Election polling has become a critical analytical domain for understanding voter behavior, forecasting electoral outcomes, and guiding strategic political decision-making. However, modern polling environments are increasingly noisy, heterogeneous, and susceptible to methodological inconsistencies, making accurate prediction a challenging task. Traditional statistical and machine learning models struggle to handle irregular sampling patterns, biased polls, and credibility variations across polling agencies. These issues result in unreliable forecasts, especially in volatile multi-party electoral settings. This study utilizes the publicly available Kaggle election-poll dataset, comprising temporally distributed party-wise vote-share estimates collected from diverse polling agencies. The dataset exhibits significant variability in quality, sampling size, and recency, making it an ideal testbed for evaluating robustness-oriented forecasting models. To address these challenges, we propose CICS-DLNet, a deep learning framework that integrates a cryptographic-inspired credibility score with a Transformer–GRU hybrid architecture to enhance temporal modeling and reliability alignment. The novelty lies in the credibility-modulated fusion layer, which dynamically adjusts feature weights based on poll trustworthiness—an approach not explored in prior election forecasting research. The model is evaluated using RMSE, MAE, MAPE, and Anomaly F1 to comprehensively assess predictive accuracy and anomaly robustness. Experimental results show that CICS-DLNet achieves the best performance across all metrics, reducing RMSE by up to 45% compared to strong baselines and significantly improving anomaly detection. The results demonstrate that incorporating credibility-aware learning substantially enhances the stability, interpretability, and accuracy of election-poll forecasting.
Roy et al. (Tue,) studied this question.