Özet:
BACKGROUND: In forecasting housing-unit prices, the conventional
hedonic model is the most common used method, and the newer artificial neural
network (ANN) models also have been used in forecasting various, different
economic and financial variables. More recently, combining forecasts models have
been introduced to enhance forecasting accuracy. The current study aims in
developing a more accurate forecasting model by combining the hedonic and neural
network forecasts of housing-unit prices.
MATERIAL AND METHODS: A total of 100 apartments in Istanbul,
Turkey were included in the study. Housing-unit characteristics were taken including
the price, the geographical location, the land size, the age of apartments, the number
of bedrooms and bathrooms, and the floor within the building. First, the hedonic and
ANN models are applied, and their forecasts are compared to detect the better model.
Second, combining models are generated by combining the forecasts of hedonic and
ANN model using different sets of forecast weights, generated by restricted and
unrestricted, weighted least squares (WLS) regression technique, respectively.
Average absolute forecast error (MAFE) of each model is calculated, and the average
difference in MAFE among all pairs of models are compared and tested, and the
superior model is the one with the lowest average absolute forecasting error (MAFE).
RESULTS: The study finds that between the ANN- and hedonic models, the
ANN model performs better. However, the ANN model was outperformed by the
combination forecast formed with restricted WLS estimated regression coefficients
as component forecast weights. In all, seven combining models were generated using
different methods in calculating the forecast weights. The unrestricted combining
iv
models were outperformed by the ANN model, while the constrained combining
models performed better than ANN model. The study finds that the restricted
combining models have the lowest MAFEs and are considered as the superior
models.
CONCLUSION: The present study successfully generates combining
forecasts models from the housing units‟ forecasts of the hedonic and neural network
models. The study finds the combining forecasts model formed with weights
generated by constrained WLS regressions generally perform the best out of all other
forecasts‟ models. Our study demonstrates that combining forecasts can improve
estimations of housing units‟ prices in Istanbul, Turkey.
