[Quantlib-dev] Marketmodel AbcdVol

[Peter C. <pca...@vo...>]

Hi Ferdinando, Mark, Klaus,

I am working in a project where we want to make use of the ql (1.1) 
implementation of the stochastic volatility libor forward model. The 
product we want to price is an exotic swap with payoff depending on 
Euribor6m and EUR CMS20y rates. The natural choice for the "rateTimes" 
therefore seems to be a 6m grid. The rates however are observed on a 
quarterly grid. My understanding is that I can choose the 
"evolutionTimes" possibly different from the rateTimes representing the 
needed fixing times. This is what I get from the description in 
evolutiondescription.hpp:

-# evolutionTimes = the times defining the rates that are to be evolved,
-# rateTimes = the times at which the rates need to be known,
-# relevanceRates = which rates need to be known at each time.
...

I believe the explanations for evolutionTimes and rateTimes are switched 
here.

The model I am using is AbcdVol(...). The observation is that the 
swaptions from the calibration basket are greatly overpriced in the 
simulation. I think this is due to the particular choice of the 
evolutionTimes grid being strictly finer than the rateTimes grid (which 
is equal to the correlation times grid coming from a 
TimeHomogeneousForwardCorrelation). This case is in my opinion not 
handled correctly when computing the covariance matrices for the 
evolution steps in the constructor of AbcdVol. To put it more precisely 
the underlying time intervals [effStartTime, effStopTime] are 
overlapping due to wrong left interval points then, thus leading to too 
high covariances (consistent with the observation of too high simulated 
swaption prices).

Here is my proposal to fix that in abcdvol.cpp

     AbcdVol::AbcdVol(
              Real a,
              Real b,
@@ -67,24 +69,25 @@ namespace QuantLib {
                     "number of factors (" << numberOfFactors <<
                     ") must be greater than zero");

          AbcdFunction abcd(a, b, c, d);
          Real covar;
-        Time effStartTime, effStopTime;
+        Time effStartTime, effStopTime=0;
          Real correlation;
          const std::vector<Time>& corrTimes = corr->times();
          const std::vector<Time>& evolTimes = evolution.evolutionTimes();
          for (Size k=0, kk=0; k<numberOfSteps_; ++k) {

              // one covariance per evolution step
              Matrix covariance(numberOfRates_, numberOfRates_, 0.0);

              // there might be more than one correlation matrix
              // in a single evolution step
              Matrix correlations;

              for (; corrTimes[kk]<evolTimes[k]; ++kk) {
-                effStartTime = kk==0 ? 0.0 : corrTimes[kk-1];
+                effStartTime = effStopTime;
                  effStopTime = corrTimes[kk];
                  correlations = corr->correlation(kk);
                  for (Size i=0; i<numberOfRates_; ++i) {
                      for (Size j=i; j<numberOfRates_; ++j) {
                          covar = ks[i] * ks[j] * 
abcd.covariance(effStartTime,
@@ -95,11 +98,11 @@ namespace QuantLib {
                          covariance[i][j] += covar * correlation;
                      }
                  }
              }
              // last part in the evolution step
-            effStartTime = kk==0 ? 0.0 : corrTimes[kk-1];
+            effStartTime = effStopTime;
              effStopTime = evolTimes[k];
              correlations = corr->correlation(kk);
              for (Size i=0; i<numberOfRates_; ++i) {
                  for (Size j=i; j<numberOfRates_; ++j) {
                      covar = ks[i] * ks[j] * abcd.covariance(effStartTime,


Do you think this is correct and complete? The testsuite shows no 
differences at least and results are much better after these changes.

Thanks a lot
Peter