public SeasonalTrendLoess buildSmoother(double[] data)
{
this.sanityCheck(data);
if (this.fPeriodic)
{
this.fSeasonalWidth = (100 * data.Length);
this.fSeasonalDegree = 0;
}
else if (this.fSeasonalDegree == 0)
{
this.fSeasonalDegree = 1;
}
var seasonalSettings = buildSettings(this.fSeasonalWidth, this.fSeasonalDegree, this.fSeasonalJump);
if (this.fFlatTrend)
{
this.fTrendWidth = 100 * this.fPeriodLength * data.Length;
this.fTrendDegree = 0;
}
else if (this.fLinearTrend)
{
this.fTrendWidth = 100 * this.fPeriodLength * data.Length;
this.fTrendDegree = 1;
}
else if (this.fTrendDegree == 0)
{
this.fTrendDegree = 1;
}
if ((this.fTrendWidth == 0))
{
this.fTrendWidth = SeasonalTrendLoessBuilder.calcDefaultTrendWidth(this.fPeriodLength, this.fSeasonalWidth);
}
LoessSettings trendSettings = this.buildSettings(this.fTrendWidth, this.fTrendDegree, this.fTrendJump);
if ((this.fLowpassWidth == 0))
{
this.fLowpassWidth = this.fPeriodLength;
}
LoessSettings lowpassSettings = this.buildSettings(this.fLowpassWidth, this.fLowpassDegree, this.fLowpassJump);
var stl = new SeasonalTrendLoess(data, this.fPeriodLength, this.fInnerIterations, this.fRobustIterations,
seasonalSettings, trendSettings, lowpassSettings);
return(stl);
}
public static Decomposition performRobustPeriodicDecomposition(double[] data, int periodicity)
{
// The LOESS interpolator with degree 0 and a very long window (arbitrarily chosen to be 100 times the length of
// the array) will interpolate all points as the average value of the series. This particular setting is used
// for smoothing the seasonal sub-cycles, so the end result is that the seasonal component of the decomposition
// is exactly periodic.
// This fit is for diagnostic purposes, so we just do a single inner and outer iteration.
var stlBuilder = new SeasonalTrendLoessBuilder();
stlBuilder.PeriodLength = periodicity;
stlBuilder.SeasonalWidth = 100 * data.Length;
stlBuilder.SeasonalDegree = 0;
stlBuilder.InnerIterations = 1;
stlBuilder.RobustIterations = 1;
var stl = stlBuilder.buildSmoother(data);
return(stl.decompose());
}
/// <summary>
/// Perform STL time series decomposition
/// </summary>
/// <param name="values"></param>
/// <param name="sWindow"></param>
/// <param name="sDegree"></param>
/// <param name="tWindow"></param>
/// <param name="tDegree"></param>
/// <param name="lWindow"></param>
/// <param name="lDegree"></param>
/// <param name="sJump"></param>
/// <param name="tJump"></param>
/// <param name="lJump"></param>
/// <param name="isRobust"></param>
/// <param name="inner"></param>
/// <param name="outer"></param>
/// <returns>Tuple of the three components</returns>
public static (double[] Trend, double[] Seasonal, double[] Residual) Fit(double[] values,
int sWindow, int sDegree = 0, int tWindow = 0, int tDegree = 1, int lWindow = 0, int lDegree = 0,
int sJump = 0, int tJump = 0, int lJump = 0, bool isRobust = false, int inner = 0, int outer = 0)
{
var builder = new SeasonalTrendLoessBuilder();
builder.PeriodLength = sWindow; // Data has a period of 12
builder.Periodic = sWindow > 0;
// setSeasonalWidth(12). // Monthly data smoothed over 35 years
builder.Robust = isRobust; // Not expecting outliers, so no robustness iterations
var smoother = builder.buildSmoother(values);
var stl = smoother.decompose();
return(stl.Trend, stl.Seasonal, stl.Residual);
}