Bayesian Real Estate Intelligence

Probabilistic modeling for multi-portal real estate market analysis
Models · Results · Quick Start · Tractability

---

A probabilistic programming framework that applies hierarchical Bayesian models, Gaussian Process spatial regression, and mixture-model anomaly detection to real estate listing data scraped from four Spanish portals.

Built on PyMC 5, ArviZ, and nutpie (Rust-based NUTS sampler).

Models

#	Model	Technique	What it does
1	Hierarchical Pricing	Multi-level partial pooling	Estimates price drivers per portal while sharing statistical strength across all four portals via group-level hyperpriors
2	Spatial GP	Gaussian Process, Matern-5/2 kernel	Learns a continuous price surface over geographic coordinates with calibrated uncertainty — no hand-crafted spatial features needed
3	Anomaly Detection	Bayesian mixture model	Classifies each listing into "normal market" vs "anomaly" components, yielding a posterior probability of being mispriced

Mathematical Detail

Hierarchical model — partial pooling across $J$ portals:

$$\alpha_j \sim \mathcal{N}(\mu_\alpha, \sigma_\alpha), \quad \beta_j \sim \mathcal{N}(\mu_\beta, \sigma_\beta)$$ $$y_i \sim \mathcal{N}!\left(\alpha_{j[i]} + \mathbf{x}_i^\top \beta_{j[i]},; \sigma\right)$$

Spatial GP — Matern-5/2 covariance over lat/lon:

$$f \sim \mathcal{GP}!\left(0,; \eta^2 \cdot k_{5/2}(d / \ell)\right), \quad y_i \sim \mathcal{N}(f(\mathbf{s}_i) + \mathbf{x}_i^\top \beta,; \sigma)$$

Anomaly mixture — two-component on price residuals:

$$z_i \sim \text{Cat}(w), \quad r_i \mid z_i{=}k \sim \mathcal{N}(\mu_k, \sigma_k)$$

Results

Hierarchical Shrinkage

Portal-level intercepts pulled toward the group mean — portals with fewer listings borrow more strength:

Group-Level Posteriors

Posterior distributions for the hierarchical hyperparameters:

Anomaly Detection

Mixture-model identifies overpriced/underpriced listings with calibrated anomaly scores:

Tractability Analysis

Comparison of MCMC (NUTS via nutpie) vs Variational Inference (ADVI), plus sampling efficiency (ESS/s):

Quick Start

git clone https://github.com/gilito11/bayesian-realestate.git
cd bayesian-realestate
pip install -r requirements.txt

# Quick demo (~6 min, synthetic data)
python demo.py --quick --no-spatial

# Full run with all three models (~20 min)
python demo.py

# With real data from PostgreSQL
python demo.py --source neon --database-url $DATABASE_URL

CLI Options

Flag	Description
--quick	Fewer MCMC draws for faster iteration
--no-spatial	Skip the GP model (slowest due to O(n^3) scaling)
--n-listings N	Number of synthetic listings to generate (default: 800)
--source neon	Load real data from Neon PostgreSQL
--output-dir DIR	Where to save plots (default: output/)

Architecture

bayesian_realestate/
├── models/
│   ├── hierarchical.py    # Multi-level partial pooling across portals
│   ├── spatial.py          # GP spatial regression (Matern-5/2)
│   └── anomaly.py          # Two-component Bayesian mixture
├── data.py                 # Synthetic data generator + Neon DB loader
├── diagnostics.py          # R-hat, ESS, divergences, model comparison
├── plots.py                # Publication-quality visualizations
└── demo.py                 # Full pipeline entry point

Tractability

Each model includes a tractability analysis comparing inference methods:

• NUTS (No U-Turn Sampler) via nutpie — exact posterior samples
• ADVI (Automatic Differentiation Variational Inference) — fast approximate posterior
• ESS/s (Effective Sample Size per second) — sampling efficiency metric
• R-hat convergence diagnostics and divergence counts

The GP model demonstrates the tractability/expressiveness trade-off: Matern-5/2 gives rich spatial structure but scales as O(n^3), requiring subsampling for large datasets.

Tech Stack

Component	Technology
Probabilistic programming	PyMC 5.x
MCMC sampler	nutpie (Rust) with PyMC fallback
Posterior analysis	ArviZ
Data	pandas, NumPy
Database	PostgreSQL (Neon serverless) via psycopg2
Visualization	matplotlib, seaborn

Data Sources

The framework operates in two modes:

1. Synthetic (default) — Generates realistic listings across 8 zones on the Tarragona coast with known ground truth anomalies. Useful for validating model recovery.
2. PostgreSQL — Connects to a live database of listings scraped from habitaclia, fotocasa, milanuncios, and idealista.

License

MIT

Author

Eric Gil — BSc Computer Science, Universitat de Lleida