Advancing influenza forecasting models by integrating key epidemiological features

Accurate, real-time forecasting of influenza activity is critical for guiding public health responses and operational planning. Despite the development of diverse statistical, mechanistic, machine learning, and ensemble approaches, seasonal influenza epidemics remain challenging to predict due to their complex dynamics. We developed an ensemble model to forecast influenza-like-illness activity in emergency departments in metropolitan France at national and regional scales, addressing three key challenges to improve forecast accuracy: i) incorporating a "holiday effect" into mechanistic models to account for the reduced transmission during school holidays, ii) refining filtering methods to leverage similarities in epidemic onset timing across seasons, and iii) developing (sub)type-specific model variants to capture the co-circulation of multiple influenza (sub)types. In total, we integrated 65 statistical and mechanistic models into 140 ensemble configurations, exploring various ensemble-building strategies. Models were assessed retrospectively through cross-validation across eight seasons, and the best-performing model was further tested on four out-of-sample seasons. Over the validation period, incorporating a holiday effect and accounting for similarities in epidemic onset timing in mechanistic models improved forecast score by up to 13% and 10% respectively. Although ensemble models were not always the top performers, they provided the most consistent accuracy across regions and seasons. The (sub)type-specific ensemble showed no consistent advantage over the non-stratified model in test seasons. The best-performing ensemble model was implemented in real time in 2025-2026 to support public health decision-making, enabling prospective evaluation. It accurately anticipated epidemic dynamics and peak timing at the national level and across most regions, demonstrating strong predictive skills.