The study of the seasonality of a historical series can have the purpose of:
- Simply estimating the seasonal component
- Eliminating it from the general course once it has been estimated
If you have to compare several time series with different seasonality, the only way to compare them is by a seasonal adjustment of them.
There are several ways to estimate the seasonal component. One of these is the use of a regression model using dichotomous auxiliary variables (dummy variables).
Suppose the existence of an additive model without a trend component:
Y(t) = S(t) + r(t)
And suppose we have measured the series on a monthly basis. The dummy variables can be defined in the following way:
- dj(t): 1 if the observation t is relative to the jth month of the year
- dj(t): 0 otherwise
Once the periodic dummy variables have been created, the seasonal component can be estimated using the following regression model:
Y(t) = β1D1 + β2D2 + ... + βnDn + ε(t)
The remaining ε(t) part of the model represents the part of the series not explained by seasonality. If a trend component is present in the series, it will coincide precisely with ε(t).