Sub-City Real Estate Price Index Forecasting at Weekly Horizons Using Satellite Radar and News Sentiment

Battle Count: 109

The paper addresses real estate index forecasting, which is relevant to real estate investment trusts (REITs), property-linked derivatives, and macro-level asset allocation. The horizon-dependent modality findings (price history dominates short-term, sentiment/SAR critical for long-term) parallel findings in financial time series. However, the focus is on physical real estate indices rather than tradable securities, and the weekly frequency is slower than typical quantitative trading horizons. The multimodal framework and ablation methodology could inform broader asset price prediction pipelines.

Moderate complexity. Requires integration of multiple heterogeneous data sources (transaction records, satellite imagery via Google Earth Engine, GDELT news API, central bank rates). Feature engineering involves SAR backscatter processing, Sentence-BERT embeddings with PCA, and careful causal lag construction. However, the core models (KNN, Random Forest, XGBoost, Ridge) are standard scikit-learn/xgboost implementations. The main complexity lies in data pipeline construction, index methodology, and the rolling cross-validation framework rather than model architecture.

Code repository and processed features will be released upon acceptance. All raw data sources are publicly accessible (Dubai Land Department Open Data, GDELT, Copernicus Sentinel-1/2, CBUAE). However, the code is not yet available, and some preprocessing steps (e.g., curated whitelist of 80 news sources, manual boundary checks) require replication effort. Hyperparameters are specified for most models.