Internet traffic prediction model

The main methods for predicting traffic in telecommunications networks are the methods and techniques of Machine Learning (ML). As a rule, within the framework of the MO approach, the network traffic predictor is considered as a tool that uses one way or another accumulated statistics over time to draw conclusions about the future behavior of network traffic [4]. However, as the analysis of the literature shows, many modern MO tools, primarily neural networks, do not work efficiently enough due to the pronounced non-linearity of traffic changes and non-stationarity. Among the tasks of forecasting, the task of predicting signs of increments (direction of change) of the process of time series is singled out separately. The article proposes to use some results of the theory of random processes for a quick assessment of the predictability of signs of increments with acceptable accuracy. The proposed fast prediction procedure is a simple heuristic rule for predicting the increment of two adjacent values of a random sequence. Knowledge of the laws of time series probability distributions is not required. The connection with this approach for time series with known approaches for predicting binary sequences is shown. The possibility of using the experience of predicting the absolute values of traffic when predicting the sign of the change is also considered.