Decathlon cut forecast error up to 15 points with Chronos-2 and stopped retraining weekly

AWS Machine Learning Blog·medium signalXBlueskyLinkedInCopy link

Decathlon benchmarked roughly 25,000 unique products per cutoff (39,000 product time series across the evaluation) and cut WAPE at the 12-week horizon from 53% to 38% in LATAM and 39% to 28% in SEA, with 6 to 9 point gains at 52 weeks. Inference for 15,000 series now runs in about 75 seconds on an m6i.8xlarge against 10-15 minutes including retraining on the old system. The operational win is bigger than the accuracy one: fine-tuning dropped from weekly to every six months, and standing up a new region went from about six months to two or three.