How It Works

Why This Page Exists

Most betting services track their record from the day they launched their premium product — not from the day they started publishing picks publicly. The clock starts on their best week.

They measure performance by win rate, which tells you how often they were right but nothing about whether they were right at the right price. A service can win 58% of its bets and still cost subscribers money if the plays were consistently overpriced.

Sharp Retriever was built around a different standard: closing line value. Every play is logged before the game starts. The closing line is recorded after. The gap between them is public. It cannot be edited retroactively. That is the standard a sharp bettor uses to evaluate their own performance, and it is the standard we hold ourselves to.

The Problem

Most betting services sell picks without ever showing you how they arrived at them. No model, no edge calculation, no verifiable track record — just a record that starts counting on their best week. Sharp Retriever was built to operate differently: every play is logged before the game starts, the methodology is documented here, and the results are public.

The Data

Odds are ingested from 15 bookmakers, including the sharp books that set the market — Pinnacle, Novig, and ProphetX. These books don't offer bonuses and don't limit winners, which means their lines reflect informed money rather than recreational action. The gap between a sharp book's line and a square book's line is the clearest signal available that one side has been bet by people who know something.

The Model

Three XGBoost models cover the three primary MLB markets: moneyline, run line, and totals. Each model was trained on Retrosheet play-by-play data from 2015 through 2024, with pitching and batting features sourced from FanGraphs. The models are retrained weekly during the season as new results accumulate. The output of each model is a probability estimate — a fair price for the outcome before the market has any influence on the number.

Edge Detection

A play is published only when two conditions are met. First, the model's fair probability must diverge from the market price by a meaningful margin — not a rounding error, but a gap large enough to cover vig and still leave positive expected value. Second, sharp book consensus must confirm the direction: if the sharp books are pricing the opposite side, the model's signal is not acted on. Both filters must agree before a play is logged.

Sizing

Plays are sized using the Kelly criterion, capped at a quarter Kelly. Full Kelly is theoretically optimal but produces drawdowns that are difficult to sustain in practice. A quarter Kelly preserves roughly 75% of the long-run growth rate while cutting variance significantly. Lower-confidence plays are published as a flat stub — a fixed fractional unit — rather than a Kelly-derived size. Stake recommendations are always expressed in units, not dollars.

Timing

Plays are published at least 60 minutes before first pitch. This is a hard rule. Line movement in the final hour before a game is often driven by sharp action, injury news, or lineup changes — chasing a line that has already moved against you is not a strategy. Publishing early also lets subscribers shop for the best available number at their book.

The Track Record

We published a plain-English guide explaining the thinking behind every layer of this system — from why win rate alone is insufficient to what the live CLV data actually shows. It is the same document we would want to read before subscribing to something like this.