For every game we estimate the chance the home team covers the point spread. The inputs are team strength ratings (Elo), the betting line itself, rest days, travel distance, indoor/outdoor and weather, division games, and “look-ahead” and “letdown” spots. A statistical model (logistic regression) turns those into a probability. We show the side it prefers and the probability as “confidence.” Most sit between 50% and 58% — close to a coin flip.
Every week we refresh the scores and retrain on all past games. Then a self-check tests several alternative versions of the model on past seasons, always predicting each season using only earlier data, so it can't peek. A new version replaces the current one only if it clears a strict bar: at least 53.5% on the test seasons, statistically significant, and at least 52.4% (the break-even rate) on recent seasons it never saw. If nothing clears the bar, we keep the old model and label picks “Lean,” not “Pick.”
Every published pick is graded in public with a win/loss badge. We also compare our numbers to the closing line, play our own AI entry in the free Pick’Em so you can see whether you can beat it, and flag injuries that could change a pick. Historically this model has landed at or below break-even (about 52.4% is needed at standard −110 pricing). We tell you that up front instead of hiding it.