Combining deep learning neural networks and spatiotemporal models for prediction and inference of residential house prices for property assessment

Property valuation is a crucial economic service that is used by local governments for the distribution of property taxes in order to fund local services. The current and widely adopted cost approach to valuation is based on estimating the land value and the depreciated cost of the building. This approach relies on third party cost information as well as individual assessor opinion and expertise and is typically enabled by a computer assisted mass appraisal system.
The sales comparison approach to property assessment uses the market to estimate value by comparing the subject to similar properties that have recently sold. This method often involves applying linear multiple regression analysis. We propose to build a machine-learning model to predict sales prices for all residential houses in Nova Scotia, using a database of features for each house. There are many different possible choices of models that could be applied to this problem. We will use two approaches: spatiotemporal modelling, and deep learning. Spatiotemporal modelling directly incorporates two of the most important predictors of house prices, and provides an interpretable model, which may be of assistance to assessors explaining and defending values if challenged in court.

Faculty Supervisor:

Hong Gu;Toby Kenney

Student:

Xinyue Zhang

Partner:

Property Valuation Services Corporation

Discipline:

Statistics / Actuarial sciences

Sector:

Real estate and rental and leasing

University:

Dalhousie University