Missing data present a persistent challenge in time-series forecasting, often degrading model accuracy and compromising decision-making reliability when not handled effectively. Classical imputation methods apply fixed global rules that fail to adapt to dynamic temporal patterns. At the same time, advanced deep learning imputers offer higher accuracy but impose substantial computational overhead and require extensive tuning. To address these limitations, this study introduces Sherwood Duel Optimisation (SDO). This socio-inspired dynamic imputation framework models gap-filling as an iterative duel between exploration and exploitation. By evaluating multiple candidate imputations and selecting the value that best aligns with local temporal behaviour, SDO provides adaptive, context-aware imputations without the complexity of deep generative models. SDO is evaluated on a real-world PM2.5 time-series dataset using four forecasting architectures: LSTM, PSO-LSTM, A-LSTM, and PSO-A-LSTM. A comparative analysis of classical, machine-learning, and deep interpolation methods shows that SDO consistently improves forecasting accuracy across MAPE, RMSE, and R 2 metrics. The SDO-Mean variant achieves the most stable performance. Significance testing confirms that SDO’s improvements are both statistically and practically meaningful. Additionally, SDO offers competitive computational efficiency, making it suitable for real-time or resource-constrained applications. These findings highlight the potential of dynamic, socio-inspired optimisation for advancing imputation in modern analytics workflows. Beyond methodological contributions, SDO provides practical value for decision-support systems in air-quality monitoring, healthcare analytics, energy forecasting, and other domains where reliable predictions depend on complete, contextually accurate data. • Present a duel-inspired method that dynamically imputes missing values in time series. • Demonstrate how socio-inspired optimisation improves accuracy with lower cost. • Show consistent performance gains across error, fit, and efficiency metrics. • Integrate seamlessly with advanced forecasting models in multiple domains. • Enable real-time decision-making in healthcare, energy, and financial systems.
Utama et al. (Sun,) studied this question.