如果未做特别说明,文中的程序都是 C++11 代码。
QuantLib 金融计算——高级话题之模拟跳扩散过程
跳扩散过程
1976 年,Merton 最早在衍生品定价中引入并分析了跳扩散过程,正因为如此 QuantLib 中和跳扩散相关的随机过程类称之为 Merton76Process
,一个一般的跳扩散过程可以由下面的 SDE 描述,
J(t) = \sum_{j=1}^{N(t)}(Y_j-1)
\]
其中 \(Y_j\) 是随机变量,\(N(t)\) 是计数过程。\(dJ(t)\) 表示 \(J(t)\) 在 \(t\) 时刻发生的跳跃幅度,如果发生一次跳,则幅度为 \(Y_j-1\);如果没有发生跳,则幅度为 0。
在应用于衍生品定价时,需要对上述 SDE 中的项做出一些特殊约定,常见的约定有:
- \(N(t)\) 是参数等于 \(\lambda\) 的 Poisson 过程;
- \(\log Y_j\) 服从正态分布 \(N(\mu_{jump}, \sigma_{jump}^{2})\);
- \(\mu = r - \lambda m\),其中 \(m = E[Y_j] - 1\),\(r\) 代表无风险利率
模拟算法
令 \(X(t) = \log S(t)\),那么
X(t_{i+1}) = & X(t_i) +
\left(\mu - \frac12 \sigma^2 \right)(t_{i+1} - t_i)\\
& +\sigma[W(t_{i+1}) - W(t_i)] +
\sum_{j = N(t_i)+1}^{N(t_{i+1})}\log Y_j
\end{aligned}
\]
在 \(X(t_i)\) 的基础上模拟 \(X(t_{i+1})\) 的算法如下:
- 生成 \(Z \sim N(0,1)\);
- 生成 \(N \sim \text{Poisson}(\lambda(t_{i+1}-t_i))\),若 \(N=0\),则令 \(M=0\),并转到第 4 步;
- 生成 \(\log Y_1,\dots,\log Y_N\),令 \(M = \log Y_1+\dots+\log Y_N\);
- 令 \(X(t_{i+1}) = X(t_i) + \left(\mu - \frac12 \sigma^2 \right)(t_{i+1} - t_i) +\sigma \sqrt{t_{i+1} - t_i} Z + M\)
那么
S_{t_{i+1}} =& S_{t_i} e^{(r-\lambda m -\frac{1}{2}\sigma^{2})\Delta t+ \sigma Z \sqrt{\Delta t} + M}\\
=& S_{t_i} e^{(r-\frac{1}{2}\sigma^{2})\Delta t+ \sigma Z \sqrt{\Delta t}}e^{(-\lambda m)\Delta t+M}
\end{aligned}
\]
其中,\(\Delta t = t_{i+1} - t_i\),而 \(e^{(-\lambda m)\Delta t+M}\) 是跳扩散相对于一般 Black-Scholes-Merton 过程的修正项。
面临的问题
目前 QuantLib 中提供的 Merton76Process
类只能配合“解析定价引擎”使用,本身不具备模拟随机过程路径的功能。究其原因,问题出在 QuantLib 的编码约定和 StochasticProcess1D
提供的接口两方面:
- QuantLib 中约定
StochasticProcess
派生出的子类仅能描述 SDE 的结构信息,也就是 SDE 的参数、漂移和扩散项的函数形式,子类不携带有关随机数生成的信息,所有随机数生成的相关信息均由 Monte Carlo 框架中其他组件控制; - 生成随机过程路径的核心函数是
evolve
方法,StochasticProcess1D
提供的接口是evolve(Time t0, Real x0, Time dt, Real dw)
。如果Merton76Process
按约定实现evolve
方法的话,形式必须是evolve(Time t0, Real x0, Time dt, const Array &dw)
,因为模拟跳需要额外的随机性,所以dw
必须是一个Array
。很明显,不匹配。
“脏”的方法
在不改变当前接口的前提下,若要实现模拟跳扩散过程,需要用比较“脏”一点儿的手段,即打破约定,让随机过程类携带一个随机数发生器,为模拟跳提供额外的随机性。
具体来说,需要声明一个 Merton76Process
的派生类,该类携带一个高斯随机数发生器。因为从数学上来讲跳扩散过程推广自一般 Black-Scholes-Merton 过程,添加了一个修正项,所以遵循“适配器模式”(或“装饰器模式”)的思想,绝大部分计算可以委托给一个 BlackScholesMertonProcess
对象,仅需要对 drift
和 evolve
方法作必要的修改。
“干净”的方法
当然,“干净”的方法要改变当前接口:
- 声明一个和
StochasticProcess1D
平行的新类StochasticProcess1DJump
,二者唯一的区别是evolve
方法,在StochasticProcess1DJump
中形式是evolve(Time t0, Real x0, Time dt, const Array &dw)
; - 将
Merton76Process
改成继承自StochasticProcess1DJump
。
实现
下面的代码实现了前面提到的“脏”的方法,因为随机数发生器的种类有很多,且没有基类提供统一的接口,所以使用了模板技术让类可以接受不同类型的随机数发生器。同时,许多计算被委托给了一个 BlackScholesMertonProcess
对象。
#ifndef MERTON76JUMPDIFFUSIONPROCESS_HPP
#define MERTON76JUMPDIFFUSIONPROCESS_HPP
#include <ql/math/distributions/normaldistribution.hpp>
#include <ql/math/distributions/poissondistribution.hpp>
#include <ql/math/randomnumbers/boxmullergaussianrng.hpp>
#include <ql/math/randomnumbers/mt19937uniformrng.hpp>
#include <ql/processes/blackscholesprocess.hpp>
#include <ql/processes/merton76process.hpp>
namespace QuantLib
{
template<typename GAUSS_RNG>
class Merton76JumpDiffusionProcess : public Merton76Process
{
public:
Merton76JumpDiffusionProcess(const Handle<Quote>& stateVariable,
const Handle<YieldTermStructure>& dividendTS,
const Handle<YieldTermStructure>& riskFreeTS,
const Handle<BlackVolTermStructure>& blackVolTS,
const Handle<Quote>& jumpInt,
const Handle<Quote>& logJMean,
const Handle<Quote>& logJVol,
const GAUSS_RNG& gauss_rng,
const ext::shared_ptr<discretization>& disc =
ext::shared_ptr<discretization>(new EulerDiscretization))
: Merton76Process(
stateVariable,
dividendTS,
riskFreeTS,
blackVolTS,
jumpInt,
logJMean,
logJVol,
disc)
, blackProcess_(
new BlackScholesMertonProcess(
stateVariable,
dividendTS,
riskFreeTS,
blackVolTS,
disc))
, gauss_rng_(gauss_rng)
{
}
virtual ~Merton76JumpDiffusionProcess() {}
Real x0() const
{
return blackProcess_->x0();
}
Time time(const Date& d) const
{
return blackProcess_->time(d);
}
Real diffusion(Time t,
Real x) const
{
return blackProcess_->diffusion(t, x);
}
Real apply(Real x0,
Real dx) const
{
return blackProcess_->apply(x0, dx);
}
Size factors() const
{
return 1;
}
Real drift(Time t,
Real x) const
{
Real lambda_ = Merton76Process::jumpIntensity()->value();
Real delta_ = Merton76Process::logJumpVolatility()->value();
Real nu_ = Merton76Process::logMeanJump()->value();
Real m_ = std::exp(nu_ + 0.5 * delta_ * delta_) - 1;
return blackProcess_->drift(t, x) - lambda_ * m_;
}
Real evolve(Time t0,
Real x0,
Time dt,
Real dw) const;
private:
const CumulativeNormalDistribution cumNormalDist_;
ext::shared_ptr<GeneralizedBlackScholesProcess> blackProcess_;
GAUSS_RNG gauss_rng_;
};
template<typename GAUSS_RNG>
Real Merton76JumpDiffusionProcess<GAUSS_RNG>::evolve(Time t0,
Real x0,
Time dt,
Real dw) const
{
Real lambda_ = Merton76Process::jumpIntensity()->value();
Real delta_ = Merton76Process::logJumpVolatility()->value();
Real nu_ = Merton76Process::logMeanJump()->value();
Real m_ = std::exp(nu_ + 0.5 * delta_ * delta_) - 1;
Real p = cumNormalDist_(gauss_rng_.next().value);
if (p < 0.0)
p = 0.0;
else if (p >= 1.0)
p = 1.0 - QL_EPSILON;
Real j = gauss_rng_.next().value;
const Real n = InverseCumulativePoisson(lambda_ * dt)(p);
Real retVal = blackProcess_->evolve(
t0, x0, dt, dw);
retVal *=
std::exp(-lambda_ * m_ * dt + nu_ * n + delta_ * std::sqrt(n) * j);
return retVal;
}
}
#endif // MERTON76JUMPDIFFUSIONPROCESS_HPP
示例
下面模拟两条曲线
#include <iostream>
#include <ql/math/randomnumbers/boxmullergaussianrng.hpp>
#include <ql/math/randomnumbers/mt19937uniformrng.hpp>
#include <ql/processes/blackscholesprocess.hpp>
#include <ql/quotes/simplequote.hpp>
#include <ql/termstructures/volatility/equityfx/blackconstantvol.hpp>
#include <ql/termstructures/yield/flatforward.hpp>
#include <ql/time/calendars/target.hpp>
#include <ql/time/date.hpp>
#include <ql/time/daycounters/actualactual.hpp>
#include "Merton76JumpDiffusionProcess.hpp"
int main() {
using namespace std;
using namespace QuantLib;
Date refDate = Date(27, Mar, 2019);
Rate riskFreeRate = 0.03;
Rate dividendRate = 0.01;
Real spot = 52.0;
Rate vol = 0.2;
Calendar cal = TARGET();
DayCounter dc = ActualActual();
ext::shared_ptr<YieldTermStructure> rdStruct(
new FlatForward(refDate, riskFreeRate, dc));
ext::shared_ptr<YieldTermStructure> rqStruct(
new FlatForward(refDate, dividendRate, dc));
Handle<YieldTermStructure> rdHandle(rdStruct);
Handle<YieldTermStructure> rqHandle(rqStruct);
ext::shared_ptr<SimpleQuote> spotQuote(new SimpleQuote(spot));
Handle<Quote> spotHandle(spotQuote);
ext::shared_ptr<BlackVolTermStructure> volQuote(
new BlackConstantVol(refDate, cal, vol, dc));
Handle<BlackVolTermStructure> volHandle(volQuote);
// Specify the jump intensity, jump mean and
// jump volatility objects
Real jumpIntensity = 0.2; // lambda
Real jumpVolatility = 0.3;
Real jumpMean = 0.0;
ext::shared_ptr<SimpleQuote> jumpInt(new SimpleQuote(jumpIntensity));
ext::shared_ptr<SimpleQuote> jumpVol(new SimpleQuote(jumpVolatility));
ext::shared_ptr<SimpleQuote> jumpMn(new SimpleQuote(jumpMean));
Handle<Quote> jumpI(jumpInt), jumpV(jumpVol), jumpM(jumpMn);
ext::shared_ptr<BlackScholesMertonProcess> bsmProcess(
new BlackScholesMertonProcess(
spotHandle, rqHandle, rdHandle, volHandle));
unsigned long seed = 12324u;
MersenneTwisterUniformRng unifMt(seed);
MersenneTwisterUniformRng unifMtJ(25u);
typedef BoxMullerGaussianRng<MersenneTwisterUniformRng> GAUSS;
GAUSS bmGauss(unifMt);
GAUSS jGauss(unifMtJ);
ext::shared_ptr<Merton76JumpDiffusionProcess<GAUSS>> mtProcess(
new Merton76JumpDiffusionProcess<GAUSS>(
spotHandle, rqHandle, rdHandle, volHandle,
jumpI, jumpM, jumpV, jGauss));
Time dt = 0.004, t = 0.0;
Real x = spotQuote->value();
Real y = spotQuote->value();
Real dw;
Size numVals = 250;
std::cout << "Time, Jump, NoJump" << std::endl;
std::cout << t << ", " << x << ", " << y << std::endl;
for (Size j = 1; j <= numVals; ++j) {
dw = bmGauss.next().value;
x = mtProcess->evolve(t, x, dt, dw);
y = bsmProcess->evolve(t, y, dt, dw);
std::cout << t + dt << ", " << x << ", " << y << std::endl;
t += dt;
}
return EXIT_SUCCESS;
}
参考文献
- 《金融工程中的蒙特卡罗方法》