Direction accuracy for the 2025-26 season sat at roughly 52% across species — slightly better than a coin flip. That is the honest number, and it is worth understanding what it means and what it does not.
Direction accuracy measures whether the model correctly predicted the direction of change between surveys — up or down. Getting the magnitude right is a harder problem that the model is not yet fully optimized for. A prediction of 50,000 Mallard when the actual count is 54,505 is directionally correct but off by 8%. A prediction of 80,000 when the actual is 20,000 is a different kind of miss.
The model had its best directional accuracy on Mallard at 55.4% and Canada Goose at 53.2%. Its worst was on Greater White-fronted Goose at 42.1% and Bald Eagle at 41.5% — species where the count dynamics are driven by factors the current model does not fully capture.
Snow Goose direction accuracy was 53.7% — roughly where you would expect for a species whose movements are driven by large-scale continental weather patterns that are partially predictable a week out and mostly unpredictable beyond that.
What does the learning loop fix? The model currently uses fixed parameters for arrival rate, departure rate, and weather multipliers. The learning loop — in development for the 2026-27 season — will calibrate these parameters against four seasons of actual survey data, allowing the model to correct systematic biases.
52% is where we start. The goal is 65% or better directional accuracy by mid 2026-27 season.
Next week: The specklebelly puzzle. Greater White-fronted Goose is the most volatile species in four years of data.