Comparative Performance Analysis of Sentiment-Based and Baseline LSTM for Forecasting Indonesian Banking Stocks

Abstract

The retail investor base in Indonesia has been expanding at a quick clip with a growing influence of social media conversations and has made public sentiment a possible predictive signal for stock movements. However, the existing studies on Indonesian banking stocks are highly dependent on single-platform sentiment and heuristic feature selection. The contribution of statistically validated multi-platform sentiment is underexplored. This study aims to compare the sentiment based and Baseline Long Short-Term Memory (LSTM) architecture in predicting PT Bank Central Asia Tbk (BBCA) stock prices. The public conversation from Twitter (X) and Stockbit with 39,942 raw comments gathered throughout 2025 and reduced to 36,512 after preprocessing was classed with a fine-tuned IndoBERT-Tweet model with an accuracy of 82.97% and a macro-averaged AUC of 0.92. Spearman's rank correlation was utilized to validate sentiment features against stock returns and insignificant lagged variations of positive and negative sentiment were removed, while significant contemporaneous positive and negative sentiment were kept. Both LSTM setups were assessed with RMSE, MAE and MAPE with same architectures and hyperparameters. The Sentiment Based model decreased test MAPE from 0.97% to 0.85%, a relative improvement of 12.3%. SHAP analysis verified that the chosen features do add to predictions in directions consistent with the Spearman results. These results suggest that a dual platform sentiment methodology based on statistics gives a meaningful complementary signal for forecasting Indonesian banking stocks. These results suggest that a dual-platform sentiment methodology based on statistics gives a meaningful complementary signal for forecasting Indonesian banking stocks.