Coverage is easy to advertise. Accuracy isn't. Below is how close our berth-wait forecasts land against real berthings — per port, recomputed monthly, straight from the backtest. Numbers here render from the public /v2/meta/validation endpoint.
| PORT | ATRACAI P50 | PORT-MEDIAN BASELINE | CALLS | VERDICT |
|---|---|---|---|---|
| Santos | 75.2 | 91.9 | 2,404 | BEATS |
| Paranaguá | 38.5 | 38.8 | 346 | BEATS |
| Vitória | 49.3 | 50.6 | 115 | BEATS |
| Itaqui | 85.9 | 99.5 | 120 | BEATS |
| Rio Grande | 59.2 | 60.8 | 62 | BEATS |
Time-based train/test split (never random for time-series). "Verdict" = beats the naïve baseline a customer could compute for free.
Our 80% band should contain the actual wait ~80% of the time. At Santos it currently holds 73% — we report the gap honestly and widen the band where it's short.
We do not publish numbered FIFO queue positions (Santos FIFO tested worse than a coin flip) or hour-of-day claims where timestamps look rounded. If a model fails its gate, it does not ship.
2026-10-30 23:20 UTC
Predictions refresh every 30 minutes; accuracy is recomputed against live actuals and republished monthly.
*Measured coverage by port (STS 73%, ITJ 87%), published live at /accuracy. The point isn't that we're perfect — it's that we tell you exactly how imperfect, in public, before you pay.
Request access