Datqorin / Canada
Evaluate sequence models across seasons
Sequence models such as LSTMs can process ordered observations. Their ability to use temporal information does not remove the need for a realistic forecast design.
Choose the sequence
Define the time interval, variables and forecast horizon. Keep timestamps consistent and document how variable-length or incomplete records are represented.
Test on later or separate seasons
Use a split that reflects the intended deployment. Avoid allowing future observations or closely related records to influence training decisions.
Compare and communicate
Compare with a simpler time-aware baseline, inspect the error distribution and report uncertainty. A successful experiment is evidence about its test conditions, not a promise about a future harvest.