How NotaSportsGuru picks every leg

Data sources

Three feeds, in order of weight. The Odds API supplies the consensus market price across major US sportsbooks for spreads, totals, moneylines and player props on NBA, MLB and NHL. Official league box-score data (NBA, MLB, NHL) provides the per-player and per-team rates the model regresses against the matchup adjustments. ESPN team and player feeds backfill recent-game form, injuries and lineup status. The combined input is what makes every prediction a comparison between a market-implied probability and a model-implied probability — the edge.

How the model thinks about a single prop

Take a player point prop, say Aaron Judge over 1.5 total bases. The model starts with Judge’s rolling 15-game total-bases rate. It scales by the opposing pitcher’s allowed total bases per appearance. It adjusts for ballpark factor (pitcher-friendly versus hitter-friendly), weather (wind, temperature) and lineup spot. The output is a projected distribution from which the model reads the probability of going over 1.5.

That model probability is then compared to the book’s implied probability from the posted line. If the model probability exceeds the implied probability, the bet has positive expected value (EV). The percentage-point gap is the model’s edge.

Why a human reviews the slate

Pure model output can be wrong for reasons the model cannot see — a last-minute lineup change, a manager’s public comment that a starter will be limited, a known umpire bias. Every morning before publication a human reviewer verifies that the top-edge legs do not have a disqualifying news angle. About 5–8% of would-be picks are removed at this step.

The review is documented: every removed pick is logged with the reason. The audit trail is intentional — model-only picks would be cleaner to defend, but history shows the human catches enough that the trade-off favors keeping the review.

What is public, what is members-only

Public: the player prop and match line picks themselves (the stat, the line, the side), the league coverage, the team and player profiles, and the audited ledger. The transparency here is deliberate — anyone can verify the track record at any time.

Members-only: the model edge percentage and confidence score on every leg, the daily Parlay of the Day before publication, and the dashboard with per-pick reasoning. Edge and confidence are the values that justify a real stake — they stay private because membership is capped and shared signal degrades the edge for everyone.

How the ledger gets graded

Every published pick is graded automatically from the official box score the morning after the game. Wins, losses and pushes update the ledger in real time. The ledger has been continuous since the project started in 2023, with every entry timestamped, line-recorded and outcome-attributed. There is no edit history because there is nothing to hide.

ROI is reported as the cumulative profit divided by cumulative stake. Hit rate is reported alongside the average price on the bets that produced it, so the two numbers can be read together.

What the model does not do

• It does not bet during live games. Live markets carry too much vig and require an execution stack that adds complexity without obvious long-run edge.
• It does not chase steam. If the market has already moved on a leg before publication, the model rejects the pick rather than betting at a closed price.
• It does not hedge for the user. Hedging is a decision for the bettor based on bankroll, not a default. The hedge calculator lives in the tools section.
• It does not take sportsbook affiliate revenue. No kickbacks, no referral codes, no recommended-book ranking. Pricing is the membership fee, and that is the only revenue line.

Where to read the math

The strategy guides explain how each of these ideas works in general — value betting, EV, bankroll management, line shopping, closing line value. The glossary is the one-paragraph reference for every term used here. The tools let you run the same calculations the model uses.