The article deals with issues related to the application of statistical methods used in the valuation process. The proposed algorithm for real estate valuation can be used in the statistical market analysis method in the process of mass appraisal. The algorithm uses a multiple linear regression model. Legal considerations indicate the need for such an algorithm for the determination of the value of representative properties. Due to the large size of the database of comparables, the proposed algorithm can be used only to appraise typical properties. A good statistical model is parsimonious, that is, it uses as few mathematical concepts as possible in a given situation. A model should extract what is systematic in the results observed, allowing for the presence of purely random deviations. The article discusses the basic principles of building a good statistical model. Attention is drawn to the number of market attributes that are entered into the model and the range of their values. As few explanatory variables as possible should be entered into the model to explain the phenomenon under study. Explanatory variables are only those characteristics of the property that differentiate prices in a given market defined and adopted by the appraiser as the basis for valuation. The article highlights the importance of taking into account market changes during the period under study.

The homogenous properties – as flats are – have the set of key features that characterizes them. The area of a flat, the number of rooms and storey number where it is located, the technical state of a building, and the state of the vicinity of the blocks of flats assessed. The database comprises 222 flats with their transaction prices on the secondary estate market. The analysed flats are located in a certain quarter of Wrocław city in Poland. The database is large enough to apply machine learning for successful price predictions. Their close locations significantly lower the influence of clients’ assessments of the attractiveness of the location on the flat’s price. The hybrid approach is applied, where classifying precedes the solution of the regression problem. Dependently on the class of flats, the mean absolute percentage error achieved through the calculations presented in the article varies from 4,4 % to 7,8 %. In the classes of flats where the number of cases doesn’t allow for machine predicting, multivariate linear regression is applied. The reliable use of machine learning tools has proved that the automated valuation of homogenous types of properties can produce price predictions with the error low enough for real applications.