Вісник № 06
Постійне посилання колекціїhttps://repository.kpi.kharkov.ua/handle/KhPI-Press/31030
Переглянути
Документ Comparative analysis of univariate time series modeling and forecasting techniques for short-term unstable data(НТУ "ХПІ", 2017) Marynych, Tetyana Olexandrivna; Nazarenko, Lyudmyla Dmytrivna; Khomenko, Nataliya HryhorivnaThe article summarizes the international experience in univariate time series modeling approaches and methodology. It aims to make empirical assessment of their relevance and forecasting power for short sample volatile data with numerous aberrant observations and structural breaks with the help of the time series R packages. The findings revealed the pitfalls of outliers’ neglection including stationarity and model misspecification, biased parameter estimates, deterioration of residuals’ properties and prediction accuracy of the models. Empirical research demonstrated the outperformance of the outlier detection methods versus robust approaches that use smaller weights for aberrant observations. We tested a method of improving the forecasting power of the ARMA models by proper identification of hidden patterns and incorporation of additional information about extraordinary events into the model. We also considered frequency domain and nonparametric methods including exponential smoothing, seasonal and trend-cycle decomposition, structural and neural networks models to make comparative forecasting diagnostics. The findings showed slightly worse accuracy of the exponential smoothing and structural state-space models for short prediction horizons and their outperformance for longer forecasting periods. Neural networks showed outstanding in-sample approximation but poor out-of-sample quality. We recommend further studying of the Bayesian regime switching models that have proven to be a comprehensive way to explore hidden patterns in data, as well as dynamic factor multivariate models that can improve explanatory and forecasting power of the time series models in various applications.