This article investigates the role of economic, financial and demographic indicators in forecasting house prices in Albania. The pool of variables is drawn from empirical studies for advanced and developing countries. To test their importance, we employ the long short-term memory network from the machine learning techniques. As the time span of observations is rather limited, the specification of models is maintained to be parsimonious and sufficient to capture the most important dynamics. As such, we compare the performance of a univariate network with models containing the most related variables such as GDP and actual rental, and then augment them with loans and interest rates, demand from non-residents, unemployment rate, urban population, cost of construction and area of building permits. The forecast ability is evaluated during the 2018-2022 period for horizons at 1, 4, 8 and 12 quarters ahead. Preliminary results suggest that multivariate, theory-driven models can help improve upon forecasts generated from the univariate network, although the latter appear in the top four best models at most horizons. Apart from GDP and rentals, costs of construction and financial indicators are some additional variables in which forecasters may have confidence on when predicting house prices in Albania.

Keywords: Machine Learning, LSTM model, house prices, Albania