2026-11-22
The Latest in Forecast Benchmark Hill Climbing
Let’s check in on the fev-bench leaderboard. It’s been three weeks since I wrote about Timestral, t0-gamma, and TimesFM-3 in The Latest in Forecast Benchmark Hill Climbing. Since then there have been two additions to the very top of the leaderboard, ci-proto and TimeH2O-v1. Additionally, Simons-297M comes in behind Timestral and t0-gamma but ahead of TimesFM-3, and iAmTime-base-forecast comes in behind TimesFM-3 but ahead of Chronos-2.
Interestingly, the now ranked-10th-overall Toto-2.0-2.5B continues to come in first for 12 out of the 15 Boomlet tasks using scaled quantile loss. Making model choices for and training on domain-specific data continues to be useful (the Boomlet tasks being derived from Datadog’s observability time series on which Toto was trained).