private double cornishFisherEps(HestonProcess process, double nu_0, double nu_t, double dt, double eps)
{
// use moment generating function to get the
// first,second, third and fourth moment of the distribution
double d = 1e-2;
double p2 = Phi(process, new Complex(0, -2 * d), nu_0, nu_t, dt).Real;
double p1 = Phi(process, new Complex(0, -d), nu_0, nu_t, dt).Real;
double p0 = Phi(process, new Complex(0, 0), nu_0, nu_t, dt).Real;
double pm1 = Phi(process, new Complex(0, d), nu_0, nu_t, dt).Real;
double pm2 = Phi(process, new Complex(0, 2 * d), nu_0, nu_t, dt).Real;
double avg = (pm2 - 8 * pm1 + 8 * p1 - p2) / (12 * d);
double m2 = (-pm2 + 16 * pm1 - 30 * p0 + 16 * p1 - p2) / (12 * d * d);
double var = m2 - avg * avg;
double stdDev = Math.Sqrt(var);
double m3 = (-0.5 * pm2 + pm1 - p1 + 0.5 * p2) / (d * d * d);
double skew = (m3 - 3 * var * avg - avg * avg * avg) / (var * stdDev);
double m4 = (pm2 - 4 * pm1 + 6 * p0 - 4 * p1 + p2) / (d * d * d * d);
double kurt = (m4 - 4 * m3 * avg + 6 * m2 * avg * avg - 3 * avg * avg * avg * avg) / (var * var);
// Cornish-Fisher relation to come up with an improved
// estimate of 1-F(u_\eps) < \eps
double q = new InverseCumulativeNormal().value(1 - eps);
double w = q + (q * q - 1) / 6 * skew + (q * q * q - 3 * q) / 24 * (kurt - 3)
- (2 * q * q * q - 5 * q) / 36 * skew * skew;
return(avg + w * stdDev);
}
/*! gaussian-assumption y-th percentile
*/
/*! \pre percentile must be in range (0%-100%) extremes excluded */
public double gaussianPercentile(double percentile)
{
Utils.QL_REQUIRE(percentile > 0.0 && percentile < 1.0, () => "percentile (" + percentile + ") must be in (0.0, 1.0)");
InverseCumulativeNormal gInverse = new InverseCumulativeNormal(mean(), standardDeviation());
return(gInverse.value(percentile));
}
/*! gaussian-assumption y-th percentile, defined as the value x
* such that \f[ y = \frac{1}{\sqrt{2 \pi}}
* \int_{-\infty}^{x} \exp (-u^2/2) du \f]
*/
/*! \pre percentile must be in range (0%-100%) extremes excluded */
public double gaussianPercentile(double percentile)
{
if (!(percentile > 0.0 && percentile < 1.0))
{
throw new ApplicationException("percentile (" + percentile + ") must be in (0.0, 1.0)");
}
InverseCumulativeNormal gInverse = new InverseCumulativeNormal(mean(), standardDeviation());
return(gInverse.value(percentile));
}
public FdmBlackScholesMesher(int size,
GeneralizedBlackScholesProcess process,
double maturity, double strike,
double?xMinConstraint = null,
double?xMaxConstraint = null,
double eps = 0.0001,
double scaleFactor = 1.5,
Pair <double?, double?> cPoint
= null)
: base(size)
{
double S = process.x0();
Utils.QL_REQUIRE(S > 0.0, () => "negative or null underlying given");
// Set the grid boundaries
double normInvEps = new InverseCumulativeNormal().value(1 - eps);
double sigmaSqrtT
= process.blackVolatility().currentLink().blackVol(maturity, strike)
* Math.Sqrt(maturity);
double?xMin = Math.Log(S) - sigmaSqrtT * normInvEps * scaleFactor;
double?xMax = Math.Log(S) + sigmaSqrtT * normInvEps * scaleFactor;
if (xMinConstraint != null)
{
xMin = xMinConstraint;
}
if (xMaxConstraint != null)
{
xMax = xMaxConstraint;
}
Fdm1dMesher helper;
if (cPoint != null &&
Math.Log(cPoint.first.Value) >= xMin && Math.Log(cPoint.first.Value) <= xMax)
{
helper = new Concentrating1dMesher(xMin.Value, xMax.Value, size,
new Pair <double?, double?>(Math.Log(cPoint.first.Value), cPoint.second));
}
else
{
helper = new Uniform1dMesher(xMin.Value, xMax.Value, size);
}
locations_ = helper.locations();
for (int i = 0; i < locations_.Count; ++i)
{
dplus_[i] = helper.dplus(i);
dminus_[i] = helper.dminus(i);
}
}
//! gaussian-assumption Expected Shortfall at a given percentile
/*! Assuming a gaussian distribution it
* returns the expected loss in case that the loss exceeded
* a VaR threshold,
*
* that is the average of observations below the
* given percentile \f$ p \f$.
* Also know as conditional value-at-risk.
*
* See Artzner, Delbaen, Eber and Heath,
* "Coherent measures of risk", Mathematical Finance 9 (1999)
*/
public double gaussianExpectedShortfall(double percentile)
{
Utils.QL_REQUIRE(percentile < 1.0 && percentile >= 0.9, () => "percentile (" + percentile + ") out of range [0.9, 1)");
double m = this.mean();
double std = this.standardDeviation();
InverseCumulativeNormal gInverse = new InverseCumulativeNormal(m, std);
double var = gInverse.value(1.0 - percentile);
NormalDistribution g = new NormalDistribution(m, std);
double result = m - std * std * g.value(var) / (1.0 - percentile);
// expectedShortfall must be a loss
// this means that it has to be MIN(result, 0.0)
// expectedShortfall must also be a positive quantity, so -MIN(*)
return(-Math.Min(result, 0.0));
}
// alternative delta type
private double strikeFromDelta(double delta, DeltaVolQuote.DeltaType dt)
{
double res = 0.0;
double arg = 0.0;
InverseCumulativeNormal f = new InverseCumulativeNormal();
Utils.QL_REQUIRE(delta * phi_ >= 0.0, () => "Option type and delta are incoherent.");
switch (dt)
{
case DeltaVolQuote.DeltaType.Spot:
Utils.QL_REQUIRE(Math.Abs(delta) <= fDiscount_, () => "Spot delta out of range.");
arg = -phi_ *f.value(phi_ *delta / fDiscount_) * stdDev_ + 0.5 * stdDev_ * stdDev_;
res = forward_ * Math.Exp(arg);
break;
case DeltaVolQuote.DeltaType.Fwd:
Utils.QL_REQUIRE(Math.Abs(delta) <= 1.0, () => "Forward delta out of range.");
arg = -phi_ *f.value(phi_ *delta) * stdDev_ + 0.5 * stdDev_ * stdDev_;
res = forward_ * Math.Exp(arg);
break;
case DeltaVolQuote.DeltaType.PaSpot:
case DeltaVolQuote.DeltaType.PaFwd:
// This has to be solved numerically. One of the
// problems is that the premium adjusted call delta is
// not monotonic in strike, such that two solutions
// might occur. The one right to the max of the delta is
// considered to be the correct strike. Some proper
// interval bounds for the strike need to be chosen, the
// numerics can otherwise be very unreliable and
// unstable. I've chosen Brent over Newton, since the
// interval can be specified explicitly and we can not
// run into the area on the left of the maximum. The
// put delta doesn't have this property and can be
// solved without any problems, but also numerically.
BlackDeltaPremiumAdjustedSolverClass f1 = new BlackDeltaPremiumAdjustedSolverClass(
ot_, dt, spot_, dDiscount_, fDiscount_, stdDev_, delta);
Brent solver = new Brent();
solver.setMaxEvaluations(1000);
double accuracy = 1.0e-10;
double rightLimit = 0.0;
double leftLimit = 0.0;
// Strike of not premium adjusted is always to the right of premium adjusted
if (dt == DeltaVolQuote.DeltaType.PaSpot)
{
rightLimit = strikeFromDelta(delta, DeltaVolQuote.DeltaType.Spot);
}
else
{
rightLimit = strikeFromDelta(delta, DeltaVolQuote.DeltaType.Fwd);
}
if (phi_ < 0)
{
// if put
res = solver.solve(f1, accuracy, rightLimit, 0.0, spot_ * 100.0);
break;
}
else
{
// find out the left limit which is the strike
// corresponding to the value where premium adjusted
// deltas have their maximum.
BlackDeltaPremiumAdjustedMaxStrikeClass g = new BlackDeltaPremiumAdjustedMaxStrikeClass(
ot_, dt, spot_, dDiscount_, fDiscount_, stdDev_);
leftLimit = solver.solve(g, accuracy, rightLimit * 0.5, 0.0, rightLimit);
double guess = leftLimit + (rightLimit - leftLimit) * 0.5;
res = solver.solve(f1, accuracy, guess, leftLimit, rightLimit);
} // end if phi<0 else
break;
default:
Utils.QL_FAIL("invalid delta type");
break;
}
return(res);
}
public FdmBlackScholesMesher(int size,
GeneralizedBlackScholesProcess process,
double maturity, double strike,
double?xMinConstraint = null,
double?xMaxConstraint = null,
double eps = 0.0001,
double scaleFactor = 1.5,
Pair <double?, double?> cPoint
= null,
DividendSchedule dividendSchedule = null,
FdmQuantoHelper fdmQuantoHelper = null,
double spotAdjustment = 0.0)
: base(size)
{
double S = process.x0();
Utils.QL_REQUIRE(S > 0.0, () => "negative or null underlying given");
dividendSchedule = dividendSchedule == null ? new DividendSchedule() : dividendSchedule;
List <pair_double> intermediateSteps = new List <pair_double>();
for (int i = 0; i < dividendSchedule.Count &&
process.time(dividendSchedule[i].date()) <= maturity; ++i)
{
intermediateSteps.Add(
new pair_double(
process.time(dividendSchedule[i].date()),
dividendSchedule[i].amount()
));
}
int intermediateTimeSteps = (int)Math.Max(2, 24.0 * maturity);
for (int i = 0; i < intermediateTimeSteps; ++i)
{
intermediateSteps.Add(
new pair_double((i + 1) * (maturity / intermediateTimeSteps), 0.0));
}
intermediateSteps.Sort();
Handle <YieldTermStructure> rTS = process.riskFreeRate();
Handle <YieldTermStructure> qTS = fdmQuantoHelper != null
? new Handle <YieldTermStructure>(
new QuantoTermStructure(process.dividendYield(),
process.riskFreeRate(),
new Handle <YieldTermStructure>(fdmQuantoHelper.foreignTermStructure()),
process.blackVolatility(),
strike,
new Handle <BlackVolTermStructure>(fdmQuantoHelper.fxVolatilityTermStructure()),
fdmQuantoHelper.exchRateATMlevel(),
fdmQuantoHelper.equityFxCorrelation()))
: process.dividendYield();
double lastDivTime = 0.0;
double fwd = S + spotAdjustment;
double mi = fwd, ma = fwd;
for (int i = 0; i < intermediateSteps.Count; ++i)
{
double divTime = intermediateSteps[i].first;
double divAmount = intermediateSteps[i].second;
fwd = fwd / rTS.currentLink().discount(divTime) * rTS.currentLink().discount(lastDivTime)
* qTS.currentLink().discount(divTime) / qTS.currentLink().discount(lastDivTime);
mi = Math.Min(mi, fwd); ma = Math.Max(ma, fwd);
fwd -= divAmount;
mi = Math.Min(mi, fwd); ma = Math.Max(ma, fwd);
lastDivTime = divTime;
}
// Set the grid boundaries
double normInvEps = new InverseCumulativeNormal().value(1 - eps);
double sigmaSqrtT
= process.blackVolatility().currentLink().blackVol(maturity, strike)
* Math.Sqrt(maturity);
double?xMin = Math.Log(mi) - sigmaSqrtT * normInvEps * scaleFactor;
double?xMax = Math.Log(ma) + sigmaSqrtT * normInvEps * scaleFactor;
if (xMinConstraint != null)
{
xMin = xMinConstraint;
}
if (xMaxConstraint != null)
{
xMax = xMaxConstraint;
}
Fdm1dMesher helper;
if (cPoint != null &&
cPoint.first != null &&
Math.Log(cPoint.first.Value) >= xMin && Math.Log(cPoint.first.Value) <= xMax)
{
helper = new Concentrating1dMesher(xMin.Value, xMax.Value, size,
new Pair <double?, double?>(Math.Log(cPoint.first.Value), cPoint.second));
}
else
{
helper = new Uniform1dMesher(xMin.Value, xMax.Value, size);
}
locations_ = helper.locations();
for (int i = 0; i < locations_.Count; ++i)
{
dplus_[i] = helper.dplus(i);
dminus_[i] = helper.dminus(i);
}
}
