How it works

The Lean — one read per game

Every game on the board gets a single verdict: a side and one token — Strong Lean, Lean, or Pass. Not six scores, not a wall of near-duplicate trends. One call you can read in a few seconds.

The spine of the read is the Market Read: a plain, descriptive observation of how the one line we track moved from open to the latest read, in the context of the market ("the line we track moved toward Home, −1.5 → −3 by kickoff"). It describes what happened to the number — it is not a prediction of the game. When the line held, a lean can still be carried by its out-of-sample receipts alone — the row says so plainly.

• Strong Lean — the read and its support line up strongly.
• Lean — a real but lighter read.
• Pass — nothing cleared our bar; the market looks efficient.

Pass is the default — when the number looks fairly priced, the honest answer is a pass. We'd rather show you a confident Pass than force a lean the numbers don't support. The slate header counts them out loud ("X games · Y leans · Z passes") so a quiet board reads as discipline, not a missing signal.

The receipts — what backs a lean

When a lean draws on history, it shows its work. Every receipt carries the same three numbers, in the same order: the record (e.g. a 24–8 out-of-sample ATS read), the win rate, and the sample size (n=32). No bare percentages — a rate with no sample behind it is noise.

Two rules keep the receipts honest:

• Pre-registered condition families. The situations behind a lean come from a fixed list we commit to in advance — not a machine that mines thousands of combinations and keeps whatever looks good after the fact. No post-hoc cherry-picking.
• Out-of-sample record, gated by a confidence interval. Each receipt is scored on data it wasn't fit to, and we put a 95% confidence interval on that out-of-sample rate. The lower end of that interval has to clear the price-aware break-even — the rate you'd actually need to profit at the price you're paying — or, for a fade, the upper end has to sit at or below the mirrored line.

At a standard −110 line you need to win about 52.4% of the time just to break even (the −10 vig is the house's cut). So a trend that hit "65%" but whose confidence band dips below 52.4% hasn't earned a place in the lean — the number is consistent with a coin flip. Anything with fewer than about 30 out-of-sample games never shows up as a receipt at all — it's kept out of the verdict entirely.

A receipt is a past record — not a prediction and not betting advice.

Closing Line Value (CLV)

CLV compares the number you bet to where the line closed. Consistently beating the close is the single best long-run indicator that you're betting the right side — win or lose on the day. The bet tracker records CLV on every pick so you can see whether you're beating the market, not just the scoreboard.

CLV here is a personal historical metric: it's based on the line you recorded at log time versus the single line we track at close. We don't verify that a bet was actually placed, we don't auto-import from any sportsbook, and your CLV is yours alone — we never pool users' results into a published "proof" number. These are past results, not a prediction.

Backtests that don't fool themselves

When you build a system in Systems Lab, we hold back the most recent completed season and test against it separately. If a system only looks good on the data it was fit to and falls apart on the held-out season, that's curve-fitting — and we show you the split so you can tell the difference. We also report how many games were dropped for missing data, so the sample is honest.

Before a system can be published to your library, it has to earn it out-of-sample. We run walk-forward testing: each completed season is graded using only the seasons that came before it, so the test never peeks at the future. Then we apply a price-aware confidence check — the conservative (95%) end of the out-of-sample win-rate has to clear the rate you'd need just to break even at a standard price. A system that was only lucky in-sample can't get past it.

And because the lab can test a lot of ideas, we correct for that. When we search many systems at once we apply a false-discovery-rate correction (Benjamini–Hochberg) over every system tested. Test a thousand random ideas and a few will look great by chance; the correction, against the honest denominator, is what tells them apart. These are past results, not a prediction.

System health: Strong, Moderate, Fragile

When one of your saved systems qualifies for a game today, your dashboard tags it with a health label — Strong, Moderate, or Fragile — so you know how much weight it carries at a glance. It's a plain reading of the same publish-gate inputs (sample size against the publish floor, whether the held-out record points the same way, and where the out-of-sample confidence interval sits against break-even). It is a presentation of how well-supported the system is — never a claim about profit.

A system built to follow the other side of a trend shows a fade qualifier ("Strong · fade"): its raw record looks like a losing one because you're betting against it, so the label reads the health of the fade, not a losing bet.

Why we retired our own scores

Week18 used to headline a set of home-grown numbers — a 0–100 Power Rating, a Convergence / Dual Signal score, and an "Edge" figure — behind a qualification gate that let a trend through at 65% over 25 games. We tested those the same way we test everything, and demoted them.

The gate was the problem. A trend hitting 65% over 25 games sounds strong, but its 95% confidence interval reaches below the ~52.4% break-even — statistically, it's consistent with a system that loses money. Headlining that as "signal" is exactly the bar The Lean was rebuilt to reject, so those raw scores no longer decide a verdict.

They didn't disappear — they were demoted to inputs. You can still inspect them: open a lean's Full breakdown and the Power Rating, convergence and matchup-edge numbers sit there as clearly-labelled raw machinery, reconciled to the one Lean. The quantitative model's raw daily selections live on the Model Picks page, framed as model internals rather than our call. ("Edge" no longer renders as a chip on those pick cards — it survives only inside the Full breakdown.)

That's the whole credibility story: we held our own scores to the same bar we hold every trend, and when the math didn't clear it, we said so and moved them behind the verdict.

What Week18 is — and isn't

Week18 is a focused sides & totals research tool: read the lean, build and backtest your own situational systems, and track every bet's true ROI and closing-line value. Lines and backtests reference a major US sportsbook (DraftKings or FanDuel where available) (no affiliation) — a number you can realistically get down at.

It is not a multi-book line-shopping scanner, a player-props model, or a +EV / arbitrage finder. If that's what you're after, we'd rather say so up front than pretend — we may add multi-book line shopping later if enough of you ask for it.

Where the numbers come from

We track one line per game. The figures are meant to reflect a major US sportsbook's number — a price you can realistically get down at — but they reach us through third-party sports-data providers, not a direct book integration. Because we track a single line, it isn't a multi-book price and it can differ from the exact number in your own sportsbook.

The line updates when it moves: our feed publishes on an actual change, not on every poll, so a steady market legitimately holds. Wherever a lean or a game shows a "Lines last updated" stamp, that's the last time the tracked line moved — cross-check it against your book before you bet, and treat anything more than a day old as stale.

Scores and schedules also come from third-party providers, can be delayed and are for informational use only. Past results are not a prediction of future outcomes — nothing here is betting advice. Please bet responsibly. If gambling is a problem, call 1-800-GAMBLER or visit ncpgambling.org. 18+ (21+ where your jurisdiction requires) — we require your date of birth at signup and refuse under-18 registrations.