Using Facebook Prophet for Time Series Forecasting: Predicting Tencent Cloud Database Storage Trends

This article introduces Facebook's Prophet time series forecasting model and demonstrates its application in predicting Tencent Cloud database storage trends.

Prophet is a powerful time series prediction model that analyzes historical data to forecast future trends. The model is particularly useful for data with temporal dimensions, such as weather or stock prices.

The author applied Prophet to predict Tencent Cloud database purchase volumes using data from January to June 2018 as the test set. The results show that as time progresses, the overall estimated storage capacity trends upward, though uncertainty also increases. The model's predictions closely matched actual values in the test set.

Prophet is structured as an additive regression model: y(t) = g(t) + s(t) + h(t) + et

The model consists of three main components:

1. Growth (g(t)) : Fits non-periodic changes using piecewise linear or logistic growth curves. Prophet automatically detects trend changes by identifying turning points in the data.

2. Seasonality (s(t)) : Models periodic changes such as weekly and yearly patterns using Fourier series for annual seasonality and dummy variables for weekly patterns.

3. Holidays (h(t)) : Accounts for the impact of holidays with non-fixed dates on the forecast values.

The practical applications of this model include:

For internal services:

• Identifying potential customers by detecting companies with rapidly increasing database purchases

• Predicting when customers need capacity expansion for better service preparation

For customer services:

• Providing SaaS analytics for customer-selected dimensions

• Analyzing CVM usage to help customers choose optimal configurations