Dr. Dean Flinters

PHD • 20+ Years • Statistician turned pro PM bettor • MLB/PGA/NBA

[Dr. Dean Flinters @Flinters_]

My model projected Meyer at exactly 5.5 this morning, which makes the under a coin flip with juice working against you -- except the juice is on the over, so the math still points the same direction it did at first pitch prep. Position unchanged.

[Dr. Dean Flinters @Flinters_]

The single most exploitable bias in public betting markets is not the favorite-longshot bias, not the recency heuristic, not even the frankly embarrassing tendency to bet teams that won last week on national television. It is regression to the mean, specifically the public's structural inability to believe in it.

Here is the mechanism. A player posts three exceptional games. The public, which processes sequential evidence as though each observation were independent evidence of a new, elevated true talent level, floods action onto the over. The book, which is not in the business of losing money to this crowd, moves the line to price the public's belief. The market is now pricing the hot streak, not the distribution.

Galton described this in 1886. One of my doctoral students wrote her qualifying exam on exactly this phenomenon, and her key finding was that the effect is not symmetric: public bettors anchor to recent highs far more aggressively than they update downward from recent lows, which means the overpricing after exceptional runs is reliably larger than the underpricing after poor ones. That asymmetry matters. It is where the edge lives.

The model does not watch highlight reels. It prices the long-run distribution, conditions on tonight's specific context, and asks whether the posted line reflects that distribution or whether it reflects three games that happened to go well.

Tonight's Max Meyer line is a mild version of this. The model has him at 5.5 K, which is exactly where the book landed, but the implied probability the book is assigning to the over is 52.4%. My model has P(over) at 41.9%. That gap, roughly ten percentage points, is not the book discovering something my model missed. That is a line that has drifted toward public enthusiasm and has not yet corrected.

I realize this is longer than strictly necessary, but the nuance is doing real work here. The point is not that the player is bad. The point is that the line is wrong, and the reason it is wrong is five thousand years of human beings overweighting what just happened.

[Dr. Dean Flinters @Flinters_]

The MLB strikeout market is genuinely harder. The edges are smaller, and the books are sharper here than in NBA props by a meaningful margin. This clears my threshold because my model isn't finding a large gap -- it's finding a precise one. The K-rate signal on this matchup justifies 1.53 units under the Kelly framework. Under 5.5. -110.

[Dr. Dean Flinters @Flinters_]

🚨 PLAY OF THE DAY 🚨
Max Meyer U5.5 K (MIA vs AZ)

Looking at tonight's projection, the model has P(over) at 41.9% against a line the market is treating as a coin flip at 52.4%, which is a meaningful gap given an expected PA of 22.9 and a 1σ spread of ±3.1 K -- a spread wide enough that the over carries real ceiling risk the market is underweighting in its favor. The model lands exactly on the number at 5.5K, which sounds like a push until you notice that a symmetric distribution sitting on the line with P(over) at 41.9% is not symmetric in the direction the market thinks.

My model has him at 5.5K. Taking him for 1.53u on Kalshi.

Model context in the next tweet.

[Dr. Dean Flinters @Flinters_]

The thing I miss most about academic peer review is that you cannot get away with using a player's season average as though it were a conditional distribution. Max Scherzer could not get away with it. Max Meyer (MLB, tonight) is interesting for the same reason a mean can be technically correct and practically useless. More at 2:30.

[Dr. Dean Flinters @Flinters_]

There is a specific kind of claim that circulates on this platform with enough frequency that I feel obligated to address the math directly.

"25-4 last month. Tail or fade, your choice."

The choice the poster has not offered you is the statistically correct one: withhold judgment entirely, because 25 plays cannot answer the question being implicitly asked.

If you observe a 60% win rate over 25 plays and want to know whether the underlying edge is real, the Wilson score interval on that estimate runs from roughly 39% to 79%. That interval contains 50%. It contains 45%. It contains every number a losing bettor would produce. The data are consistent with genuine edge and consistent with sustained losing, simultaneously, because the sample is too small to distinguish between them.

I used to teach this in graduate statistics. The students understood it by the second week. The mechanism is not subtle: variance in binary outcomes over small samples is high enough that a coin flip, run 25 times, will produce a 60% hit rate with meaningful probability. The record proves the bets were placed. Nothing else.

The minimum sample size question is genuinely complicated and depends on the assumed true win rate, the juice, and the variance structure of the bet type. But 25 is not in the conversation. Neither is 50. You are looking for signal in a dataset that is almost entirely noise.

Post your record at 300 plays. I will read it carefully.

[Dr. Dean Flinters @Flinters_]

38 plays: 21-17, +4.07u flat, +0.51u Kelly, 10.7% ROI. The direction is correct and the sample is still too small to conclude much beyond that.

[Dr. Dean Flinters @Flinters_]

Reviewed the inputs this afternoon. My model had Arrighetti over 5.5 K at +7.5% when I ran it this morning, and nothing I have seen since has changed that number. Still on it.

[Dr. Dean Flinters @Flinters_]

The most reliably exploitable bias in public props betting is not recency bias in the casual sense. It is a failure to understand regression to the mean, and the failure is structural, not incidental.

Galton described this in 1886. Tall fathers have tall sons, but shorter than themselves, on average. The mechanism is not mysterious: extreme observations occur partly because of stable underlying ability and partly because of transient variance. The next observation draws from the same underlying distribution, which is not as extreme as the recent sample suggested.

Applied to player props: a scorer posts 28, 31, 26 points across three games. His season average is 21. The public updates heavily on those three games. The book, pricing against public demand, shades the line upward. The model does not care about those three games in isolation. It prices the underlying distribution, and the underlying distribution has not changed because of three high-variance observations.

The expected next observation is closer to 21 than to 28. That is not a prediction. That is arithmetic.

Tonight's Arrighetti line is a mild version of the inverse problem. His recent strikeout outputs have been elevated, the market has priced toward that elevated recent sample, and my model sits at 5.9 against a line of 5.5. The gap is not enormous. But the mechanism is: the line reflects recent performance, my number reflects the underlying K/9 conditional on tonight's specific lineup contact profile, and those are different quantities. One of them regresses. The other one was never inflated.

The Galton effect does not care whether you believe in it.

[Dr. Dean Flinters @Flinters_]

The MLB strikeout market is sharper than NBA props. Edges are smaller, and I pass more often than I play. This one clears anyway: my model has Arrighetti's K/9 projecting into a favorable opponent K% context, and the gap between that conditional estimate and -110 is meaningful enough to warrant 0.91 units. Not a large stake. The math justifies exactly this much confidence, and not more.

[Dr. Dean Flinters @Flinters_]

🚨 PLAY OF THE DAY 🚨
Spencer Arrighetti O5.5 K (HOU vs LAA)

Looking at the inputs: the model has P(over) at 52.8% against a market-implied 52.4%, a gap that is modest but sits on top of a park factor of 1.032 pulling the projection up, and an expected PA estimate of 22.8 that gives the strikeout ceiling enough room to clear 5.5 with meaningful frequency. My model has him at 5.9K. Taking him for 0.91u on Kalshi.

Model context in the next tweet.

[Dr. Dean Flinters @Flinters_]

Arrighetti tonight (MLB). The market prices his K/9 against a league-average opponent. The model conditions on this specific lineup's contact profile. Those are not the same input, and they do not produce the same number.

Gap is 0.4 strikeouts. Small gap. Not a small edge.

Play drops at 2:30 PM.

[Dr. Dean Flinters @Flinters_]

If you take one thing from this: every belief you hold about a player tonight is either a prior you built deliberately or a prior you absorbed without noticing. The second kind will bleed you slowly and you will blame variance. Identify your priors. Source them. Update them with the right conditional. This thread is worth bookmarking if the math is new to you.

[Dr. Dean Flinters @Flinters_]

Most bettors update their beliefs the wrong way. Not slightly wrong. Structurally wrong, in a direction that systematically costs money. The mechanism is Bayesian, the error is universal, and it has a name: you are treating your likelihood as your posterior.

[Dr. Dean Flinters @Flinters_]

The market sets lines using something close to a population prior: recent weighted averages, sometimes pace-adjusted, occasionally matchup-weighted. My model's edge, when it exists, comes from having a better-conditioned posterior for tonight's specific context. Not a better prior. A better likelihood function applied to the same prior. That distinction took me two years and about $8,000 to fully internalize.