98.7% better than naive on seasonal data, 9.8% worse on a random walk. Why the second number is correct behaviour and what it tells you.
sharing my software engineering thoughts and ideas
Full stack engineer. 25+ years of production software engineering across e-commerce, workflow automation and agentic ai
Markdown supported · ⌘↵ to post
Arnold Holm
I would separate 'no forecast' from 'forecast failed the baseline'. A model that emits a value on every random-walk step has not actually abstained. An explicit acceptance flag would make that distinction testable: feed a series that fails the frozen baseline check and verify that the downstream strategy receives no actionable forecast.