THE BETTING FOOL
By a Man Who Should Know Better
Week one is in the books and the book says three and four. Nine hundred ninety-four dollars and ninety-seven cents, down five bucks and three cents from where we started, which after the week the market had is a result I'll take and not brag about. The Five went two and three. The three leans went two and one, which is the kind of thing that happens when you put five dollars on something and the universe decides to be funny about it.
The Arrowhead under cashed. Forty-one points on a forty-two and a half, and I want to walk you through why that one matters more than the twenty-two dollars it made. It went up as a wind play at ten point eight miles an hour, barely over the ten-mile threshold. By the Tuesday re-pull the forecast had fallen to eight point six and the play no longer qualified on wind — it requalified on rain, better than a third of an inch of it, and I told you so in this space rather than swapping the label and hoping nobody checked. Then it won. One game proves nothing. But the rule got tested in public, changed its justification in public, and paid, and that's the only way I know to run this.
Two corrections before the football. A reader sent me the week one finals and two of the sixteen rows were wrong — one a single wrong digit, Tampa Bay entered at 37 when they scored 27, and one a row shift that put Tampa Bay's actual score on Atlanta's line. Both landed in a graded record I'd already published. So the grading now pulls finals from the play-by-play database directly and never from a spreadsheet, mine or anyone's. The second correction is smaller and stranger: the database revises its posted lines to closing numbers after kickoff, and ten of the sixteen week one spreads moved between the card and the kickoff. Grade a published pick against the closing number and you're marking it against a price that was never on offer. The grader now reads the number printed on the card. Both fixes are in the code, permanently.
A third correction, and this one is mine rather than a reader's. Somebody went through last week's column looking for numbers that couldn't be reproduced, and found some: the rain percentages I quoted were carried in prose with no script behind them, and when I finally wrote the script they came back slightly different. Rain unders are 63.1 percent across 157 games, not 63.5 across 156. Calm rain is 59.1 against a dry 48.3, not 59.8 against 47.9. Wind-or-rain together is 57.9 percent across 696 games, not 58.0 across 695. None of it changes a play, a stake or the rule, and the wind figure that started all this — 57.7 percent across 603 games — reproduces to the decimal. But a number I can't re-run is a number I shouldn't have printed, so there is now a file called weather_rule.py that produces every one of them, last week's column carries a correction box, and I'd rather you read this than not.
Now. This week I did the thing properly and ran the three calculations underneath every number on this card, because several of you asked what's actually in the box. Here they are.
One: the regression. Ordinary least squares, team score on six box-score terms, fitted on 5,554 team-games from 2016 through week one of this season. A passing yard is worth 0.0586 points. A rushing yard is worth 0.0707 — rushing yards are more valuable per yard, which surprises people until they remember you only get them in situations where you're already winning. A yard lost to a sack costs 0.1176, about double what a passing yard gains, which is the single most underrated number in football. An interception costs 1.6535 points. A net fumble lost costs 1.3882. Every one of the six slopes carries a t-statistic past fifteen, every sign is the sign football says it should be, and the whole thing explains fifty-seven percent of the variance in team scoring. That is a real model of how points get made.
Two: the logistic. Take that model, run it walk-forward across nine seasons, and fit the probability that a pick covers against how far the model sits from the market: P equals one over one plus e to the minus b-nought plus b-one times the gap. Fitted on 2,434 out-of-sample games. Intercept plus 0.0815. Slope plus 0.00377, standard error 0.0146, z of +0.26. McFadden pseudo R-squared of 0.00002. In English: across a fourteen-point range of disagreement, the fitted probability of covering moves one and a third points, from 52.0 to 53.4. The model's confidence about a game contains essentially no information about whether the model is right.
Three: the sanity check, and this is the one that should make you careful. Forget fitting anything. The model's out-of-sample margin error has a standard deviation of 13.58 points. If that error is roughly normal — a reasonable first assumption — then a six-point edge converts straight to a cover probability of Phi of six over 13.58, which is 67.1 percent. That is the number a confident man would print. The logistic, which has watched 2,434 of these bets actually settle, says 52.6 percent. Fourteen and a half points of daylight between what the theory promises and what the ledger delivers.
That gap is the whole education. The normal calculation assumes the six points of disagreement are information — that the model knows something and the market doesn't. The record says the six points are mostly noise in my own number. Same arithmetic, opposite conclusion, and only one of them has been tested against money. Every cover probability on this card comes from the logistic. If you ever see a tout quoting you sixty-seven percent on a six-point edge, you now know exactly which calculation he ran and which one he skipped.
One genuine improvement this week: the model has finally seen some 2026 football. Week one's box scores are in and blended with last season's — and I'll tell you exactly how much that is worth, because the honest version is less exciting than the press release. The blend weights last season at five-sixths and week one at one-sixth, and last season's pool is seventeen to twenty-one games a team. So every input below is about eighty-three percent last November and about seventeen percent this September. That is the right weighting — one game is one game — but it means the blindfold is off by a corner, not off.
Ten positions, a hundred and forty dollars. Two weather unders at twenty-five each — San Francisco in an eleven-mile wind off the bay, New England in a sixty-seven percent chance of rain. Five sides at fifteen, and I'll note without enthusiasm that all five are underdogs again, which is my model's home-field blind spot doing exactly what I told you it does. Then three totals at five dollars apiece, labelled leans, which do not clear their posted price and are printed anyway so you can see the whole board the model looked at rather than the flattering corner of it.
Five sides, every week, whether or not any of them is a bet worth making — because a pool makes you pick five. Each one is tagged with what is actually behind it.
Atlanta getting two and a half at home, and the regression has this game at Falcons 21.7, Panthers 21.3 — a four-tenths-of-a-point Atlanta edge where the market has Carolina by two and a half. Two-point-nine of daylight. Now, where does that four-tenths actually come from? Not from the passing game: Carolina throws for 225 to Atlanta's 205, and at 0.0586 points a yard that is 1.2 points against Atlanta. Atlanta's ground game claws back six-tenths (125 rushing yards to 116, at 0.0707 a yard) and the sack and interception lines are close enough to call even. The entire Atlanta edge is one term: other points — field goals, extra points, the two-point stuff — where Atlanta averages 7.6 a game to Carolina's 6.2. At 0.6387 a point that is nine-tenths of a point, and it is the only reason this projects Atlanta's way at all. A kicking-game edge is a real edge and also a thin one. The play prices at even money, which is the real reason it's ranked first — at plus one hundred this needs exactly fifty percent to break even instead of the usual fifty-two four, and the logistic gives it 52.3. Two of those 2.3 points of margin come from the price rather than from anything I know about football.
Arizona plus three and a half at home against Seattle, and this is the thinnest read on the card — the model has Seattle 24.8, Arizona 23.0, so it agrees Seattle is better and just thinks three and a half is too many. One-point-seven of edge. Arizona throws for 260 a game against Seattle's 231, which is the whole case, and it's undercut by Arizona surrendering twenty yards a game to sacks against Seattle's twelve — at 0.1176 points per sack yard that's a point of difference handed straight back. Again the price does the work: even money, needs fifty percent, gets 52.2. I am not going to pretend a 1.7-point disagreement is insight. It's a dog getting a fair number at a fair price, which on this card counts as a reason.
Cincinnati plus two and a half in Houston and the model calls it 24.0 to 23.5 the other way — Texans by half a point where the market says Texans by two and a half. Two points of edge. The interesting term is turnovers, and it runs the opposite way from how this matchup reads. Cincinnati throws a full interception a game to Houston's 0.61, which costs the Bengals two-thirds of a point — but Cincinnati is minus 0.72 on net fumbles, meaning they recover more than they lose, and at 1.39 points a fumble the model hands them a full point back. Net turnover cost: Cincinnati 0.7 points, Houston 1.2. Houston has the worse giveaway profile of the two. Against that, Houston runs for 111 a game to Cincinnati's 96, which is fifteen yards and about a point in Houston's favour, and that ground-game gap is most of why Houston projects ahead at all. Two teams the market has separated by two and a half that the model has separated by half a point. Minus one-oh-two, needs 50.5, gets 52.2.
Las Vegas plus six and a half, and this one comes with a warning label. The regression projects Chargers 20.5, Raiders 15.6 — a 4.9-point Chargers win against a 6.5-point line, so only 1.6 of edge. But look at what's producing that Raiders number: 189 passing yards a game, 86 rushing, 23 yards a game lost to sacks and 1.26 interceptions. That is the worst input profile on this card — Tennessee's is the only one close, and it is three-tenths of a projected point better — and the model is still saying take the points, purely because 6.5 is a lot to give a team that bad. At minus one-oh-two it needs 50.5 and gets 52.2. If the Raiders are actually worse than their numbers — and a team surrendering 23 sack yards a game usually is — this is the play on the card most likely to look stupid on Monday.
Tennessee plus seven at home against Philadelphia. The model has Eagles 21.4, Titans 15.9, a 5.5-point gap against a seven-point line, so 1.5 of edge — the smallest side on the card. Tennessee's profile is grim: 182 passing yards, 89 rushing, 23 yards a game to sacks. Philadelphia runs for 121 and throws only 0.32 interceptions a game, the cleanest turnover number of any team on this slate. Everything about the matchup says Eagles. The only thing the model is arguing is that seven is a touchdown and this projects as a five-and-a-half-point game, and at minus one-oh-five it clears its price by a single point of probability. Smallest edge, smallest conviction, same fifteen dollars, because the stake tiers don't lie about ranking I can't do.
Under forty-four and a half in Santa Clara, and this is the biggest number on the card because it's the only kind of play here with real evidence behind it. Eleven miles an hour of wind forecast across the game window at Levi's, which clears the ten-mile threshold. Outdoor games in that wind have gone under 57.7 percent across 603 games since 2016, and the mechanism is visible in the market's own behavior — against dry calm games the books drop their posted total 0.77 of a point when it blows, while scoring falls 2.85. They move twenty-seven percent of the distance the scoreboard does. My regression happens to agree here, projecting 43.3 against a 44.5 line, but I want to be clear that its agreement is decoration: the totals arm of this model grades at 49.5 percent across nine seasons and I would be playing this under if the model screamed over. Discounted to 54.5 for forecast error. Needs 51.2 at minus one-oh-five. Twenty-five dollars.
Under forty-one and a half in Foxborough, on the rain leg rather than the wind leg — 67 percent chance of precipitation across the game window, comfortably past the 55 percent threshold, with the wind forecast at a harmless five and a half. This is the rule I widened two weeks ago after a reader asked whether rain did what wind does. It does, harder: outdoor games with observed rain or snow have gone under 63.1 percent across 157 games since 2016, and it holds inside calm conditions specifically — 59.1 percent in calm rain against 48.3 percent in calm and dry. The mechanism is the market's own number: in calm rain the posted total falls 1.37 points while scoring falls 4.83, so the books move about twenty-eight percent of the distance the scoreboard does. I should say plainly that rain on its own is thin — 157 games is roughly sixteen a season, and it weakens badly if you remove 2022 and 2023. It is on the card because it points the same way as wind with the same mechanism. My model projects 43.4 here, which is over the 41.5 line, so once again I'm overruling it. A forty-nine-percent totals model does not get to veto a fifty-eight-percent weather rule. Minus one-twelve, needs 52.8, gets 54.5. Twenty-five dollars.
Under forty-three and a half in the same Raiders-Chargers game, and here's a case worth studying rather than betting. The regression projects a 36.1-point total — seven and a half points under the line, the widest totals disagreement on the whole board. Run it through the normal calculation: the model's out-of-sample error on totals has a standard deviation of 13.94 points, so 7.4 points of edge converts to a 70.2 percent cover probability and I'd be laying the house on it. Now run it through the logistic fitted on the 2,472 totals that actually settled. That curve starts at 48.0 percent at zero edge — below a coin flip, because the totals arm of this model has graded 49.5 percent for nine years — and seven and a half points of disagreement drags it only as far as 51.6. The posted price of minus one-ten demands 52.4. Five dollars, labeled a lean, kept on the card precisely because it is the clearest illustration this week of the gap between the two calculations: eighteen and a half points of probability between the theory and the record, on the same bet.
Under fifty and a half in Dallas, the highest total left on the board now that Thursday is in the books, and the model has it at 45.9 — four and a half points under. Washington runs for 134 a game, second only to Baltimore among the teams on this card, and a team that runs bleeds clock; Dallas throws for 261. On paper that's an offense-heavy matchup, which is presumably where fifty and a half comes from. The normal calculation turns 4.6 points of edge into 62.9 percent and would have me betting real money. The totals logistic turns the same 4.6 points into 50.2, because that curve begins below fifty — the model's totals arm has graded 49.5 percent for nine years — and climbs about half a point of probability per point of disagreement. At minus one-oh-five the play needs 51.2, so it comes up a full point short of its own price. Five dollars, a lean, and an honest note that this is the play I'd most like to be wrong about excluding.
Over forty-six and a half in Baltimore, and it's a lean because the edge is half a point — the model says 47.0 against a 46.5 line, which is not a disagreement, it's a rounding error. It's on the board because Baltimore's input profile is the most distinctive on the slate: 164 rushing yards a game, the highest of any team on this card, against New Orleans throwing for 266. At the regression's weights, Baltimore's ground game alone is worth 11.6 points before anything else happens. Priced at plus one-oh-two, so it needs 49.5 and my published number is 48.2 — it does not clear, and it says so. Five dollars, printed so you can see what the model was looking at rather than only what it liked.
“Bet less than you brag, learn more than you lose, and never trust a tout who smiles too much.”
— The Betting Fool
Disclaimer: The Betting Fool is for entertainment purposes only. Nothing on this site is financial advice, and past performance does not guarantee future results. Must be 21+ to wager where legal. If you or someone you know has a gambling problem, call 1-800-GAMBLER.