Changelog

What has changed, month by month. Kept short and honest — improvements and the numbers behind them, not internal noise.

Updated September 2026

September 2026

Next-port accuracy raised to 77.9% events-weighted (from around 76%), through new prediction layers that use the declared destination and the origin port together, learn aliases for messy declarations, and weight recent history more heavily.
Port View: each arrival now shows an expected stay duration and a departure window, not just an arrival time.
A historical accuracy figure is now shown in the app alongside the live-scored one, so a model improvement is visible immediately rather than after weeks of new predictions resolving.
Port View layout reworked for readability — fewer columns, each carrying one idea.

The Port View launched — a live arrivals list for any of roughly 460 ports, each vessel with its load state, likely need, contact and onward move.
Distances across fuel, ETA and the map switched to real sea routes, following shipping-lane density, replacing straight-line estimates with a flat mark-up. This corrected a systematic over-statement on many Baltic–North-Sea legs.
Second- and third-port-ahead prediction rebuilt to return a ranked top-three, which covers roughly half of second-port calls.
Arrival and stay estimates made load-aware — a laden vessel sails slower and stays longer, and the timing now reflects that.
Back-testing moved to the server, making per-vessel history and accuracy views fast.

The fuel and cost model was calibrated against a real reported consumption figure, replacing an assumed value that ran high.
Contact and technical enrichment added, using grounded web search that returns cited candidates for a human to confirm.
Port-name normalisation and declared-destination parsing overhauled, recovering a large amount of previously unmatched history.