DiamondIQ
An MLB predictive-analytics platform that publishes its own accuracy record, read live from production data. We set out to build a model that hit 73% — then proved across roughly 4,900 games that our own target was mathematically impossible, and rebuilt the product around the finding.
./visit --live diamondiq.live ↗- Accuracy, every game
- ~57%
- Games backtested
- ~4,900
- Data sources
- 8
The market runs on numbers that aren't real
Sports prediction services routinely advertise 70–90% win rates. Those figures are marketing artifacts — cherry-picked windows, retroactive selection, undisclosed sample sizes, or counting only the games the service chose to call. That creates a trap for an honest builder: publishing a truthful 57% next to a competitor's fictional 85% looks like weakness rather than integrity. We started where everyone starts, assuming a sufficiently sophisticated model could reach around 73%.
So we measured it properly
Every completed MLB game across the 2024 and 2025 seasons plus 2026 to date — roughly 4,900 games — was pulled from the MLB Stats API and run through a point-in-time backtest. Point-in-time means the model only ever sees data that existed before first pitch. No look-ahead, no hindsight, no quietly excluded games.
We disproved our own target
The best achievable point-in-time accuracy across full seasons was 54–57%. Roughly a quarter of MLB games are one-run outcomes no model can call — that is a property of baseball, not a limitation of the model. For calibration, professional betting markets, the most capitalised forecasters in existence, top out around 58–60% on the same task. The second finding explained years of frustration in one line: every team-quality metric we were tuning turned out to be collinear. We had been tuning one signal four times and expecting compound gains.
The one lever that moved independently
Starting-pitcher ERA gap behaves orthogonally to team quality, and accuracy climbs sharply as the gap widens. Critically, in the widest bucket the better-ERA pitcher's team is home less than half the time — so this is not home-field advantage recounted under another name. It is genuinely new information, and it became the backbone of the rebuilt product.
Two numbers, not one
Rather than inflate a single figure, the product reports two and is explicit about which is which. Every game on the slate gets a prediction — no skipping, no quietly omitting hard games to protect the average, because a service that only predicts what it is confident about has a flattering record and no utility. Separately, a high-confidence tier fires only when the starting-pitcher ERA gap is at least 2.5 and two independent signals back the same side. It is provisional: backtested across three seasons, not yet validated in live play.
No skipping. The durable level, against a 54–57% achievable ceiling.
107/160 across three seasons. Fires on ~2–3% of games. Provisional, not yet live-validated.
A system that scores itself
Predictions are scored against real outcomes automatically, and the accuracy number on the homepage is read live from that ledger — not hardcoded by a marketing team. If accuracy drops, the headline drops with it. When every AI provider is unavailable the system writes nothing rather than falling back to a statistics-only guess, because a missing prediction is honest and a fabricated one corrupts the record permanently. The blended output is more accurate than every single one of its ingredients measured individually, which is the entire justification for ensemble modelling, demonstrated on live production data.
- 8 data sourcesMLB Stats, Statcast, splits, standings, odds, weather, park factors, injuries
- Statistical model50/50 baseline, weighted signal adjustments read live from the database
- AI blendClaude reads the same factor set and returns its own pick and reasoning
- Calibrated pickConfidence means an actual estimated win probability, on a 50–70 scale
- Self-scoring ledgerEvery result recorded, accuracy broken down per signal