Methodology
How the engine works, and how it scored
Every win percentage on the grid comes from a model we can show you. This page explains how that model is built, then reports how it did in a walk-forward backtest of the 2014–2025 NFL seasons — including where it fails to beat the market.
Results
Did it work? Past seasons replayed week by week.
Never behind the market or the chalk
Fifteen seasons, 2011–2025, replayed on the board as it stood in each week.
6 of 15 seasons the engine won it all, running ten tickets with rebuys through week 4 at the default cap
15 seasons replayed week by week, with no knowledge of any line, result or pick share that had not happened yet
0 seasons the engine lagged the market or the chalk
- 4.92 weeks a single engine ticket survived on average, against 3.25 for the market’s own pick, over 2014–2025. Half a season further.
- 85.2% of the engine’s top weekly picks won, over the same years.
Every season is rebuilt as the board actually stood in each week and replayed with the engine that prices this grid, so a future week carries only the line posted at the time. Winning a season means a ticket was still alive at the final week, running ten tickets at the default cap of 70% on one team. Across the six caps tested the survivor counts range from three to six seasons and are not statistically separable at this sample size. Past results are not a forecast of your pool.
Is it more accurate than the market?
Walk-forward, favorites only, — games.
Brier score and log loss both measure how well a probability matches what happened — lower is better. "Top pick hits" is how often the single highest-rated team of the week actually won.
Where the edge actually shows up
One ticket, pick the best unused team each week, until it loses.
Survivor is not a forecasting contest, it is an allocation problem: you can only spend each team once. Two models can be equally accurate and still last very different amounts of time, because they disagree about when to spend a good team.
What a book of entries actually does
Every ticket on a different team, rebuys open through week 4.
"Reached the final week" counts a season where at least one of your tickets was still alive at the end. Rebuys are the extra entries bought after an early loss — real money, so they are shown, not hidden.
The same ten entries, spread differently
This is the setting that decides whether more entries are worth anything. Ten tickets that all follow the same rule are one ticket you paid for ten times. Spreading them is what the Strategy tab is for.
How it works
What goes into each number, and why.
The four numbers on the grid
The Grid lets you switch the win-probability source. They are genuinely different models, not re-skins of one number:
- Vegas Moneyline — the consensus sportsbook moneyline with the bookmaker's margin removed, so the two sides sum to 100%. This is the market's own estimate and the baseline everything else is measured against.
- Empirical 2010–2025 — forget the moneyline: at this exact spread, home or away, how often did that team actually win? Read straight off every posted NFL spread since 2010, interpolated between half-point steps, with thin buckets dropped.
- Logistic — a fitted curve through spread and home-field, so it stays smooth where the empirical table is noisy.
- AI Assisted — the blend we default to and the one described below.
How the AI Assisted number is built
It starts as 45% market, 55% empirical table. The market is sharp, but it prices a point spread — and at some spreads, teams win noticeably more or less often than the spread implies. The empirical half captures that.
Then a set of situational adjustments apply: line movement across the week, divisional games, rest and bye weeks, time slots, neutral sites, and where we are in the season. Each one was measured against the blend before it was allowed in. Two deliberate constraints keep them honest:
- Fades at full strength, boosts at half. Warning signs (line movement against a favorite, Monday home spots, mid-sized favorites off a bye) held their full measured size out of sample — half-faded favorites were still losing more than their number said. Positive bumps did not earn the same trust, so they stay discounted by half.
- Capped at 8 points. However many rules fire on one game, their total is bounded, so no single matchup can be dragged far from the market.
For weeks that have not arrived yet, the number is additionally shrunk toward 50% — more the further out it is — because a line ten weeks away is not the line you will actually get. The drift table below is where those shrink factors come from.
Every situational rule, with its receipts
Favorites since 2014.
"They won" is the actual win rate of favorites in that situation. "Blend said" is what the model expected before the rule. The gap is the size of the effect. A take rule should show a positive gap, a fade rule a negative one.
Sample size is the column to read first. A 9-point gap on 27 games is a hint; a 3-point gap on 753 games is a finding.
Why future weeks are shrunk
How far a spread moves between a look-ahead and kickoff.
Measured on archived weekly grids: for every game, the spread as posted n weeks early against the spread it closed at.
A week from now the number barely moves. Ten weeks out, a three-point favorite is still favored by three about four times in five — good enough to plan around, not good enough to treat as fact. That is why a week-14 cell is quieter than a week-1 cell.
Why pool size changes the plan
Chance the pool is still contested, by entrants.
Computed from how fast the public field thins — about 74% of entrants survive a typical week. Below the line, saving a stud for a late week is insurance on an event that mostly never happens, so once a pool's field size is set on the Pools tab, the planner discounts future value by these numbers. In big contests the discount is zero and the plan is unchanged.
Limits
What the backtest can and cannot tell you.
Limits, and what we are not claiming
A backtest that only flatters the product is not evidence. The honest boundaries:
- The rules were found in this data. The empirical table is refit each season using only earlier years, so the blend is a true walk-forward. The situational rules are not: their constants were chosen from the full history, so the gaps in that table read optimistically. That is why boosts are halved — and why fades run at full strength only after their out-of-sample gaps matched the full constants.
- Twelve seasons is a small sample. The survival results come from twelve numbers. In most of those seasons the engine and the market picked nearly the same path.
- Some years kill everyone. Where both columns read zero, the week-1 favorite lost and nothing else mattered.
- It is not a forecast of your pool. Results depend on your pool's size, rules, and what everyone else picks. A tool cannot promise a season.
- Survivor Board is planning software. We do not operate contests, accept wagers, or pay prizes.