Effect of frequency-dependent test length on prediction performance in monthly/quarterly time series analysis

Abstract

While time series analysis is of great importance in predicting the future based on past and present data, it also has a great importance in improving the performance of models used in the same process. In this study, the effect of the test data length selected depending on the frequency length on the performance was investigated by using seven monthly and four quarterly datasets accessed as open-source. The research question investigated in this article is that whether and how the selection frequency-based test length for forecasting time series data, such as monthly or quarterly, are superior to the traditional selection. While statistical-based models such as autoregressive integrated moving average (AUTO.ARIMA), HOLT-WINTERS, Seasonal and trend decomposing time losess (STLF), theta method forecast (THETAF), and exponential smoothing state-space model with Box-Cox (TBATS) were used for time series forecasting analysis, on the other hand, deep learning models such as neural network autoregression (NNTAR), multiplayer perceptrons (MLP) and extreme learning machine (ELM) were used. In addition, the evaluation of forecasting performance relies on widely recognized metrics such as mean absolute percentage error (MAPE) and root mean square error (RMSE). The empirical studies conducted and reported in this article show that test data length selected as frequency multiples for monthly and quarterly time series has a positive effect on performance in forecasting analysis. © 2025 Elsevier B.V.