Residual Property Prediction Tool

Summary

A public-facing property valuation and analysis tool hosted at sokrates.is/predictions designed to demonstrate Sokrates’ AI capabilities to the Icelandic market. It utilizes a multi-layered residual model to decompose real estate prices into structural, spatial, temporal, agent-specific, and presentation-driven components.

Details

The Residual Property Prediction Tool is a strategic marketing asset for Sokrates, designed to capture the attention of Icelandic SMEs and executives by providing high-utility data on the local real estate market. The tool moves beyond simple price estimation by providing a “residual decomposition” of a property’s value, allowing users to see exactly which factors are driving a price up or down.

Model Architecture

The system employs a stacked modeling approach where each layer captures a specific dimension of value:

• Structural Baseline: Fundamental attributes including square meters, room count, property type, and construction year.
• Spatial Premium: Neighborhood-level effects and geographic desirability within specific postal codes.
• Temporal Curve: Seasonality and broader market temperature/inflation trends.
• Agent Effect: The measurable impact of specific real estate agents or agencies on the final sale price relative to the baseline.
• Presentation Residual: The “alpha” generated by listing quality, including description sentiment, photo quality, and listing strategy.

The core predictive models utilize gradient boosted trees for high-speed inference (sub-millisecond), while the presentation layer incorporates advanced NLP.

Data and Training

The tool is powered by a dataset synthesized from two primary sources:

1. fastinn.is: Approximately 739,000 listing records, including asking prices, descriptions, and metadata, scraped via residential proxies.
2. Official Registry: Approximately 62,000 registered purchase agreements dating back to 2007, providing actual sale prices.

These datasets are joined via the fastanúmer (property registry number). For the semantic optimization of listing descriptions, the project utilizes a Viking 7B Icelandic language model. This model is fine-tuned using Direct Preference Optimization (DPO) on pairs of descriptions ranked by their “presentation residual” score, effectively teaching the AI to write descriptions that statistically correlate with higher sale prices.

Marketing and Strategy

The tool is hosted on the main sokrates.is domain to consolidate SEO authority. The business logic follows a “recursive marketing” loop: the prediction tool is a product of Sokrates’ AI, and the tool itself is optimized by an autonomous research agent (autoresearch), serving as a live proof-of-concept for the company’s “outsourced AI department” value proposition.

While the data is sourced from fastinn.is, the project maintains a policy of “partnership over permission,” with plans to engage the platform owners directly to frame the tool as a complementary service rather than a competitor.

Related

• Viking 7B
• fastanumer
• Eidos
• sokrates.is
• Basis Genesis Engine