Live odds ingestion
We ingest live odds from multiple sportsbooks on a regular schedule and score every available matchup.
Methodology
How Galaxy Sports Edge actually evaluates a matchup.
No black box. The pipeline is four phases — Ingest, Score, Publish, Calibrate — and every public surface is gated by what the data can honestly support.
- Phase 01
Ingest
The data pipeline runs on a regular cadence (every 30 minutes during peak hours, more often as a slate approaches kickoff). Every odds row is stamped with bookmaker count and a freshness timestamp. If a market is thin, the data-quality score reflects that on the pick card.
Inputs
Live lines from multiple sportsbooks, on a schedule.
Outputs
Normalized markets — spread, total, moneyline — timestamped per fetch.
- Phase 02
Score
The scoring engine computes implied probabilities, weighs sharp line movement, evaluates head-to-head context, and looks at venue form. Each contributor is exposed in the factor breakdown so a serious reviewer can see exactly what moved the dial.
Inputs
Normalized markets + historical team-game logs (after the model-history gate opens).
Outputs
A confidence range, an edge projection, and a risk profile per side.
- Phase 03
Publish
A pick is only published when the public-picks gate is open, and a numeric confidence value is only shown when the calibration gate is open. Until then, confidence is presented as a label (Lean / Strong / Top Pick) and the page surfaces a "collecting baseline data" note.
Inputs
Engine output + the readiness gates.
Outputs
The pick card you see — selection, confidence band, risk, reasoning, freshness, and factor breakdown.
- Phase 04
Calibrate
The model only learns from real outcomes paired with the signal state at the moment a pick was made — never from its own prior reasoning text. When a proposed weight change improves out-of-sample calibration, it lands in a versioned model bump that a human reviews and merges. Customers see the model version on every pick card.
Inputs
Settled outcomes from real games, paired with the engine state at prediction time.
Outputs
A calibration proposal. Not a silent weight change.
The audit trail behind every signal
Every published card is tied back to live markets, timestamped data, factor scoring, and the gates that keep weak signals off the board.
Bookmaker coverage as a transparency signal
Each pick is scored against the bookmakers that had a market for the game at the time of scoring. We surface the bookmaker count as a transparency signal.
Data freshness on every pick
Where available, each pick shows the timestamp of the odds data it was scored against so you can judge freshness for yourself.
Calibrated confidence presentation
Confidence is expressed as a label or score depending on the platform's current confidence-display mode. Numeric scores are only shown once calibrated against settled outcomes.
Until we have enough settled outcomes to calibrate against, confidence is shown as a label, not a number.
Risk level on every pick
Each pick carries a risk level reflecting bookmaker consensus, market depth, and known volatility factors.
Factor breakdown for subscribers
Subscribers with the right entitlement can see a factor-by-factor breakdown of how each pick was scored.
Public performance is gated, not advertised
Public performance statistics are only displayed after the platform has accumulated enough settled, canonical picks to compute them honestly.
When you see win-loss numbers on the Performance page, you'll also see the period, sample size, model version, and the exact win-rate definition.
Readiness gates
The customer surface is gated by what the data supports.
These flags are public-facing on purpose. They're the difference between a tout site and a system you can audit.
- CANONICAL_HISTORY_ENABLEDCanonical history
- Once on, every new pick and team-game log is recorded as canonical — eligible to count toward the public record. Bootstrap-era rows stay flagged and never enter customer-visible stats.
- DERIVED_MODEL_HISTORY_ENABLEDDerived model history
- Once on, the scoring engine starts using head-to-head, venue-form, and ATS-form signals. Only canonical logs feed in — never bootstrap rows.
- PUBLIC_PICKS_ENABLEDPublic picks
- Once on, the /api/picks endpoints return picks publicly. Until then, the picks endpoints respond honestly with a not-yet-ready state instead of fabricated data.
- PERFORMANCE_STATS_ENABLEDPerformance stats
- Once on, the dashboard and the public Performance page display the record and win-rate. Requires at least 100 canonical settled picks before the operator may even consider flipping it.
- OUTCOME_LEARNING_ENABLEDOutcome learning
- Once on, settled canonical picks become eligible data for the next calibration cycle. This gates data collection only — weight changes still require an explicit model version bump and a human review.
The full gate sequence and its prerequisites are documented in the operator runbook in this repo, not just in marketing copy.
See how it looks on a real card.
The Picks page is the same engine, the same calibration policy, and the same disclosure stack — just one slate at a time.
Sports betting carries risk. Only wager what you can afford to lose. If you or someone you know has a gambling problem, call 1-800-GAMBLER.