the price we locked in was two points better than where the line closed.

ROI is the second-order metric. It will be noisy over any window short of a full season, and it can diverge from CLV when the sample is small.

Across the 4,444 bets in the 2024–2025 backtest, mean CLV is +2.89pp, with 60.8% of bets positive. That is the durable number. Whatever ROI you see on the track record is what that CLV happened to convert to.

We have caught and fixed errors in our own pipeline before — the kind of mistake that would have inflated those numbers — and what you see on the track record is post-fix.

Why the closing line

Holding ourselves to the close only makes sense if the close is actually sharp. So we checked. Across three seasons we measured the closing line's own reliability — when its price implies a 60% chance, does that side win about 60% of the time? It does. The closing line is well-calibrated: a hard benchmark, not a number we chose because it flatters us.

That sets the bar for what we stake. A market earns a bet only when our calibrated probability is measurably more accurate than the close — and only when that edge survives out-of-sample, on a season the model never trained on. Some markets clear the bar. Others don't, and we don't bet them, however tempting the line looks. The market mix on the track record is the set that beats the close — nothing else.

How we bet

Bets are placed by rule, not by feel.

• Bets are placed only when the model's edge over the market is large enough — there is a floor, and most days most games do not clear it.
• Sizing is a fixed fraction of the mathematically optimal bet, with hard per-bet and daily caps. Never all-in, never doubling down to chase a loss.
• Single-leg bets only. No parlays. Parlays compress edge faster than they expand variance, even at conservative leg counts.
• Some days the model rejects every game on the slate. Zero-bet days are a feature, not a bug.

None of these rules are tuned to maximize backtest ROI. They are tuned to survive the drawdowns the model will produce at our sample size.

Limitations

The numbers on the track record assume 5% vig and a flat $100 stake. Sharper books will produce better results; fish books worse. Flat stake strips out compounding, so the displayed ROI reads as average edge per bet, not as a bankroll outcome.

Even when the model has edge on average, individual bets lose. A 5% edge over a near-coin-flip line still loses close to half the time. Losing days, losing weeks, and drawdowns spanning hundreds of bets are statistically normal — not evidence the system is broken. This product does not predict winning days. It produces calibrated probabilities and follows a disciplined betting rule. Variance does the rest.

Two things the model handles poorly: extreme weather games where late lineup scratches move projected runs more than the model accounts for, and the early weeks of a season, when rolling stats have not yet converged. We do not place bets until those rolling windows have enough data to be reliable — typically two weeks into the regular season.