ANNA4REV

The real estate sector is one of the largest economic sectors. Thus, a reliable real estate valuation plays an important role. To fulfil this role, resilient market values accompanied by statements regarding their uncertainty are needed. Research in the field of real estate valuation focus among others on the development of estimation methods that lead to improved market values and corresponding uncertainties. The sales comparison approach is a widespread valuation method, which is based on the comparison of performed transactions. It uses a behavioural approach to relate target parameters (i.e. the property value) to several characteristic parameters of the real estate. Traditionally, effects and causes of real estate submarkets are modelled mathematically with a linear regression. This method reaches its limits when it comes to assessing large spatial submarkets that exhibit nonlinear dependencies and want to include submarkets with few transactions. Artificial neural networks (ANN) are a fundamental method in the field of machine learning. They can be understood as a non-linear mapping of a set of input variables into output quantities. This mapping takes place in a nested structure of processing units, so-called nodes. The weights characterize the connections between the nodes and are estimated as free parameters as part of a training phase. In this project we employ ANN to establish an non-linear relationship between characteristic and value-impacting parameters of the real estate and its value. Thus, the above mentioned limitations of traditional methods shall be overcome. From a methodological point of view, the estimation of the weights is being further developed in the project. A Kalman Filter (KF) based approach is used for this scope. KF is a sequential estimation method allowing a rigorous propagation of uncertainties and therefore, enabling the statistical assessment of the estimated weights. The overall aim of the project is to introduce ANN with the extended Kalman Filter (EKF) for automated real estate valuation as a hybrid model. For this purpose, sample size, complexity, parameter estimation, acceptance and applicability for real estate valuation including low purchase price locations with ANN+EKF will be researched and further developed. ANN+EKF require large amounts of data. The data scarcity of real estate valuation is countered by data augmentation methods, e.g., aggregation or inclusion of asking prices. Quantifying the bounds of minimal prediction risk requires evaluation of complexity. This motivates the study of the Vapnik- Chervonenkis dimension as a complexity measure. The obtained results are validated mathematically and by experts, and the acceptance of ANN+EKF is analyzed. A special focus is on the applicability for markets with few transactions in Austria and Germany.