public override double LogLikelihood(Vector <double> parameter, double penaltyFactor, bool fillOutputs)
{
//var sigmaSquared = new Vector(values.Count);
//double mVar = GetVariance(parameter);
//var allLLs = new Vector(values.Count);
//for (int t = 0; t < values.Count; ++t )
// allLLs[t] = conditionalLL(t, sigmaSquared, parameter, mVar);
Vector <double> pbak = Parameters;
if (values == null)
{
return(double.NaN);
}
if (parameter != null)
{
Parameters = parameter;
}
var sigmaSquared = Vector <double> .Build.Dense(values.Count);
double mVar = GetVariance(Parameters);
double logLikelihood = 0;
var allLLs = Vector <double> .Build.Dense(values.Count);
for (int t = 0; t < values.Count; ++t)
{
allLLs[t] = conditionalLL(t, sigmaSquared, Parameters, mVar);
logLikelihood += allLLs[t];
}
if (fillOutputs)
{
var rts = new TimeSeries {
Title = values.Title + "[GARCH Res]"
};
predictiveStdDevAtAvail = new TimeSeries {
Title = values.Title + "Pred.StDev[AA]"
};
for (int t = 0; t < values.Count; ++t)
{
double rx = values[t] / Math.Sqrt(sigmaSquared[t]);
rts.Add(values.TimeStamp(t), rx, false);
if (t > 0)
{
predictiveStdDevAtAvail.Add(values.TimeStamp(t - 1), Math.Sqrt(sigmaSquared[t]), false);
}
}
Residuals = rts;
GoodnessOfFit = logLikelihood;
}
if (parameter != null)
{
Parameters = pbak;
}
var llp = new LogLikelihoodPenalizer(allLLs);
return(llp.LogLikelihood - llp.Penalty * penaltyFactor);
}
public override double LogLikelihood(Vector <double> parameter, double penaltyFactor, bool fillOutputs)
{
Vector <double> allLLs = null;
Vector <double> pbak = Parameters; // save the current one
if (values == null)
{
return(double.NaN);
}
if (values.Count == 0)
{
return(double.NaN);
}
if (parameter != null)
{
Parameters = parameter;
}
Residuals = null;
double loglikelihood = 0;
if (!DataIsLongitudinal())
{
double[] rs;
double[] forecs;
double[] resids = ComputeSpecialResiduals(values, out rs, 1, out forecs);
TimeSeries rts;
TimeSeries unstdrts;
if (fillOutputs)
{
rts = new TimeSeries {
Title = values.Title + "[ARMA Res]"
};
unstdrts = new TimeSeries {
Title = values.Title + "[ARMA Res]"
};
oneStepPredictors = new TimeSeries {
Title = values.Title + "[Predic]"
};
oneStepPredStd = new TimeSeries {
Title = values.Title + "[Pred. Stdev.]"
};
oneStepPredictorsAtAvailability = new TimeSeries {
Title = values.Title + "[Predic. AA]"
};
for (int t = 0; t < values.Count; ++t)
{
rts.Add(values.TimeStamp(t), resids[t] / Sigma, false);
unstdrts.Add(values.TimeStamp(t), resids[t], false);
double stdev = Math.Sqrt(rs[t]);
double fx = values[t] - resids[t] * stdev;
oneStepPredictors.Add(values.TimeStamp(t), fx, false);
oneStepPredStd.Add(values.TimeStamp(t), stdev * Sigma, false);
fx = t < values.Count - 1 ? values[t + 1] - resids[t + 1] * Math.Sqrt(rs[t + 1]) : forecs[0];
oneStepPredictorsAtAvailability.Add(values.TimeStamp(t), fx, false);
}
Residuals = rts;
unstandardizedResiduals = unstdrts;
}
// now it is easy to compute the likelihood
// this is a straightforward implementation of eqn (8.7.4) in Brockwell & Davis,
// Time Series: Theory and Methods (2nd edition)
allLLs = GetLikelihoodsFromResiduals(resids, rs);
for (int t = 0; t < values.Count; ++t)
{
loglikelihood += allLLs[t];
}
}
else
{
// do nothing for now!
return(double.NaN);
}
if (fillOutputs)
{
GoodnessOfFit = loglikelihood;
}
if (parameter != null)
{
Parameters = pbak; // then restore original
}
var llp = new LogLikelihoodPenalizer(allLLs);
return(llp.LogLikelihood - llp.Penalty * penaltyFactor);
}
