온라인 기반 부동산 심리지수(Real Estate Sentiment Index)의 개발과 아파트 매매가격 반응 분석- 머신러닝 예측과 Shapley Additive Explanations 해석을 활용한 접근 -

This study develops an online real estate sentiment index (RSI) using unstructured textual data and compares its explanatory power with that of the traditional survey-based index (CSI) in predicting apartment prices in Korea. The analysis employs monthly data from January 2016 to December 2024. Sentiment classification is performed using the KR-FinBert-SC model, built on KoBERT. The RSI is constructed by reclassifying neutral documents based on softmax probabilities and computing the difference between the number of positive and negative documents. The empirical methods employed include univariate regression, Granger causality tests, random forest modeling, Shapley additive explanations (SHAP) analysis, and co-integration-based regression. Results indicate that the RSI significantly outperforms the CSI in both explanatory power (R²=0.4359 vs. 0.0079) and predictive accuracy. Granger causality tests show a statistical relationship significant with a 2‒3 month lag. The RSI also exhibited leading behavior at all 16 identified price turning points, including those driven by interest rate shocks. SHAP analysis confirmed the RSI’s nonlinear influence and structural contribution to price prediction. Overall, the RSI demonstrates strong potential as an early warning indicator, a behavioral policy tool, and a complement to macroeconomic variables by providing a real-time, emotion-sensitive measure of market sentiment.