Deep Learning-Driven Sentiment Analysis for Electoral Outcome Prediction

Abstract

Social media platforms offer a unique, real-time window into public opinion, creating novel opportunities for measuring electoral sentiment. This study employs a mixed-methods approach to critically evaluate the predictive power of X (formerly Twitter) sentiment analysis for electoral forecasting, using Nigeria’s 2023 presidential election as a case study. We analysed a dataset of 136,500 tweets and conducted semi-structured interviews with 15 domain experts. Long Short-Term Memory Recurrent Neural Network (LSTM-RNN) was benchmarked against traditional classifiers, Naïve Bayes, and Support Vector Machines (SVM). The optimised LSTM-RNN model achieved superior performance (93.8% accuracy, 95.0% F1-score). However, a moderately strong yet statistically insignificant correlation with official election results underscores the limitations of sentiment analysis as a standalone predictive tool. Qualitative insights contextualize this finding, highlighting the volatility of online discourse, challenges of multilingualism, and critical issues of representativeness, including urban bias and the influence of misinformation. While demonstrating the considerable promise of deep learning, our results reveal significant pitfalls, advocating for hybrid predictive frameworks that integrate real-time sentiment tracking with demographic weighting and multimodal data validation. Beyond forecasting, this research underscores the utility of social media analysis for promoting governance and participatory democracy in multilingual and resource-constrained contexts.