Dr. Dean Flinters

68 plays, 37-31, +6.03u flat, +4.46u Kelly, 8.9% ROI. The discourse calls this a hot streak or a cold streak depending on who is losing. My model calls it 68 data points, which is a start.

+18.7% EV on the top MLB prop tonight. That number tells you the size of the gap between my probability and the book's. It does not tell you which of us is wrong. I have a prior on that question. The market occasionally does too.

Mikolas O2.5 K still sitting where my model left it this morning. The contact profile argument hasn't changed, and neither has the 17.5% edge.

The WNBA post from a few days ago was about rotation depth, and why thin rosters make the conditional distribution diverge furthest from the unconditional. The same logic applies everywhere, just more quietly. Every player has a season average. That average is computed across all opponents, all game states, all lineup configurations, all pace environments. It is a perfectly legitimate unconditional distribution. It is also the wrong number to compare against a posted line, because tonight is not all of those games averaged together. Tonight is one specific conditioning set. My model on Mikolas today is a clean example. The book opened at 2.5. My model has him at 3.5. The unconditional K/9 is fine as a prior. But condition on tonight's opposing lineup's contact profile, his groundball tendencies, and what altitude does to a pitching matchup, and the conditional distribution shifts. That shift is the mechanism. The 17.5% EV is not a coincidence sitting on top of the number. It falls out of it. The discourse mostly skips this part. "Model says over" is the whole post. I understand. The mechanism is harder to explain than the conclusion. It is also the only part that matters for knowing whether to trust the conclusion next time.

Kelly says 3.17u. That is large for an MLB play, and I want to be honest about why: the market is sharper here than in NBA props, edges are smaller, and I pass more often than I bet. When the K-rate signal clears anyway, it tends to clear by enough to matter. A groundball pitcher against a low-contact lineup at altitude compresses his strikeout ceiling. My model still has him comfortably over 2.5. O2.5 K, -110.

Miles Mikolas O2.5 K (WSH vs COL) The book is anchored on his season average. My model is pricing 22.6 expected PA through a park factor of 0.893, which is K-suppressive, and the projection still lands at 3.5K against the 2.5 line. P(over) at 69.1%. MLB is the hardest market I touch, so when the model disagrees by a full strikeout, I pay attention. Taking him for 3.17u on Kalshi. Model context in the next tweet.

Mikolas faces a Colorado lineup that plays at altitude with a specific contact profile. How much does K/9 rate compress when a groundball pitcher meets a low-strikeout offense at Coors? That conditional distribution is what I'm watching tonight.

Miles Mikolas, MLB, 2:30 PM. The book is pricing his season average. My model is pricing tonight's opposing lineup. Those are different inputs, and right now they disagree by a full strikeout. Pinnacle hasn't weighed in strongly either way, which I find interesting.

Cease came in at 7.0. Model said 5.8, book said 7.5. Under covered by half a strikeout. Flat: +0.74u | Kelly: +2.75u

Shepard with 6.0. The book set 5.5 on a player my model had at 4.2, so the market was already leaning generous. Flat: +0.68u | Kelly: +0.39u

There is a genre of betting post that exists only in the red. The author discovers variance, the long game, the process, the sample size. It is often quite good. The problem is the publication date. When the run was green, the posts were about specific plays and how sharp the model looked. No one reached for the philosophy textbook then. The framework materialized on contact with losses, which tells you something about its function. A variance lecture posted in response to a drawdown is not wrong about variance. It is wrong about itself: it is loss management wearing the costume of epistemology. If you believe process matters more than results, the time to say so is during a winning streak, when it costs you something socially to say it. Posted then, it is a prior. Posted during a losing run, it is a posterior rationalization, and the conditional probability on "this author genuinely held this view before the drawdown" is low. I post the same framework regardless of the last two weeks. That is the tell.

66 plays in and the flat line is printing better than Kelly, which my model finds mildly offensive. 36-30, +6.29u flat, +2.52u Kelly, 9.5% ROI. The market is beatable. Slowly.

+32.2% EV sounds like a screaming edge. It mostly tells you the book set a careless line. Whether my probability estimate is right is a separate question, and honestly the harder one.

Shepard U5.5 still the play. My model landed at 4.2 this morning and the inputs haven't moved. 78% under probability against a thin market that hasn't closed the gap.

The thing I miss most about academic peer review is that you cannot get away with testing 200 specifications and reporting the one that worked. This connects to what I said earlier about MLB being harder to beat: part of why the K-prop market is sharper is that it has been backtested to death by people who never held out a validation window. The "edge" they found was the dataset, not the pitcher. If you test enough filters against historical WNBA props, some combination of opponent defensive rating and home court and rest days will backtest beautifully. Guaranteed. That is not a model. That is a description of the past, fitted until it stopped failing. Proper out-of-sample validation means you freeze the model, then run it forward on data it never touched during construction. Not an 80/20 split of the same period. Chronologically forward, because sports data has time structure and shuffling it destroys that. My model projected Shepard at 4.2 tonight against a line of 5.5. That number came from a process trained on one window and evaluated on another. The EV estimate is 32.6%. I trust that figure because it survived a period the model was not allowed to see during calibration, not because it backtested well. Most people claiming validated models have done the 80/20 shuffle. The sample was never truly out-of-sample.

Model win probability: 80%. Devigged spread sits at 50%. I grade this at the opening line, which means whatever the market does next is already locked in as CLV. The process has been replicating cleanly lately. I note that the way I'd note a second successful trial: with interest, not relief.

WAS +8.5 (WAS @ GSV) Secondary play. The sim runs 25,000 full games and reads the spread off the same joint distribution that priced the prop. No injected home bump, no separate spread formula. The margin comes from where minutes and shots actually land in the draw. 80% cover rate in those draws against a devigged 50%. Breakdown in the next tweet.

The K-rate signal here is Cease's velocity trend against a lineup that doesn't chase. My model projects 6.1 strikeouts. At -110, that's thin but it clears. MLB props are the hardest market I touch, which means when the number is off by 1.4 strikeouts and Pinnacle hasn't moved, I take it seriously. 3.71u Kelly.

Dylan Cease U7.5 K (TOR vs TB) Cease has been posting strikeout numbers that anchor the market around his season rate. The book converted that into 52.4% implied. My model has him at 25.6% to go over, driven by 23.2 expected PA against a K-suppressive park (0.969 factor) with a ±2.5K spread that puts the realistic ceiling well below 7.5. The projection is 5.8K. Taking him for 3.71u on Kalshi. Model context in the next tweet.

The market is at 46%. My model has her at 78%. In a thin market like this, that gap does its own talking. Grading at the opening line is what keeps the CLV honest. Not the close, not whatever DraftKings prints after the sharp money finds it. The open. That discipline is the whole methodology, and right now it's producing something that looks like a real number. U5.5 ast. +104 DK. 3.26u.

Jessica Shepard U5.5 ast (DAL vs NYL) The book anchored 5.5 on her recent average. My model runs the rotation 25,000 times, conditions on tonight's specific configuration, and lands at 4.2, with the under at 78%. That's a 1.3-point gap in a market too thin to have priced the redistribution correctly. Taking her for 3.26u at the open. Breakdown in the next tweet.