Home
All conferencesWhat NBA Edge Lab is, how it works and where to start.
Where does a simple model disagree with the betting market?
NBA Edge Lab compares bookmaker odds with its own probability estimates for NBA games and player props, and ranks the gaps by how much they can be trusted. It is a research and learning tool: it never places bets and never connects to a sportsbook.
How it works
- Data. Games, results and odds come from a synthetic demo season or, in live mode, from The Odds API and BALLDONTLIE. Loading is explicit, resumable and protects your API credits.
- Market baseline. Odds from several books are converted to probabilities with the bookmaker margin (vig) removed, then averaged into a consensus.
- Model. Elo-style team ratings with home advantage estimate win, spread and total probabilities. Player props use recent per-minute form and projected minutes.
- Blend. The model is mixed with the market. The weight shrinks when the model has little data, so it cannot overrule the market on thin evidence.
- Recommendations. For each option the app computes edge, expected value (EV), fair odds, a capped fractional-Kelly stake and a confidence level, then ranks rows as Top, Good, Marginal or Verify.
- Explainability. The ? icon on every row rebuilds the calculation step by step and explains what it means in practice.
- Backtest and calibration. A walk-forward paper-trading simulation reports ROI, closing line value, drawdown, log loss and Brier score.
What it is
- A transparent, inspectable research model: every number can be traced.
- A way to compare prices across books and spot outliers.
- A paper-trading sandbox to test ideas without risking money.
What it is not
- Not betting advice and not a promise of profit. A large edge is usually a model error, not an opportunity.
- No bet placement and no sportsbook integration. Book links only open the site.
- Injuries and lineups are not integrated: always confirm a player is active.
Where to start
- Data & Providers. In live mode, click Load history once, then Refresh now for fresh odds. In demo mode, data is already there.
- Recommendations. Start with the Top badges, open the ? to see why, and use the book link to check the price at the source.
- Dashboard. Model overview, launch gates and the upcoming slate.
- Backtest. Run it to see how the approach would have behaved on past games, and run the props calibration.
- Compliance and Audit. Public output stays locked until every gate passes; admin actions are logged.
Gamble responsibly. U.S. availability varies by state; 21+ where required. Help: 1-800-GAMBLER.
Model approach
- Market baseline: devigged two-way probabilities.
- Team strength: Elo-style ratings with home advantage.
- Spread/total: normal margin/total distributions.
- Blend: model probability + market probability.
- Risk: fractional Kelly with hard caps.
Launch gates
Public publishing stays locked until every gate passes.
Upcoming synthetic slate
Edges are compared with devigged market probabilities. Demo data is synthetic.
How to use this table
- Set the filters. Min edge and Min EV hide weak opportunities; Market limits rows to moneyline, spread, total or player props; Book keeps only the book offering the best price. Click Refresh to apply.
- Start with the badge. Rows are ranked by trust, not by raw EV. Top = solid edge, several books agree, good confidence. Good = one minor reservation. Marginal = several reservations. Verify = edge above 10%, more likely a model error than an opportunity. Hover a badge to see the reasons.
- Find the bet. Matchup is Away @ Home. Pick and Line show the side and spread/total number being evaluated; Odds are decimal.
- Compare the probabilities. Model is the blended model's chance of the pick winning; Market is the average chance across books after removing the vig (consensus).
- Read the value. Edge = Model − Market. EV = expected return per unit staked. Green means positive, red means negative. Prefer rows where both are positive.
- Check the price. Fair is the break-even decimal odds implied by the model. If the listed Odds are higher than Fair, the price has value. Book is where the best price was found, with the number of books compared in parentheses; the odds shown are that best price.
- Size and trust. Stake is a capped, fractional-Kelly share of a paper bankroll. Conf is how much confidence the model has in the estimate; treat low values with caution.
Player props: Pick shows the player and Over/Under, Line is the stat threshold, and the model projects the stat from the player's recent games and minutes. Injuries and lineups are not integrated: confirm a player is active before trusting a prop. Unders are not offered for low-minute players.
Tip: sort your attention by EV first, then confirm with Edge and Conf. These are research outputs on synthetic demo data by default, not betting advice or guaranteed outcomes. Early in a season the model has few games per team, so its weight is reduced automatically and edges stay small.
Research recommendations
Uses opening odds for decisions, closing odds for CLV, and updates ratings only after results.
Player props model calibration
Checks the props distribution against realized stats for rotation players (walk-forward, line = floor(mean)+0.5). There is no historical prop-odds feed, so this validates the model, not a betting edge.
Live data …
Demo data
Synthetic NBA-like games, odds, and results for product evaluation.
Provider adapters
Requires API keys. Verify licenses before commercial use.
Schema
teams · games · odds_snapshots · predictions · paper_bets · audit_log · settings
U.S. public launch gates
These are product guardrails, not legal advice. Counsel must review state-by-state requirements.