Model Probability & Confidence
Model probability
The model probabilityThe model's calibrated probability that the picked side wins. This is distinct from the market-implied probability — the difference between the two is the edge.is the output of the machine learning model after calibration — the model's best estimate of the probability that the picked team wins the game.
It is distinct from the implied probabilityThe win probability encoded in a betting line. Convert American odds: negative odds → |odds|/(|odds|+100); positive odds → 100/(odds+100). E.g. -150 ≈ 60%, +130 ≈ 43%. in the market odds. The gap between model probability and market implied probability is the edgeThe gap between the model's win probability and the market-implied probability at bet time. Positive edge means the model thinks the outcome is more likely than the price implies..
Calibration
A model is calibrated if its stated probabilities match real-world frequencies. If the model says 60% on 100 games, a well-calibrated model should win roughly 60 of those 100 games.
SharpSDA trains a separate isotonic regression layer on top of the base model specifically to improve calibration. This is what makes the probability useful for Kelly sizing — an uncalibrated model can produce misleading stakes even when the directional pick is correct.
Confidence
Confidence is an internal model score distinct from the probability. Where probability answers "how likely is the picked side to win?", confidence answers "how certain is the model in this output?" — a reflection of how cleanly the input features point to one side.
Confidence is an admin-level metric used for quality control and track record analysis. It does not appear on public pick cards.
Brier score
The Brier score measures probability forecast accuracy. It is the mean squared error between the model's stated probabilities and the actual binary outcomes (0 = loss, 1 = win). Lower is better.
A naive model that always predicts 50% gets a Brier score of 0.25. SharpSDA targets Brier ≤ 0.245 on its holdout test set as a gate before a model version can be promoted to serve live picks.