Time series forecasting algorithm: scenarios and parameter configuration-Simple Log Service(SLS)-阿里云帮助中心

Introduction

The forecasting algorithm is based on the Prophet model, which decomposes time series data into trend, seasonal, and holiday components. Each component is fitted and forecasted independently, then combined into a final prediction. Prophet uses a linear or logistic function for the trend component, a Fourier series for the seasonal component, and an indicator function with a Gaussian distribution for the holiday component. For more information, see Forecasting at scale.

Scenarios

The forecasting algorithm uses online machine learning to forecast each metric per entity. It supports the following general time series forecasting scenarios:

Forecasting machine-level metrics, such as CPU utilization, memory usage, and disk read and write speed.
Forecasting business metrics, such as QPS, traffic volume, success rate, and latency.
Forecasting golden metrics. For example, golden metrics that describe the request quality of a domain name include the average response latency, number of requests, and number of failed requests per minute.

Parameter configurations

Configure the algorithm in the Algorithm Configurations step of the Create Time Series Prediction Job wizard. For more information, see Create a time series forecasting job.

Parameter	Description
Time series prediction period
Period configuration	The period of the time series, measured in days. Use decimals for periods shorter than a day, such as 2.4 days. Time series forecasting jobs by default consider the impact of Day, Week, and Year periodicities on the trend. You do not need to configure these periodicities.
Holiday configuration
Country	Select the country where your data originates. The job considers that country's national holidays in its predictions.
Other holidays	Specify any other holidays or event dates that may affect the time series trend. Include holidays from both the historical training data and the future prediction period.
Prediction configuration
Length of series to predict	The length of the future time series you want to predict. The time unit affects the number of data points in the forecast. For example, to forecast time series data for the next 2 hours, if you set Forecast Length to 2 hours, the forecast contains only 2 data points, which is one data point per hour. If you set Forecast Length to 120 minutes, the forecast contains 120 data points, which is one data point per minute.
Confidence Level	The confidence level for the prediction, from 0.5 to 0.99. A higher value widens the range between the upper and lower bounds, increasing the probability that actual values fall within them.
Number of samples	The number of samples to use, from 0 to 100. A higher value improves the accuracy of the predicted upper and lower bounds.
Prediction frequency	How often the job runs to generate a new prediction.
Observation period	The duration of historical time series data to use for each prediction run.
Scheduling Settings
Time Range	The start time for the time series forecasting job.