When not to bet

The point estimate is the least interesting output

Every model produces a single number: 61 percent, say. That number is a summary of a distribution, and on its own it hides the only thing that decides whether there is a bet — how wide that distribution is. Two models can both output 61 percent, one meaning "between 59 and 63" and the other meaning "somewhere between 50 and 72". Against a market at 57, the first has a bet and the second has nothing.

Where the width comes from

Sample uncertainty

Models are fitted to finite history. Common game states have thousands of examples and tight estimates. Unusual ones — fourth-and-goal from the seven, down eleven, nine minutes left — have a handful, and the estimate wobbles accordingly.

Input uncertainty

The model does not know everything true about the game. It may not know a lineman is playing hurt, that the wind picked up, or that the bullpen is short after a doubleheader. Each unknown widens the band, and the ones you cannot see widen it most.

Structural uncertainty

The model form itself is a simplification. Reality is not logistic regression. Some of the gap between the model and the truth cannot be closed by more data, only acknowledged.

The overlap rule

Here is the whole decision procedure. Take your band. Take the de-vigged market probability. If the market falls inside your band, there is no bet. Not a small bet — no bet. The disagreement is inside the noise, and acting on it is paying vig to express a coin flip.

Only when the entire band sits on one side of the market price do you have a claim worth staking. Model at 61 with a band of 58 to 64, market at 55: the whole band clears, and there is a real edge to size. Model at 61 with a band of 54 to 68, market at 55: the market is inside, and the correct action is to close the tab.

Why this feels wrong and is right

The overlap rule kills most of your bets. That is the point. In a market with roughly 4.5 percent hold, you need to be right by a meaningful margin just to break even, and genuine mispricings are not common — especially in-game, where the book is watching the same feed you are. A tool that finds an edge in every game has calibrated its confidence to your appetite rather than to reality.

Passing is a position. It has an expected value of exactly zero, which is better than the negative expectation of a bet placed inside the noise band. Over a season, the discipline of taking zero instead of slightly-negative is worth more than any single winner.

Situations where the answer is no before you even look

• You are chasing. Down for the day and looking for a way back is a state, not a strategy. It reliably makes bands feel narrower than they are.
• The data is stale. A feed that has not updated in a minute during a live game is describing a game that no longer exists.
• Rare game states. If the model is extrapolating rather than interpolating, the band it reports is probably too narrow.
• Right after a big play. Markets and models both need a moment. So do you.
• You cannot state the edge in a sentence. If you cannot say "the model says X, the fair market says Y, and the gap is Z", you do not have a reason.
• The stake would hurt. A bet you cannot afford to lose distorts every decision that follows it, including the hedge.

Sizing what survives

When a bet does clear, size it against the conservative end of your band, not the point estimate. A fractional Kelly stake computed from the low edge is smaller and far more robust to the model being somewhat wrong — which it always is, somewhat. Full Kelly on a point estimate assumes your model is exactly right, and no model is.

How to know the rule is working

Log every pass alongside every bet, and record what the market closed at. If your passes were mostly correct — the closing line drifted toward the market and away from your number — the band is doing its job. If you keep passing on games where the close moved decisively toward you, your bands are too wide, and you can tighten them deliberately rather than by feel. Either way, the discipline improves only if you write it down.