{"type": "article-journal", "title": "Recency Bias in Google Trends: A Fitting-Power Criterion for Cropping the Recent Tail", "author": [{"family": "Ashrafnejad", "given": "Arash"}], "issued": {"date-parts": [[2026]]}, "container-title": "Journal of Seasonality", "publisher": "ArtaQuest Foundation", "URL": "https://artaquest.org/papers/google-trends-recency-bias.html", "abstract": "Google Trends' most recent observations are the least reliable - sampled from incomplete data and revised later. We quantify how much recent data a time-series model should discard using a pure in-sample fitting-power criterion: crop the most recent C months of data, refit on the remainder, and record R2, sweeping C from 0 to 6 years with no hold-out or test set. Across 309 monthly series the median R2 rises from 33% (no crop) to about 55% by a knee near 12 months, then continues to climb gently to 65% by 72 months - with no interior optimum inside the window, so cropping recent data only ever helps within the range tested. The recent year-and-a-half carries the most unfittable noise; the model conservatively drops one year. We give the open data and a one-click reproduction."}