Forecasting Daily Online Attention to K-Pop Artists: A Retrospective Comparison of Baselines, Ridge Regression and Random Forests

Abstract

Daily public attention measures offer a low-cost basis for audience monitoring, but forecast improvements must be assessed against strong simple methods and with chronological information boundaries. This study compares persistence, seasonal naive and moving-average baselines with ridge regression and random forests using English Wikipedia user-classified pageviews for BTS, BLACKPINK, TWICE, Stray Kids, SEVENTEEN and aespa. The primary dataset contains 10,404 artist-day observations from January 2022 to September 2026. Parameters were selected using eight expanding quarterly validation blocks in 2024–2025. Retrospective evaluation used 1,638 artist-days in January–September 2026, with models fixed before the first forecast origin and lagged inputs updated as observations became available. Log-target ridge regression achieved a mean absolute error of 512.50 pageviews, compared with 603.18 for persistence, a 15.03% reduction (paired 14-day block-bootstrap 95% interval: 6.14%–21.51%). The development-selected log-target forest achieved 551.13 pageviews, an 8.63% reduction (−2.95%–18.20%). Both improved the point-estimate error for all six primary artists, although monthly gains were uneven. Forest error was stable across five random seeds; removing the latest count raised MAE by 33.50%. Only 26 large-traffic artist-days accounted for 38.33% of forest absolute error. Seven-day forecasts were less accurate, and transfer to two artists excluded from fitting did not establish a reliable advantage over persistence. These findings support transparent, lightweight forecasting for aggregate attention monitoring while showing that nonlinear models need not outperform a well-specified linear comparator. Pageviews measure article visits, not unique fans, music consumption or campaign effects.