DOI: 10.16984/saufenbilder.1844087 ISSN: 1301-4048

Noise-Reduced Artificial Neural Networks for Forecasting Housing Demand in Ankara, Türkiye

Bayram Ali Temel
Show PDF Cite
Reliable forecasting of housing-demand dynamics is important for city-scale planning processes and the development of housing-supply strategies. However, the volatile and nonlinear structure of housing-sales data can limit the predictive performance of conventional forecasting models. This study develops a noise-reduced artificial neural network (NR-ANN)-based forecasting model for Ankara using housing sales, population, gross domestic product (GDP), and housing unit price data for the 2013–2024 period. To reduce the effects of short-term fluctuations in the housing-sales series, moving-average transformations (MA-5 and MA-7) were applied, and their effects on model performance were evaluated. The results show that the MA-7-based (80, 60, 40) NR-ANN model achieved a Nash–Sutcliffe efficiency (NSE) of 0.8933 in the month-based external test. Under a stricter temporal evaluation, leave-one-year-out year-block validation yielded a pooled NSE of 0.7607 and revealed substantial variation in model performance across years. Permutation-based feature-importance analysis indicated that the model predictions were most sensitive to the housing unit price variable. Projections for the 2025–2030 period were evaluated under low, baseline, and high housing-unit-price growth scenarios, together with forecast uncertainty bands and a comparison with a linear-trend benchmark model. The NR-ANN model produced lower errors and higher NSE values than the linear-trend benchmark under both the month-based external test and year-block validation. Overall, the findings indicate that moving-average-based noise reduction can improve ANN forecasting performance under the examined data structure; however, frequency mismatch, interpolation, and interannual performance variability require the results to be interpreted cautiously alongside alternative validation outcomes and scenario assumptions.

More from our Archive