Calibrated. Tracked. Transparent.

EdgeK scans every MLB starting pitcher every game day and compares the model's strikeout projection against every sportsbook line. When the model thinks the implied probability on a line is wrong by 5% or more, that line shows up on the Value Bets tab.

Three scans a day (10am, 1pm, 4pm ET) so we catch lineup announcements, line movement, and late soft prices. Every bet gets logged. Every result gets settled. Every aggregate ROI number on this site is a sum of those real, dated, time-stamped bets.

Beyond the bets, every starting pitcher gets a public prediction on the Predictions tab: predicted K count, 80% confidence interval, talent tier, and how confident the model is to within ±1 and ±2 strikeouts. No edge, no problem: you can still see what the model thinks. Yesterday's predictions (with actual Ks and per-game MAE) live on Yesterday's Results.

Strikeout counts per pitcher per game follow a Negative Binomial distribution (same family as Poisson) but with the spread tunable per pitcher. A finesse arm with a tight K distribution gets a different α (dispersion) than a high-variance power pitcher.

For each pitcher-day, the model predicts a mean K count (μ) and looks up the pitcher's volatility (α). Together those give us a full probability distribution over how many strikeouts they're likely to throw. That becomes the bet probability for any given line, plus the 80% confidence interval you see on the Value Bets table.

Predicted probabilities get a final Beta calibration pass to keep them honest (with a 0.93 cap to avoid overconfident extreme bets), and the prediction intervals get a conformal recalibration against the full backtest so when we say "80% range," empirical coverage actually hits 80%.

Here's the honest version most services won't tell you: we don't claim to predict strikeouts better than the market. Nobody reliably does. Sportsbooks set strikeout lines about as sharply as any model can — our average miss on a pitcher's K count is statistically tied with the closing line, and that's the ceiling, not a shortcoming.

The edge isn't in the center of the prediction — it's in the shape of the distribution and the price. We model each pitcher's full range of outcomes with their own volatility, then bet only when the odds on offer disagree with that distribution. Two models can agree a pitcher averages 5.5 Ks and still disagree on the odds for the over — that gap, not the average, is where the value lives.

We used to lead with closing-line value (CLV) as proof of that gap. We stopped: on player props, a headline CLV number is mostly the sportsbook's vig, not real edge — strip the vig out and our number was much smaller than what we'd been showing. Rather than keep a metric that flatters itself, we cut it. What we show instead is settled ROI on real, dated, time-stamped bets — slower to look impressive, harder to fake.

EdgeK reports model output. It is NOT financial advice, gambling advice, or a guarantee of future returns. Past performance — even calibrated past performance — doesn't promise tomorrow's results. Bet within your bankroll, set limits, and call 1-800-GAMBLER if betting becomes a problem.