Introduction
STL: Seasonal-Trend decomposition procedures based on Loess
Loess (locally weighted regression)
This method was proposed by Cleveland in the paper "STL: A seasonal-trend decomposition procedure based on loess" in 1990.
The above picture is from the snowdroptulip blog .
Features
STL has a simple design, which includes a series of applications of the loess smoothing method; this simple design allows the analysis of the properties of the process, and can also achieve fast calculations, even for long time series, and a large number of trends and seasons The smoothness of sex can also be calculated quickly.
Other features of STL are:
- Regarding the amount of seasonality and trend smoothing, this is an almost continuous way, from a very small amount of smoothing to a very large amount of smoothing;
- Robust estimation of trend items and seasonal items without being distorted by abnormal behavior in the data;
- The period of the season item can be specified as an integer multiple of the sampling interval arbitrarily greater than one;
- Can decompose time series with missing values;
How to interpret the results of a time series plot
1. Find outliers and mutations
Outliers
Outliers can have a disproportionate effect on time series models and can produce misleading results.
When there is an abnormal value, it is necessary to determine the cause, whether the data input is wrong or true.
Mutations
Find mutations in the sequence or mutations in the trend. Please try to determine the cause of similar changes.
2. Find trends
Trends are long-term increases or decreases in data values. The trend can be linear or it can show a certain degree of curvature. If the data shows a certain trend, you can use time series analysis to model the data and generate forecasts.
3. Look for seasonal patterns or cyclical movements (periodical changes)
The seasonal pattern is the repeated rise and fall of the data value in the same time period. Seasonal patterns always have a fixed, known time period.
On the contrary, cyclic movement refers to the rising and falling data values that are repeated irregularly.
In general, cyclical motion is longer and more variable than seasonal patterns.
4. Evaluate whether seasonal changes are additive or multiplicative
If the magnitude of the seasonal change is fixed, the seasonal change is additive. If the magnitude of the seasonal change increases as the data value increases, the seasonal change is a multiplicative change. The additional variability may make multiplicative seasonal changes more difficult to predict accurately.