Inside Ball Street: How Our AI Finds Edges the Books Don't Want You to See

Every two minutes, our pipeline ingests a fresh snapshot of odds across 90+ sportsbooks on 4 continents. That's not a marketing number — it's the actual count of distinct books in our silver layer. US retail, European sharps, betting exchanges, Australian books. All of them.

Step 1: Ingest Everything

We pull from The Odds API and ESPN, parsing every event, every market, every line. NBA, NHL, NCAAB — moneylines, spreads, totals, and player props. Each snapshot lands in our bronze layer, gets normalized in silver, and feeds the gold detection engine.

Since launch, we've processed over 350 million individual price points. That's not historical data we downloaded — that's real-time odds we've captured, timestamped, and stored.

Step 2: Detect Sharp Movements

Our gold layer clusters simultaneous line movements across books. When 3+ books move in the same direction within a 5-minute window, that's a signal. When one of those books is Pinnacle or Betfair, the signal gets stronger.

We weight each book dynamically based on its historical reliability — how often its moves predict the final line. Sharp books get higher weights. Retail books add confirmation but don't drive signals.

Step 3: Score Across 31 Dimensions

This is where Ball Street separates from everything else. Each detected movement gets scored across 31 proprietary dimensions — signal strength, book consensus, timing, market conditions, and more. We can't disclose the specific dimensions (they're our edge), but the scoring is comprehensive.

Step 4: Walk-Forward Validation

Every rule in our system is trained on historical data, then tested on data the model has never seen — rolling forward week by week. If a pattern worked from January through March, we test it on April. Then May. Then June. If it doesn't hold out-of-sample, it doesn't ship.

We also apply FDR correction (Benjamini-Hochberg) — the same statistical method used in medical trials — to ensure we're not just finding patterns in noise.

Step 5: Fire the Pick

When a live event matches a validated rule at the right time, the pick fires automatically to Discord. No human touches it. The pipeline runs 24/7, ingesting, detecting, scoring, and delivering. Whether we're watching or not.

The show goes on.