@Override
public void save(final int t, final AugmentedPredictionErrors pe) {
DataBlock U = pe.getTransformedPredictionErrors();
Matrix L = pe.getCholeskyFactor();
DataBlock D = L.diagonal();
Matrix E = pe.E();
int nvars = E.getColumnsCount();
n += nvars;
LogSign sld = D.sumLog();
det += sld.value;
Q.subMatrix(0, nd, nd + 1, nd + 1 + nvars).copy(E.subMatrix());
Q.row(nd).range(nd + 1, nd + 1 + nvars).copy(U);
ec.tstoolkit.maths.matrices.ElementaryTransformations.fastGivensTriangularize(Q.subMatrix());
}
@Override
public boolean collapse(AugmentedState state) {
if (state.getInfo() != StateInfo.Forecast) {
return false;
}
if (!isPositive(Q.diagonal().drop(0, 1))) {
return false;
}
// update the state vector
Matrix A = new Matrix(state.B());
int d = A.getColumnsCount();
Matrix S = new Matrix(a());
LowerTriangularMatrix.rsolve(S, A.subMatrix().transpose());
DataBlock D = b().deepClone();
LowerTriangularMatrix.lsolve(S, D);
for (int i = 0; i < d; ++i) {
DataBlock col = A.column(i);
state.a().addAY(-Q.get(d, i), col);
state.P().addXaXt(1, col);
}
state.dropAllConstraints();
return true;
}
private SubMatrix a() {
return Q.subMatrix(0, nd, 0, nd);
}
// time-varying trading days
// @Test
public void demoTD() {
TsData s = Data.X;
CompositeResults rslts = TramoSeatsProcessingFactory.process(s, TramoSeatsSpecification.RSA5);
PreprocessingModel regarima = rslts.get("preprocessing", PreprocessingModel.class);
SeatsResults seats = rslts.get("decomposition", SeatsResults.class);
assertTrue(seats != null && regarima != null);
if (regarima.isMultiplicative()) {
s = s.log();
}
int[] calPos =
regarima.description.getRegressionVariablePositions(ComponentType.CalendarEffect);
UcarimaModel ucm = seats.getUcarimaModel();
// compute the full decomposition...
SsfUcarima stoch = new SsfUcarima(ucm);
ExtendedSsfData xdata = new ExtendedSsfData(new SsfData(s, null));
xdata.setForecastsCount(s.getFrequency().intValue());
Matrix x =
regarima
.description
.buildRegressionVariables()
.all()
.matrix(new TsDomain(s.getStart(), xdata.getCount()));
RegSsf xssf = new RegSsf(stoch, x.subMatrix());
Filter filter = new Filter();
filter.setInitializer(new DiffuseSquareRootInitializer());
filter.setSsf(xssf);
DiffuseFilteringResults fr = new DiffuseFilteringResults(true);
fr.getVarianceFilter().setSavingP(true);
fr.getFilteredData().setSavingA(true);
filter.process(xdata, fr);
Smoother smoother = new Smoother();
smoother.setSsf(xssf);
smoother.setCalcVar(true);
SmoothingResults sm = new SmoothingResults();
smoother.process(xdata, fr, sm);
Smoother lsmoother = new Smoother();
lsmoother.setSsf(stoch);
lsmoother.setCalcVar(true);
SmoothingResults lsm = new SmoothingResults();
ExtendedSsfData xldata =
new ExtendedSsfData(new SsfData(regarima.linearizedSeries(false), null));
xldata.setForecastsCount(s.getFrequency().intValue());
lsmoother.process(xldata, lsm);
int spos = stoch.cmpPos(1);
DataBlock Z = new DataBlock(xssf.getStateDim());
double[] v = new double[xdata.getCount()];
double[] c = new double[xdata.getCount()];
double[] svar = sm.componentVar(spos);
double[] slvar = lsm.componentVar(spos);
int start = regarima.description.getRegressionVariablesStartingPosition();
for (int i = 0; i < v.length; ++i) {
Z.set(spos, 1);
for (int j = 0; j < calPos.length; ++j) {
Z.set(stoch.getStateDim() + calPos[j], x.get(i, calPos[j]));
}
v[i] = sm.zvariance(i, Z);
Z.set(spos, 0);
c[i] = sm.zvariance(i, Z);
System.out.print(svar[i]);
System.out.print('\t');
System.out.print(slvar[i]);
System.out.print('\t');
System.out.print(c[i]);
System.out.print('\t');
System.out.println(v[i]);
}
System.out.println(sm.P(50));
System.out.println(sm.P(svar.length - 1));
System.out.println(regarima.estimation.getLikelihood().getBVar());
}