Modeling Investor Sentiment Dynamics in the Indonesian Capital Market Using IndoBERT-Based Natural Language Processing

Abstract

Social media plays a role in shaping market perceptions. It is important to understand how investor sentiment evolves over time, particularly in emerging capital markets where retail participation grows rapidly. This work presents a framework for modeling investor sentiment dynamics in the Indonesian capital market using IndoBERT, applied to 3,288 Indonesian-language financial tweets collected between March 2021 and March 2024. Continuous sentiment scores are derived from the probability difference between positive and negative class outputs, achieving an overall accuracy of 76.6% with F1-scores of 0.820, 0.725, and 0.652 for Positive, Neutral, and Negative classes, respectively. The sentiment time series is further analyzed using a 7-day moving average and rolling volatility metrics. Three distinct behavioral phases are identified: (i) a predominantly optimistic period in early 2021; (ii) a sustained negative reversal from late 2021 to early 2022, driven by global macroeconomic uncertainty; and (iii) a volatile recovery phase from 2023 onward. Mean sentiment volatility of 0.584 indicates persistent dispersion in investor opinion, with monthly analysis confirming a positive bias punctuated by episodic negative reversals. These findings demonstrate that NLP-based temporal sentiment modeling can capture meaningful behavioral dynamics and provide a foundation for future research linking sentiment indices to price movements.