8.2 Components of Time Series
There are three primary components of time series data. Each component reveals helpful information about the dataset, increasing your understanding of the data and how to best model it. The components we will discuss are listed below.
- Level
- Trend
- Seasonality
- Noise
Level
Level is the baseline value of the series if it were a straight line. It is the baseline value that the pattern would be if there were not trend, seasonality, or noise. This is what the forecast would be if the value of the outcome variable was stable over time.
Trend
In time series forecasting, trend refers to the overall direction and shape of the data. As you now have a brief introduction to MLR, you know that data can be linear or nonlinear. The trend in a time series dataset is seen as an overall pattern to the data that represents the change that occurs from one observation to the next. The video below demonstrates how to visualize trend in time series data.
Seasonality
Seasonality is seen in time series as a pattern within a given year that repeats itself on an annual basis. Examples of seasonality include quarterly trends, monthly trends, or trends having to do with the actual season of the year (i.e. Fall, Spring, Summer, etc.). Another common seasonal pattern occurs around holidays, such as an increase in spending around Christmas time.
Noise
Noise refers to the random variation inherently present in time series, due to causes that are not accounted for, and perhaps should not be accounted for in your model. Noise is seen in a plot of your model’s residuals. If there is a pattern to the residuals, this means that there is still pattern or meaning that should be modeled. However, if there is no pattern to the residuals, then it is likely noise.
The image below shows a time series dataset, Jobs, broken down into the various components. The top graph is simply the data plotted over time, or a line plot of the time series data. The details of the seasonal ups and downs in the top line do not stand out because of the long length of the line. The second graph shows just the trend of the data, with the seasonal pattern removed. The third graph then shows the seasonal pattern, with the trend removed. And finally, the fourth graph shows the residuals. Do they look like noise to you?