Researchers Develop Machine Learning Processes for the Automatic Valuation of Real Estate
Procedures currently used in real estate valuation require comprehensive manual data preparation and processing, sometimes also an inspection on site, which makes the entire process time-consuming and expensive.
Together with the company DataScience Service GmbH, researchers of the St. Pölten UAS and the Kufstein University of Applied Sciences are developing multimodal machine learning processes that use various types of data to predict real estate characteristics such as building size and renovation status – fully automatically.
Currently High Costs and High Risks
For now, even modern and automated models for real estate valuation rely on manual data input, which is not only accompanied by high costs but also bears the risk of input errors – or even fraud when incorrect data are entered on purpose in order to fake a higher market value.
Moreover, the data are often derived from a rather small number of partly redundant and incomplete data sources, which limits the content of available information and impedes the subsequent valuation of the property.
Predicting Real Estate Characteristics
In the project IMREA (Intelligent Multimodal Real Estate Analysis), researchers are currently developing multimodal machine learning methods to exploit the complementary nature of real estate-related data.
The innovative power of the project lies in a machine learning-based approach that combines and merges numerous complementary data sources (such as images, floor plans, broker exposés, contractual texts, expert reports and categorical data) with the purpose of automatically predicting real estate characteristics such as the size and properties of a building, its age and condition.
Automatically Generated Reports Replace Time-Consuming Valuation
“At the moment, there is no largely automated approach that takes into consideration the different types of real estate data as complementary factors and automatically extracts the information from different data sources. However, this would considerably increase the information value of every property and allow for automatic valuations that are more precise. According to our estimate, at least half of all real estate valuations could be replaced by automatically generated reports in practice – which would free real estate valuers for dealing with the more complex cases”, explains Matthias Zeppelzauer, Head of the research group Media Computing at the Institute of Creative\Media/Technologies at the St. Pölten UAS.
Reducing Duration and Costs of Credit Checks
The innovative procedure allows for large amounts of metadata to be automatically extracted and is intended to improve the data basis for real estate valuation in future.
This would considerably reduce the duration of the credit check and the valuation costs involved. Furthermore, the automatic analysis methods enable the continuous monitoring of properties, thereby creating the basis for automatic early warning systems for large real estate portfolios.
“Most existing methods either work with one data source (modality) – usually structured data – or model several modalities separately. The main objective of this project is to develop multimodal machine learning processes with the ability to extract real estate-related information and parameters from heterogeneous input data such as text, images and half-structured data, which is supposed to result in more precise and up-to-date automatic real estate valuation models. In addition, we plan to concentrate on learning from partly incomplete data and insufficient modalities as well as the prediction of several attributes at once, also referred to as multi-tasking learning”, states Zeppelzauer.
Competitive Advantages for Banks
“When it comes to banks, IMREA will actually be a game changer in terms of competition – in times of increasing competitive and cost pressures on top of increasing demands from the supervisory authorities, such as the new EBA Guidelines on Loan Origination and Monitoring”, says Wolfgang Brunauer, CEO of DataScience Service GmbH, a leading provider of automated real estate valuation systems. “For real estate experts, this project will bring the decisive competitive advantage.”
The first prototypes should already be made available to the customers of DataScience Service GmbH before the end of this year.
IMREA – Intelligent Multimodal Real Estate Analysis
The project IMREA runs from 1 February 2021 to 31 January 2024 and is financed within the framework of the Bridge programme of the Austrian Research Promotion Agency (FFG). Partners in the project are Fachhochschule Kufstein Tirol Bildungs GmbH (lead) and DataScience Service GmbH.
- ImmoAge – Visual Age Prediction of Real Estate
- ImmBild - Location Assessment by Computer Vision
- InfraBase – Automatic Building Footprint Segmentation