How the engine works

The pipeline

Every game runs through the same pipeline. Sandman Edge continuously pulls live odds, team form, pace, situational splits, betting splits, star-player status, injuries, and weather (where it matters) into one refreshed data layer. From that, a statistical model produces a baseline projection — built from a 50% starting point plus roughly two dozen weighted factors — before a language model adds qualitative context and returns a structured pick for each market. The result is saved and later graded against the actual outcome, which feeds back into calibration.

The five-anchor score projection

Rather than trusting a single view of a team, the engine projects each side’s score from five anchors and blends them:

Those projections generate a moneyline, spread, and total estimate, which the engine compares against the market line to compute the value on each side.

Different math for different sports

Score distributions vary by sport, so the totals model does too. Basketball totals behave roughly normally, so NBA and NCAAB over/unders use a Normal-distribution model (Normal CDF). Hockey and baseball score in lower-frequency events where variance closely tracks the mean — classic Poisson territory — so NHL and MLB totals use Poisson convolution. Golf is modeled separately at the player level: a strokes-gained model with course-fit adjustments produces win, top-finish, and make-cut probabilities rather than moneyline/spread/total markets.

Calibrated confidence

Each pick includes a confidence number expressing the engine’s calibrated probability for that side. Confidence is graded independently across markets — a high-confidence moneyline read does not automatically raise confidence on the spread or total, because those are different bets with different math. The engine continuously tracks projections against outcomes and recalibrates, so over large enough samples a 65% confidence pick should perform roughly in line with that probability.

The PASS discipline

Many services push action on every game. Sandman Edge doesn’t. When a projection does not clear a sport’s confidence threshold, the engine returns NO PICK — a PASS. A skipped game is the model saying the line looks fair and there’s no meaningful value to target. PASS calls are not counted as losses, because no wager was ever recommended. The same discipline applies to spot-specific traps: when statistical confidence is extremely high on a spread favorite, the engine demands stricter supporting evidence before backing it, because high-confidence favorites cover less often than bettors assume.

Edge detection & closing-line value

When the model’s disagreement with the market is large enough, the pick receives an EDGE tag — those are the spots where the line looks most inefficient, surfaced so they stand out from routine picks. The engine also captures closing-line value (CLV): the gap between the number it picked at and where the market ultimately closed. Consistent positive CLV is one of the strongest long-run signals that an approach is beating the market, independent of any single result.

FAQ