Generalization Limits in Noisy Time Series Returns

Machine learning models are often evaluated under the assumption that additional complexityyields improved predictive performance. However, financial time series pose unique challengesdue to non-stationarity, heavy-tailed noise, and adaptive market behavior. In such environments,forecasting models may struggle to generalize beyond historical data.This paper examines whether increasing model complexity meaningfully reduces forecast errorin financial time series. Rather than proposing a new model, we focus on understanding whycommonly used forecasting approaches often fail to outperform simple baselines. The analysis isframed through the bias–variance tradeoff, a central concept in statistical learning theory.