Methodology

1. Reference data

The system aggregates prices and signals from a curated set of references chosen for accuracy, not visibility. We don't publish the list — that's intentional, see §7.

If you want to know whether a specific book was covered for a given pick, the answer is in the transparency log. Every published pick lists the books that priced it at publication.

2. EV calculation

EV%    = (p_true × decimal_odds − 1) × 100
p_true = 1 / no_vig_implied_odds

True probability is obtained by removing the bookmaker margin from sharp consensus lines at the moment of market open, before late action displaces them. Default: multiplicative devig method. Logarithmic and Shin variants are exposed under advanced filters.

The EV figure shown next to each pick is the edge against the best price available at publication — not the price you'll see at click time. Lines move.

3. CLV (Closing Line Value)

CLV% = (bet_odds / close_odds − 1) × 100

CLV measures whether you beat the line at close. A bettor with persistent positive CLV has a real edge regardless of short-term win rate. Target: CLV > +1.5% averaged over 500+ settled bets.

CLV is the only metric we trust for long-horizon evaluation. Hit rate, profit, and ROI on small samples are noise.

4. Sharp signal

Three independent inputs blended into one normalized score (0–100):

• —Directional agreement across our reference books
• —Reverse line movement (line moves against the public)
• —Steam detection (synchronized moves across multiple books within a narrow window)

Score > 55: sharp side aligns with model. > 75: unanimous. < 40: not published.

5. Publication threshold

EV    > +2.0%
Sharp > 55
Books ≥ 3 offering the price

Parlays are not algorithmically recommended. Combined edge degrades multiplicatively and most parlay structures hide correlation that the math doesn't catch.

6. Known limitations

• —Closing-line data carries ~30s latency from the underlying market.
• —Injury and lineup data is pulled from public feeds. Late scratches and warmup decisions may not propagate in time.
• —Free accounts see odds with a 15-minute display delay.
• —The model does not estimate correlation across legs in custom parlays.
• —Small-sample warning: EV figures over fewer than 200 settled bets are statistically meaningless.

7. Why we don't publish the source list

• Operational

  Naming the reference books invites them to throttle or block our access. The stack was assembled to survive that pressure — naming it defeats the point.

• Honest

  The value isn't which sources we use, it's how they're combined, devigged, and timed. A vendor list wouldn't help you reproduce the signal.

• Boring

  A public source list reads like a brochure. We're shipping a tool, not selling logos.