Abstract
To investigate the impact of real estate market sentiment on demand forecasting, this paper constructs a Weibo sentiment index incorporating emotional polarity and verifies its predictive advantage for market demand. Based on the Bidirectional Encoder Representations from Transformers - Bidirectional Long Short-Term Memory (BERT-BiLSTM) and the characteristics of China's housing market, we classify sentiments in crawled Weibo texts and train a sentiment analysis model specifically for the Chinese real estate domain. This model accurately extracts positive, neutral, and negative sentiment features to build a high-frequency sentiment index. Simultaneously, Internet Concern index is constructed using Baidu search data as a non-directional sentiment proxy variable. Further adopting the Autoregressive Distributed Lag Mixed Data Sampling model (ADL-MIDAS), we compare the predictive performance of these two sentiment indices alongside macroeconomic variables on market demand. Experimental results show that: (1) The BERT-BiLSTM model achieves 78.5% accuracy in sentiment classification, with its F1-score outperforming traditional methods (SVM, LSTM, etc.) by over 30%; (2) The Weibo sentiment index yields a Root Mean Squared Forecast Error (RMSFE) of 1.6%-4.7% under the ADL-MIDAS framework, significantly lower than the Internet Concern index (6.7%-7.0%). The study demonstrates that integrating deep learning with high-frequency social media sentiment indiex containing emotional polarity can more effectively capture market expectation fluctuations, while simultaneously yielding superior performance for real estate market demand forecasting.