Time series forecasting was for a long time based on the principle that simple methods in forecast accuracy exceed machine learning methods. However, simple models cannot use the various mutual dependencies and information offered by time series, content-related or similar to those that are subject to prediction. The occurrence of massive data is also connected with the collecting of enormous amounts of time series, but in using simple, traditional methods, their high potential for enhancing forecast accuracy remains untapped.
The opportunity to bridge this gap comes with neural networks. To process sequential data, recurrent neural networks are used in forecasting that can use mutual dependencies of points in time. Among recurrent neural networks, long short-term memory neural networks are considered as especially successful in time series forecasting. The paper focuses on the building and optimization of this neural network type. Our purpose was to improve forecast accuracy in time series forecasting, understanding at the same time why and to what degree individual factors contribute to this improvement.
The forecasting was applied to the number of clicks on Facebook ads. First, we analysed various combinations of time series processing, discovering that they might influence forecast accuracy improvements. Neural networks learned using a group of related time series and the forecasts were compared to the traditional time series forecasting approaches ARIMA, ARIMAX and VAR. We also researched a number of options to improve forecast effectiveness with neural networks using similar time series.
As expected, we found that with adequate data processing, long short-term memory neural networks achieve greater forecast accuracy compared to the traditional models. We demonstrated that forecasting can be further improved using similar time series, however, this approach is not always helpful.
|
