Open Access
Subscription Access
Forecasting Non-Stationary Time Series Method of Allocation Patterns
Importance. This paper proposes a method of forecasting of essentially non-stationary time series, i.e. of series that changing structure, variables or model, the coefficients for the variables. Because of non-stationarity of processes that generate economic indicators, most economic time series falls in this category.
Objective. Currently, structural analysis of essentially non-stationary time series can be represented in two directions: 1) a breakdown of such series on the segments in which the properties of the component little changed, after which the analysis of patterns, one approach which is consistent allocation component of the time series for each segment and 2) identify patterns. Found in the history of areas with similar dynamic characteristics allow to receive quite accurate forecasts of future values of the series. In the work in the second direction was studied algorithms of the method of k-nearest neighbor with constant and variable length pattern and different metrics vicinity.
Methods. To study the variations of the algorithm was built prediction model data with varying period and forecasts of real cyclic series production of electricity from coal and natural gas.
Results. The most accurate forecast model data was built using the algorithm with the conversion of the abscissa, the algorithm with fixed-length pattern gave high errors.
Conclusions and Relevance. When dealing with economic time series change as the model coefficients, so and its structure. Reduces the number of variables and their influence disappears from the model, while others appear. For such series forecasting patterns should be used with transformations on the abscissa.
Objective. Currently, structural analysis of essentially non-stationary time series can be represented in two directions: 1) a breakdown of such series on the segments in which the properties of the component little changed, after which the analysis of patterns, one approach which is consistent allocation component of the time series for each segment and 2) identify patterns. Found in the history of areas with similar dynamic characteristics allow to receive quite accurate forecasts of future values of the series. In the work in the second direction was studied algorithms of the method of k-nearest neighbor with constant and variable length pattern and different metrics vicinity.
Methods. To study the variations of the algorithm was built prediction model data with varying period and forecasts of real cyclic series production of electricity from coal and natural gas.
Results. The most accurate forecast model data was built using the algorithm with the conversion of the abscissa, the algorithm with fixed-length pattern gave high errors.
Conclusions and Relevance. When dealing with economic time series change as the model coefficients, so and its structure. Reduces the number of variables and their influence disappears from the model, while others appear. For such series forecasting patterns should be used with transformations on the abscissa.
Keywords
Variable Pattern Length, Forecasting Non-Stationary Time-Series.
User
Font Size
Information
Abstract Views: 116
PDF Views: 0