network interconnected devices has brought about a large amount of easy-to-access time series data, and more and more companies are interested in mining this data, thus obtaining valuable information and making corresponding data decisions. As a data-based company, Esensoft has been working for years to find how to manage growing data and help partners make predictive analytics of the market.
What is time series prediction?
Before understanding the time series prediction, look at the following figure:
This is the prediction of the Dow Jones Index in June 2012 (red line marking). This is a typical application scenario for time series prediction.
time series is a data form related to time changes that are often encountered in real life. For example, Beijing's PM2.5 index, stock price of a certain stock, daily sales of a certain product on Taobao, etc. Time Series Analysis The purpose of is to mine the information and patterns implicit in the time series, and to evaluate the sequence data and predict the subsequent trend of the series.
Simply put, time series prediction is to find the data change pattern in the historical time series to predict the trend of time without occurring .
Time series prediction method
Time series can have several forms: continuous sequences arranged by minutes, hours, days, quarter, month, and year, and the sequence is a numerical sequence, not a nominal sequence. For example, the daily air quality, the net profit of each quarter, and the annual sales, there should be no order or overlap between them. It should be narrated strictly according to the timeline. According to its characteristics, the
time series has the following manifestations and produces corresponding prediction methods:
- Long-term trend changes: is affected by some basic factors, and the data changes according to time as a certain tendency, which steadily grows or decreases according to certain rules. The analytical methods used are: moving average method , exponential smoothing method , model quasi-sum method, etc.
- Seasonal periodic changes: is affected by factors such as seasonal change, and the sequence changes regularly according to a fixed cycle, also known as commercial cycle. Analytical method used: seasonal index.
- cycle changes: cycle fluctuations change.
- Random changes: sequence changes caused by many uncertainties. The methods used include AR, MA, ARMA models, etc.
data mining platform WonderDM has built-in Holt-Winters (Cubic Exponential Smoothing Algorithm ), which is suitable for time series prediction of seasonal period changes. In addition, through the built-in R algorithm package, the ARIMA algorithm can also be used to realize the time series prediction of random changes.
time series prediction steps
2017 has passed. Students who care about national economic development must be looking forward to China's growth in GDP in 2017. Next, we will introduce in the data mining platform WonderDM to use the Holt-Winters algorithm to predict how my country's GDP performance in 2017 is to see if it matches the actual situation.
1, prepare data
We collected data sets of China's GDP from 1990 to 2016 from the Internet. Explore the data change trend through graphical Linear graph:
From the above figure, we can see that GDP has an exponential growth trend with year, and there is a certain regularity.
2, training model
first create a mining process, select the time series Holt-Winter, and enter the mining process interface.
We select the prepared " China GDP" dataset, select the YEAR_ field as the timestamp field and the GDP_ field as the prediction field, because there is not too much data in our dataset, set all data to participate in the prediction (the maximum number of rows in the training data is greater than the actual number of rows). Because GDP is per year as one cycle, the seasonal cycle is set to 1. The number of forecast periods is 1, and only 2017 is predicted. The timestamp format is year yyyy. After setting up, click the "Train Model" menu to view the trained prediction model.
As shown in the figure above, the system predicts that in 2017, my country's GDP will be 11.43 trillion US dollars. The solid line on the figure is historical data, and the dotted line is prediction data. It can be seen that since 2006, the predicted data has been lower than the actual data. At the same time, we can see that the average absolute error of the predicted value is 22.95%, which is a bit too large, but the confidence interval does not. All of these means that the forecast GDP data for 2017 may be inaccurate.
requires us to adjust the parameters to make the prediction value more accurate. Click on the advanced parameter panel of model parameter , and adjust the following parameters:
- horizontal smoothing coefficient is used to balance the fluctuations in the horizontal direction. If the value is centered, it means that the influence of the closer and farther observations is balanced;
- trend smoothing coefficient is used to balance the fluctuations in the trend. The smaller the value, the smaller the trend fluctuations;
- seasonal smoothing coefficient is used to balance the fluctuations in the season. The larger the value, it means that some predictions of the current season are only based on the most recent observations. The three coefficients of
are valued at 0-1. The larger the value, the greater the influence weight. We adjusted these three coefficients to 0.4, 0.5, 0.9 respectively. Let’s take a look at the adjusted prediction effect.
You can see that the historical prediction value after adjusting the parameters is closer to the historical actual value, and the average absolute error is reduced to 6.65%, and the confidence interval is also within a reasonable range. Therefore, our final forecast of GDP in 2017 is US$11.9 trillion, and the possible GDP range is between US$11.19 trillion and US$12.71 trillion, which is very close to the official announcement of US$12723.8 billion, and the forecast is accurate.
time series prediction allows us to gain insight into future trends through historical data. Time series analysis is commonly used in macro-control of the national economy, regional comprehensive development planning, enterprise management, market potential forecast, meteorological forecast, hydrological forecast, , earthquake precursor forecast, crop disease and disaster forecast, environmental pollution control, ecological balance, astronomy and oceanography. After reading this article, are you starting to be ready to predict tomorrow's stock price?
