2026-10-14T15:39:28Z https://api.figshare.com/v2/oai

oai:figshare.com:article/34048438 2026-10-01T17:33:29Z category_15 category_734 category_931 category_811 category_64 category_132 portal_5 item_type_1 month_year_10_2026

<p>Comparison of now-casting model performance and bias-variance trade-off.</p> Sara Mesquita (17691590) João Loureiro (480403) Cláudio Haupt Vieira (25157779) Lília Perfeito (20553240) Joana Gonçalves-Sá (3173976) Neuroscience Biological Sciences not elsewhere classified Information Systems not elsewhere classified Mathematical Sciences not elsewhere classified Cancer Infectious Diseases xlink "> disease integrating digital epidemiology consistently outperforming models combines model evaluation casting increasingly relies apply simple supervised feature curation informed feature selection unsupervised methods unstable performance training periods traditional data results suggest online activity often noisy offer fast method offers inexpensive insights increased stability improve prediction hybrid approach frequently suffered different flu curated subset behavioral noise behavioral insights behavioral insight behavioral confounders adaptable framework 2009 h1n1 19 datasets <p>Predictive performance of Linear Regression (circles) and Random Forest (squares) using different search-term clusters (All, C1, C2, C3) during the 2009 Flu and COVID-19 pandemics. Panels a), b), and d) correspond to flu, while c) and e) to COVID-19. For flu, error bars represent standard deviation across multiple seasons (P1–P6), reflecting model variability. Panel a) highlights C1’s superior performance over “All”. For COVID-19, standard deviations are not shown due to the rolling window approach, which results in overlapping training data and non-independent <i>R</i>² estimates. Instead, panel c) presents the <i>R</i>² of the mean prediction across models. Panels d) and e) illustrate the bias-variance trade-off. Random Forest has lower variance but slightly higher bias than Linear Regression. In flu (d), C1 reduces variance compared to “All”, suggesting that selecting fewer, more relevant terms improves stability. In COVID-19 (e), bias remains high due to the pandemic’s unpredictability. While here C1 also lowers variance relative to “All”, this effect is weaker than in flu, where seasonal patterns allow for more consistent predictions. LASSO shows high bias and low variance.</p> 2026-10-01T17:32:53Z Image Figure 10.1371/journal.pone.0358063.g002 https://figshare.com/articles/figure/_p_Comparison_of_now-casting_model_performance_and_bias-variance_trade-off_p_/34048438 CC BY 4.0