Beyond Aggregation - Real-Time Analytics - page 486

forecast(1.0); }
public double update( double y) {
double e = error(y);
level = level + trend + a*e;
trend = trend + a*b*e;
return e;
}
The seasonal component of this model was later added in what has become
known as the Holt-Winters forecast. There are two types of seasonal models
used in Holt-Winters: additive and multiplicative. Both require two extra
parameters: a seasonal period s and a smoothing parameter g . In this
example,thelengthoftheperiodispassedinwiththeinitialestimatesofthe
seasonal contribution to the forecast:
public class SeasonalForecast extends HoltForecast {
double [] s;
double g = 0.0;
long t = 0;
public SeasonalForecast( double level, double trend,
double a, double b, double [] s, double g) {
super (level, trend, a, b);
this .s = s;
this .g = g;
this .t = 0;
}
}
The period of the seasonality in this method must be selected ahead of time,
althoughitisusuallyfairlyintuitive.Formanyoftheclassicalapplicationsof
this method, the period is either 4 or 12 depending on whether the data are
quarterly or monthly outputs. For real-time applications, the period is often
something like hourly or daily, leading to periods of 24 or 7, respectively.
For the additive seasonal forecast, the seasonal estimate for each time
period in the future is added to the output of the Holt forecast:
public class AdditiveForecast extends SeasonalForecast
{
public AdditiveForecast( double level, double trend,