How Signal Labs models sports picks.
We build daily picks across MLB, NBA, NHL, NFL, College Football, College Basketball, and PGA. Every pick is the output of a documented mathematical process — not opinion. This page explains exactly how.
1. The blend
Every pick starts as two independent probability estimates: our proprietary model and the current market probability implied by sharp sportsbooks. We weight the blend because pure-model approaches over-confide on small samples and pure-market approaches add no edge above the market itself.
The blended probability is then compared against the implied probability of the line we're betting into. The difference is the edge, measured in percentage points (pp).
2. The edge floor
We only publish picks once they clear our edge floor. The exact threshold is calibrated empirically against settled results. Picks with smaller edges may be honest in expectation but do not have enough margin against variance to publish.
3. Diamond tiers
Confidence is expressed in diamonds (◆). Tier maps to edge magnitude:
| Tier | Edge range | Use case | Suggested stake |
|---|---|---|---|
| 3◆ | +3.5 to +5pp | baseline play | 1.0u |
| 4◆ | +5 to +7pp | strong conviction | 1.5u |
| 5◆ | +7pp and up | elite edge | 2.0u |
4. Hard guardrails
Three filters block picks before they can publish:
- Heavy-favorite block: No 5◆ at -175 or worse. No 4◆ at -200 or worse. Vig-adjusted edge math punishes long-odds favorites because of the implied juice; we cap the tier.
- Dog-edge inflation gate: Edges above 6pp on DFS Projection Sources get an extra audit against the histogram. Our model can over-fit on undervalued DFS Projection Sources; the gate prevents publishing on noise.
- Pre-publish consistency audit: Every approved pick is run against its outcome histogram, model probability direction, and edge sign. If any of those disagree, the pick is blocked + admin-alerted.
5. Settlement and public ledger
Every pick is settled at 8 AM ET the day after the game finishes. Results post directly to /track-record as a public ledger — every Win, Loss, Push, and Void recorded with the original tier, edge, opening line, and closing line.
Closing-line value (CLV) is computed against the consensus closing line at first-pitch / tip-off / kickoff. Positive CLV means we got better odds than the market settled on; negative CLV means we bet into a market move against us. CLV is the most honest measure of model quality — it's independent of variance.
6. Data sources
Every projection is cross-checked against at least two independent sources before any pick publishes:
- MLB: MLB Statcast (via pybaseball), MLB Stats API, Retrosheet historical, UmpScorecards umpire data, weather APIs for game-time conditions, our own ballpark geometry + HR factor models.
- NFL + College Football: NFLverse, CFBData (CFBD), ESPN scoreboards, FPI for opponent adjustment.
- NBA + College Basketball: NBA Stats API, ESPN scoreboards, basketball-reference for historical, balldontlie as fallback.
- NHL: NHL Stats API, ESPN scoreboards.
- PGA: External Golf Data for strokes-gained per round, ESPN field updates, course-fit cross-references.
All historical data is backfilled per sport: MLB 12 years, CFB 13 years, NHL 13 years, NFL 11 years, NBA 10 years (via ESPN scoreboards), Golf 11 years target via External Golf Data.
7. Calibration
Models are calibrated against actual outcomes via expected calibration error and proper scoring rules. Calibration files refresh nightly using yesterday's settled picks and outcome data. The historical calibration improves projection accuracy on the next day's slate.
The DFS optimizer's field-aware Monte Carlo sim was calibrated against 137 ingested DraftKings contest CSVs so cash %, top-1%, and top-0.1% percentages reflect actual GPP outcomes — not inflated marketing math.
8. What we don't do
- We don't guarantee winning picks. Anyone who does is lying.
- We don't cherry-pick wins. Every pick is logged the same day, win or lose.
- We don't manipulate the record. picks_log.json is the canonical record — append-only.
- We don't bet for users. We are not a sportsbook. We publish quantitative commentary.
- We don't recommend illegal gambling. Subscribers must verify legality in their jurisdiction. 21+ only.
9. Transparency commitment
If you ever want to know why a specific pick was published — open the pick on the dashboard and click "Why we like it" to see the model probability, market implied probability, edge calculation, tier rationale, and the data sources used. Every pick is traceable from input to output.
For questions about methodology, email hi@signallabsports.com.
Last updated: 2026-05-28 · This page is part of the public methodology documentation. The model code, data sources, and calibration files are internal but every claim on this page is independently verifiable from the public track record.