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

Baris Arat; Hasan Fehmi Ates; Emre Sefer

cs.LGFeb 20, 2026

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

Authors: Baris Arat, Hasan Fehmi Ates, Emre Sefer

Organizations: Ozyegin University, Istanbul, Turkey

Abstract

Reliable real estate price indicators are typically published at city level and low frequency, limiting their use for neighborhood-scale monitoring and long-horizon planning. We study whether sub-city price indices can be forecasted at weekly frequency by combining physical development signals from satellite radar with market narratives from news text. Using over 350,000 transactions from Dubai Land Department (2015-2025), we construct weekly price indices for 19 sub-city regions and evaluate forecasts from 2 to 34 weeks ahead. Our framework fuses regional transaction history with Sentinel-1 SAR backscatter, news sentiment combining lexical tone and semantic embeddings, and macroeconomic context. Results are strongly horizon dependent: at horizons up to 10 weeks, price history alone matches multimodal configurations, but beyond 14 weeks sentiment and SAR become critical. At long horizons (26-34 weeks), the full multimodal model reduces mean absolute error from 4.48 to 2.93 (35% reduction), with gains statistically significant across regions. Nonparametric learners consistently outperform deep architectures in this data regime. These findings establish benchmarks for weekly sub-city index forecasting and demonstrate that remote sensing and news sentiment materially improve predictability at strategically relevant horizons.

Figures & tables

Figure 1: Weekly price indices for 19 Dubai sub-city regions (2015–2025, rebased to 100). Regional trajectories diverge substantially from the citywide aggregate (bold).

Figure 2: Overview of the multimodal forecasting framework integrating transaction data, news sentiment, SAR imagery, and interest rates for sub-city price index prediction.

Tag	Included modalities (feature blocks)
P	Prices ( $12$ lags)
PC	Prices ( $12$ ) + Counts ( $12$ )
PCS	Prices ( $12$ ) + Counts ( $12$ ) + Sentiment ( $15$ )
PCSB	Prices ( $12$ ) + Counts ( $12$ ) + Sentiment ( $15$ ) + SAR ( $15$ )
PCSBI	Prices ( $12$ ) + Counts ( $12$ ) + Sentiment ( $15$ ) + SAR ( $15$ ) + Interest Rates ( $12$ )
PCSBIG	Prices ( $12$ ) + Counts ( $12$ ) + Sentiment ( $15$ ) + SAR ( $15$ ) + Interest Rates ( $12$ ) + Global context ( $12{+}12$ )

Table 1: Additive modality tags and included feature blocks.

Figure 3: Spatial distribution of selected Master Projects in Dubai.

Panel of Figure 4 — Figure 4: Quantile trimming diagnostics for transaction-based index construction. Relaxed trimming ([5th, 95th]) improves citywide cycle fidelity relative to a macro benchmark, while more aggressive trimming ([25th, 85th]) yields a smoother regional monitoring signal with reduced sensitivity to idiosyncratic transactions.

Panel of Figure 5 — Figure 5: Sentiment feature diagnostics. (a) Weekly NSI captures market sentiment shifts including the COVID-19 period. (b) Low correlation between lexical NSI features and semantic PCA components confirms complementary information.

Figure 6: Per-horizon KNN performance by modality configuration. Lines show mean MAE across 19 regions. Beyond 14 weeks, multimodal configurations stabilize while price-only models deteriorate. Dashed lines show Naive-12Mean and ARIMA( P ) baselines.

Model	Group	P	PC	PCS	PCSB	PCSBI	PCSBIG
Ridge	Short	1.988 $\pm$ 1.269	2.248 $\pm$ 1.448	2.260 $\pm$ 1.435	2.446 $\pm$ 1.544	2.608 $\pm$ 1.634	2.912 $\pm$ 1.836
	Medium	3.426 $\pm$ 1.704	4.005 $\pm$ 1.968	3.987 $\pm$ 1.829	4.254 $\pm$ 1.818	4.428 $\pm$ 1.818	4.687 $\pm$ 1.820
	Long	4.373 $\pm$ 2.130	5.135 $\pm$ 2.525	5.055 $\pm$ 2.208	5.097 $\pm$ 2.220	5.324 $\pm$ 2.065	5.278 $\pm$ 1.980
RF	Short	2.905 $\pm$ 1.771	2.887 $\pm$ 1.738	2.876 $\pm$ 1.707	2.796 $\pm$ 1.560	2.949 $\pm$ 1.666	2.839 $\pm$ 1.574
	Medium	3.858 $\pm$ 1.884	3.687 $\pm$ 1.829	3.560 $\pm$ 1.821	3.376 $\pm$ 1.686	3.484 $\pm$ 1.773	3.218 $\pm$ 1.557
	Long	4.310 $\pm$ 2.253	3.928 $\pm$ 1.936	3.738 $\pm$ 1.850	3.486 $\pm$ 1.789	3.583 $\pm$ 1.902	3.260 $\pm$ 1.580

Table 2: Mean

\pm

std MAE across regions by horizon group, modality configuration, and model. Values are averaged over regions and horizons within each group, after averaging MAE across cross-validation folds. Groups: Short (2, 6, 10 weeks), Medium (14, 18, 22 weeks), Long (26, 30, 34 weeks). Bold: best configuration per model-group pair. Shaded: best across all models. Naive-12Mean and ARIMA are univariate price-only baselines reported as single summary rows. Modality codes: P=prices, C=counts, S=sentiment, B=SAR backscatter, I=interest rates, G=global context.

Figure 7: Long-horizon MAE distributions across sub-city regions. Horizontal gray lines indicate Naive-12Mean baseline. KNN and Random Forest show the largest multimodal gains with tighter regional dispersion.

Model	Friedman	Significant Transitions (Bonferroni-corrected $p$ )
Ridge	$p=0.00606$	P $\rightarrow$ PC : 0.00483; others n.s.
RF	$p<0.001$	PCS $\rightarrow$ PCSB : 0.00134; PCSBI $\rightarrow$ PCSBIG : 0.00483
KNN	$p<0.001$	PC $\rightarrow$ PCS : 0.00261; PCS $\rightarrow$ PCSB : $5.7\times 10^{-5}$
LSTM	$p=8.82\times 10^{-4}$	PC $\rightarrow$ PCS : 0.00847; others n.s.
XGB	$p<0.001$	PC $\rightarrow$ PCS : 0.00027; PCS $\rightarrow$ PCSB : 0.00210; PCSBI $\rightarrow$ PCSBIG : 0.00586

Table 3: Within-model significance tests across modality configurations (

h\geq 26

). Friedman tests assess overall variation; Wilcoxon signed-rank tests with Bonferroni correction assess adjacent transitions. “n.s.” denotes non-significant after correction.

Figure 8: Modality contributions relative to transaction baselines (KNN). Panels show incremental additions of sentiment, SAR, and interest rates. Negative $\Delta$ MAE indicates improvement; largest gains arise from combined sentiment and SAR.

Figure 9: Ablations excluding regional price history (KNN). (a) $\Delta$ MAE relative to price-only baseline. (b) $\Delta$ MAE relative to full multimodal configuration. Sentiment and SAR provide value primarily when combined with transaction data.

Sentiment ( PC $\rightarrow$ PCS )
Feature Variant	Mean $\Delta$ MAE	$p$ -value
NSI only (lexical)	$-0.124$	$0.0348$
PCA only (semantic)	$-0.193$	$0.0069$
Full sentiment block	$-0.264$	$0.0047$
Built environment ( PCS $\rightarrow$ PCSB )
SAR backscatter	$-0.429$	$1.8\times 10^{-4}$

Table 4: Representation ablations for sentiment and built-environment modalities (KNN,

h\geq 26

). Reported values are mean

\Delta

MAE relative to the no-modality baseline for each transition, with one-sided paired tests.

Figure 10: SAR and NDBI comparison across three Dubai subregions (2016 vs. 2025). SAR backscatter (left columns) shows consistent sensitivity to built-up development, while optical NDBI (right columns) exhibits greater variability.

Figure 11: Out-of-sample 34-week forecasts for the Dubai Marina price index (2022–2025). The price-only model displays higher volatility and delayed trend response, while the multimodal configuration produces smoother forecasts that track realized values more closely.

Time Series Forecasting

Time series forecasting predicts future values from ordered observations using statistical, neural, and probabilistic models to support decisions under uncertainty.

See all 858 papers

S2S-JEPA: Predicting the Predictable at Subseasonal-to-Seasonal Timescales
Oct 2, 2026Chenyu Dong, Gianmarco MengaldoWeather Forecasting Spatiotemporal Forecasting
Department of Mechanical Engineering, College of Design and Engineering, National University of Singapore, 9 Engineering Drive 1, 117575, Singapore.
Earth Embeddings Reveal Diverse Urban Signals from Space
Apr 3, 2026Wenjing Gong, Udbhav Srivastava, Yuchen Wang +6Geospatial Representation Learning Remote Sensing Benchmarks
Texas A&M University, College Station, TX, USA. · The University of Alabama, Tuscaloosa, AL, USA. · Emory University, Atlanta, GA, USA.
HeatCast: A Benchmark for Neighborhood-Scale LST Forecasting across 124 U.S. Cities
Aug 7, 2026Jesus Guerrero, Isaac Corley, Leon Najafirad +2Remote Sensing Spatiotemporal Forecasting
University of Texas at San Antonio · San Antonio, TX, USA · Taylor Geospatial +1

Oct 2, 2026J/K move · Enter open · S save

S2S-JEPA: Predicting the Predictable at Subseasonal-to-Seasonal Timescales

Chenyu Dong, Gianmarco Mengaldo

Department of Mechanical Engineering, College of Design and Engineering, National University of Singapore, 9 Engineering Drive 1, 117575, Singapore.

Weather Forecasting Spatiotemporal Forecasting Time Series Forecasting Joint-Embedding Predictive Architecture

Open paper AskPDF arXiv

The subseasonal-to-seasonal (S2S) timescale, roughly from two weeks to two months ahead, is a critical forecast window for sectors such as agriculture, energy, and water management. Yet, it is widely known as the `predictability desert'. Recent AI weather models excel up to two weeks ahead but deteriorate beyond, largely because they are trained to predict fine-scale details that are neither predictable nor essential at S2S timescales. We argue that a more physically grounded objective is to forecast only the slowly varying components that remain predictable. Computer vision reached the same conclusion with the Joint-Embedding Predictive Architecture (JEPA), which predicts in latent space, discarding unpredictable details. In this work, we introduce S2S-JEPA, which brings the JEPA paradigm to S2S forecasting. It is tailored to this task through design elements from state-of-the-art AI weather models. S2S-JEPA achieves comparable skill to the gold-standard ECMWF physics-based ensemble and surpasses it on multiple metrics at weeks 5 to 6.

Abstract

Time Series Forecasting

Explore similar work

S2S-JEPA: Predicting the Predictable at Subseasonal-to-Seasonal Timescales

Earth Embeddings Reveal Diverse Urban Signals from Space

HeatCast: A Benchmark for Neighborhood-Scale LST Forecasting across 124 U.S. Cities