THE BETTING FOOL
By a Man Who Should Know Better
Ten positions a week, in two sections. The Five is five sides, ranked, published every week without exception — because a pick-five pool makes you submit five whether or not any of them is a bet worth making. The Board is up to five totals. A game can appear in both; a side and a total are separate questions.
A pool pick and a wager are different games. A pool pick competes against the other entrants, so a 51% side is a perfectly rational entry. A wager competes against the vig, so the same side is a losing bet. The card refuses to blur that, which is what the tags are for.
WEATHER UNDER — an outdoor game with a forecast of 10 mph of wind or more, or forecast rain, on the under. 2.5% of bankroll.
This started as a wind rule and was widened in September 2026 when a reader asked whether rain did the same thing. It does, harder. Wind of 10 mph or more: 57.7% under across 603 games. Observed rain or snow: 63.1% across 157 — and that is not the wind arriving alongside it. In calm rain it is 59.1% against a dry 48.3%; in windy rain 68.8% against a dry 56.4%. The two stack. Every weather percentage in this section is reproduced by model/weather_rule.py — outdoor games only, 2016–2025, pushes dropped. The figures elsewhere on this page come from the backtest and the line-movement study, not from that script.
What makes it believable is that both are the same failure, visible in the market's own numbers. The posted total steps down as conditions worsen and never steps far enough:
| Conditions | Posted total | Actual | Gap |
|---|---|---|---|
| Dry and calm | 45.11 | 45.92 | +0.81 |
| Wind 10+, dry | 44.34 | 43.07 | −1.27 |
| Rain, calm | 43.74 | 41.10 | −2.65 |
| Rain and wind 10+ | 42.67 | 38.98 | −3.69 |
In calm rain the number falls 1.37 points while scoring falls 4.83 — the market moves about 28% of the distance, almost exactly the fraction measured independently for wind, where it moves 27%. Taken together, wind-or-rain goes 57.9% across 696 games, z of +4.17, above break-even in eight of ten seasons. It is played whichever way the model leans.
Two honest limits. Rain on its own is thin — 157 games, roughly sixteen a season, and it falls to 58.1% (z = +1.76) if 2022 and 2023 are removed; it rides because it agrees with wind for the same reason, not because it stands alone. And the history is measured on observed weather, while a card has to be built on a forecast. The thresholds are deliberately demanding: 10 mph of wind, 55% chance of rain, or 0.20 inches expected across the game window. A naive calculation says an 18% chance of rain clears the vig; that number assumes perfectly calibrated forecasts and treats a passing shower as equivalent to a soaking, so it is not used.
BEST BET — the model's cover probability clears what the posted price demands by at least one full percentage point, in the worst quarterback scenario as well as the likely one. 1.5%. (Before Week 6 any margin counted. Clearing by less than a point is inside the noise of the estimate.)
LEAN — it clears by less than a point, or not at all. Published for the pool players and to show what the model was looking at, staked at a token size, and marked on its face as a losing proposition. 0.5%.
Stakes go by tier and nothing else. Across all sixteen sides in the 2026 opener the model's cover probability spanned 52.04% to 52.91% — 0.87 of a point between its best idea and its worst, less than the gap between a −105 and a −112 price. It cannot rank its own plays, so sizing by its enthusiasm would be sizing by noise.
Every column carries an UPDATED time in Eastern at the top of the page and on the printable card. Weather-driven plays go stale: the Arrowhead under in the 2026 opener was published as a wind play at 10.8 mph and requalified on the Tuesday re-pull, on rain, when the wind forecast fell to 8.6. When that happens the column says so in the text rather than quietly swapping the justification, and the time stamp tells you how old the numbers you are reading actually are.
The headline record and the bankroll answer to the bets — the wind unders and the plays that clear their price. The Five keeps its own pool record alongside, covering all five sides including the leans. Everything published is graded either way; the two are simply not the same measurement and are not added together.
Updated October 6, 2026. Since Week 4 the numbers come from a Ridge regression — a regression that is penalized for leaning too hard on any one input — of a team's scoring on 27 inputs: thirteen offensive measures (passing yards, rushing yards, yards lost to sacks, interceptions, net fumbles, non-touchdown points, points scored, expected points added per play overall and on passes and runs, plays per game, pass rate and success rate), the same thirteen measuring what the opposing defense has allowed, and a home-field term. Each input is the team's season-to-date average blended with last season's until six games are in. The penalty is chosen by cross-validation on every game since 2017 (5,116 team-games as of Week 5), and it comes out large, which is the model telling us how little signal there is in a single week of football.
Why Ridge. Each week the model is raced against the alternatives on 2018–2025, predicting each game using only games already played. Average miss on the final margin: Ridge 10.24 points, Lasso 10.24, plain regression 10.35, XGBoost 10.40, and the closing line from Las Vegas 9.83. The best model is still about four-tenths of a point a game worse than the market. The more complicated the tool, the worse it did. On totals the model is worse than the market too (10.90 against 10.42), so model totals are never bet; totals are bet only on the weather rule.
What it still does not do. It has no strength-of-schedule term beyond what the defensive inputs carry, and it cannot see injuries on its own. Quarterback changes are priced separately in points, using a shrunken estimate of each passer's expected points added per dropback, and applied only when the likely starter has taken less than 75 percent of the team's dropbacks this season. Other absences are priced from the weekly injury reports and snap counts, in small deductions capped at four points. Both adjustments are judgment layered on a model, and the column says which games they touch.
Updated October 13, 2026. Four changes. (1) Final scores come from the final box score; the earlier training set dropped points scored on the last play of some games. The race above was re-run on the corrected data and the order did not change. (2) Cover probability comes from a symmetric curve, 1 / (1 + e−0.0218 × edge), where the edge is the model's disagreement with the line in points: each point is worth about half a percentage point, floored at 44% and capped at 62%. (3) A bet must clear its price by a full percentage point. (4) Neutral-site games get no home-field term, and where the starting quarterback is in real doubt every scenario is priced and weighted by its probability.
The test that matters. Does the model's disagreement with the closing line predict who covers? It does not. Trained forward, one season at a time, on 2018–2025, the side it preferred covered 50.8% of 2,175 games, and the correlation between the size of the disagreement and the size of the miss was 0.001. So the model is used to choose among close calls and to find where a quarterback or an injury could move a number, and never to claim that the line is wrong. The weather rule is the one edge on this site with a record behind it, and it is why most of the money sits on the Board.
Weeks 1–3 were published from the earlier version: a plain regression on six box-score terms, with no home-field term and no defensive inputs. Their columns describe it as published.
Every play shows the Fool's number next to Vegas's number, the edge between them, a cover probability, and — the one that decides whether a bet is worth making — NEEDS TO WIN: the win rate the posted price demands. A bet at −110 needs 52.4%. At −120 it needs 54.5%. A cover probability printed without it is half a sentence.
None at the moment, and the reason is worth saying rather than hiding. Projecting a player's yardage is easy enough; comparing it to a posted prop line is not, because no data source this column has access to carries prop lines. Publishing a projection against an invented line would be making up the only number that matters. Props come back the day there's a real feed of prop prices.
Every week has a one-page PDF, linked at the top of the column. It's generated from the same file as the page, so the two cannot drift apart — same picks, same prices, same stakes, same bankroll.
Over 2,434 games the original model hit 52.4% against the spread. Break-even at −110 is 52.38%. Backing every road underdog blindly over the same games returns 51.8%, so the model's contribution above a mindless heuristic is under a point. Separately, its picks beat the closing line by about a sixth of a point — real, statistically solid, and roughly a tenth of what it takes to overcome the juice. None of that is a reason to bet the mortgage. All of it is on the site because you'd want to know.
The Betting Fool is for entertainment purposes only and is not financial advice. Past performance does not guarantee future results, and the model is wrong plenty — see the archive for proof. Must be 21+ to wager where legal. If you or someone you know has a gambling problem, call 1-800-GAMBLER.