The Lab · Our method · updated 2026-10-22

Why we stopped posting long shots

On October 17 our feed posted a +650 college underdog as a "19.7% edge." It lost — but that isn't why we changed the rules. Every pick we had ever published was an underdog, and the arithmetic said it would be.

On the morning of Saturday, October 17, our X feed posted this:

NCAAF | Sacramento State ML +650 … Edge +19.7%

Sacramento State lost to Ohio, 48–7.

This article is about why we changed our rules that day. It isn't because the bet lost. A +650 underdog is supposed to lose most of the time, and the long shots we've published are, as it happens, up. We changed the rules because of what the post revealed about how our picks were chosen.

What happened

College football joined our board that Saturday as a trial. It was switched on with the NFL's pick rules — our mistake; the trial was meant to grade games, not publish picks. At 8:00 AM the board published five college picks:

Two won, three lost, +0.39 units on 1.00 staked. The feed led with the biggest number on the card, which is how a 13%-to-win college underdog became the day's headline.

Every pick was an underdog

The college card was the extreme case, not the exception. Here is every pick our board ever published, July through October:

• 14 picks. All 14 at plus money — every one an underdog.
• 8 priced longer than +200. They went 4–4, +1.44 units.
• The other 6, from +110 to +200, went 2–4, +2.76 units.

That isn't a finding about underdogs. It's one line of arithmetic:

EV = (your probability − the price's break-even) × (1 + b)

The second term is the payout multiplier, and it grows with the price. A pick had to clear an EV bar — 3% for baseball — and the disagreement needed to clear it shrinks as the price lengthens:

A flat EV bar is a filter for long prices. Charlotte's "+5.3% edge" was less than half a point of disagreement, multiplied by 11.5.

And the disagreement itself was rarely a forecast. With no signal promoted live, our board's probability is the consensus of the books, fee removed. A "disagreement" was one book's price sitting longer than the rest — a real thing to notice, worth what line shopping is worth, and not a prediction about the game.

Why a winning record didn't change our minds

The long shots are up 1.44 units. That tells us almost nothing. At the board's own probabilities, those eight picks should have won about one and a half times between them; winning four is roughly a one-in-twenty run of luck, and going 0–8 would have been about one in six. It would take hundreds of long shots to tell skill from noise. (How many it takes.)

Even our best grade flatters them. Closing line value on the five college picks averaged +10.8%, with four of five beating the close. But CLV is measured in payout terms, so it runs through the same multiplier: a long price inflates it exactly the way it inflates EV. Sacramento State's price beat the closing consensus by 4.3% even though the market moved away from it before kickoff, because the book we priced it at was an outlier the whole way.

Meanwhile the costs of publishing long shots are concrete:

• Mistakes multiply too. The 7.5× that turned 2.6 points into "+19.7%" turns a 2.6-point mistake into −19.7%. Long prices magnify error exactly as efficiently as edge.
• A pick reads as "bet this." A +650 shot loses about 87 times in 100 when it's priced perfectly. Leading a public feed with one invites people to bet something that will almost always lose, on the strength of a word — "edge" — that overstated it.
• Thin markets disagree more. College football has far more games and far less money per game than the NFL. A book sitting off the field there is more likely stale than smart.

The rule since October 17

• No published pick may be priced longer than +200 or shorter than −200. That caps the multiplier on anything we publish at 3×. (At the other end, −200 means risking two to win one; we don't publish anything shorter than that either.)
• College football is graded on the board but never picked. Every game gets a probability and a grade, in public. None becomes a pick.
• The five college picks stay on the record. They were published; they're graded, wins and losses alike. Deleting them would make the record look better and mean less.
• The board still shows every price, long shots included. It just doesn't call them picks.

What this doesn't tell you

Whether long shots are bad bets. Some are fairly priced, a few may be underpriced. The band is a rule about what we're willing to publish, not a claim about underdogs.

Whether +200 is the right line. It's where the multiplier reaches 3×, and we chose it on the day; we haven't proved it's better than +150 or +250. Like the 3% floor, it's a judgment that the record will test in public.

Whether the picks inside the band are any good. Six picks went 2–4 for +2.76 units — all of it, and then some, from one +115 winner. That's not a sample; it's an anecdote with a unit count. When nothing clears, no pick is the honest output.

Model output is informational and entertainment content, not betting or financial advice. If you bet, bet what you can afford to lose.

← Back to The Lab · How the model works · The record