A deterministic duel-based imputation framework for reliable AI-driven time-series forecasting

Abstract Reliable operation of AI-driven time-series forecasting systems is often challenged by missing observations, particularly in sensor-based monitoring environments where data incompleteness disrupts temporal continuity and degrades predictive reliability. Although numerous imputation techniques have been proposed, most evaluate reconstruction quality based on statistical similarity rather than their impact on downstream forecasting performance. This study proposes a deterministic duel-based imputation framework that integrates forecasting objectives directly into the imputation decision process. Instead of relying on a single reconstruction strategy, the framework generates multiple candidate imputations. It selects the most suitable values through deterministic pairwise comparisons using a forecasting-oriented composite loss function combining Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE), and the coefficient of determination (R 2 ). The framework is evaluated using the Beijing PM2.5 dataset and tested across four recurrent forecasting architectures: Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), Bidirectional LSTM (Bi-LSTM), and Gated Recurrent Unit (GRU). Experimental results show that the proposed approach reduces MAPE by up to 70% and RMSE by 12–15% compared with conventional imputation methods while achieving R 2 values above 0.95 across architectures. These findings demonstrate that forecasting-oriented imputation can improve temporal structure, reduce error propagation, and enhance the reliability of AI-driven forecasting systems.