Optimizing Property Valuation with Deep Learning

The Challenge

ImmobilienScout24, Germany's largest real estate marketplace, needed to provide instant, accurate property valuations at scale. Their existing regression models used only structured features (size, location, age) and missed important signals hidden in listing descriptions, images, and market trends.

Our Approach

We built a multi-modal deep learning model that combines structured data, NLP analysis of property descriptions, and computer vision features from listing photos. The model was designed to serve predictions in under 50ms to power real-time pricing widgets.

• Text features — a fine-tuned German BERT model extracts quality signals from listing descriptions (renovation status, premium finishes, neighborhood context).
• Image features — a CNN trained on property photos estimates interior quality, natural light, and maintenance state.
• Ensemble — a gradient-boosted tree combines all feature families with traditional structured data for the final prediction.

The Result

The multi-modal approach improved valuation accuracy by 25% compared to the structured-data-only baseline. The model now scores over 12 million properties monthly at sub-50ms latency, powering the platform's valuation widget used by millions of users.