Software Failures Forecasting by Holt - Winters, ARIMA and NNAR Methods
Vitaliy S. Yakovyna, Bohdan Uhrynovskyi, Oleksandr Bachkay · 2019 IEEE 14th International Conference on Computer Sciences and Information Technologies (CSIT) · 2019
This paper describes the results of software failures prediction using regression (Auto-Regressive Integrated Moving Average, ARIMA), smoothing (Holt - Winters) and neural networks (Neural Network Auto-Regressive, NNAR) models. GitHub service was used as a data source, and the failures time-series of the Kubernetes project was chosen for this study. Both cumulative and noncumulative, 7- and 14-days' time series were used. The best forecasting accuracy was reached using the NNAR method for noncumulative 14-days' time-series.