OU Process

The general definition of an Ornstein-Uhlebeck (OU) process is as the solution of an SDE of the form.

(49)\[\begin{equation} d x_t = -\kappa x_t dt + d z_t \end{equation}\]

where \(z\), with \(z_0 = 0\), is a Lévy process. As \(z_t\) drives the OU process, it is usually referred to as a background driving Lévy process (BDLP).

The OU process can be integrated into the formula (see Appendix below).

(50)\[\begin{equation} x_t = x_0 e^{-\kappa t} + \int_0^t e^{-\kappa\left(t-s\right)} d z_{\kappa s} \end{equation}\]

Gaussian OU Process

The Gaussian Ornstein-Uhlebeck process is an OU process where the BDLP is a Brownian motion with drift \(d z_{\kappa t} = \kappa\theta dt + \sigma dw_{\kappa t}\). Substituting this into the OU SDE equation yields:

(51)\[\begin{align} dx_t &= \kappa\left(\theta - x_t\right) dt + \sigma dw_{\kappa t} \\ x_t &= x_0 e^{-\kappa t} + \theta\left(1 - e^{-\kappa t}\right) + \sigma \int_0^t e^{-\kappa\left(t-s\right)} d w_{\kappa s} \end{align}\]

\(\theta\) is the long-term value of \({\bf x}_t\), \(\sigma\) is a volatility parameter and \(w_t\) is the standard Brownian motion.

In the interest rate literature, this model is also known as Vasicek model.

Marginal and moments

The model has a closed-form solution for marginal distribution, which equal to the normal standard distribution with the mean and the variance defined by

(52)\[\begin{align} {\mathbb E}[x_t] &= x_0 e^{-\kappa t} + \theta\left(1 - e^{-\kappa t}\right) \\ {\mathbb Var}[x_t] &= \frac{\sigma^2}{2 \kappa}\left(1 - e^{-2 \kappa t}\right) \end{align}\]

which means the process admits a stationary probability distribution equal to

(53)\[\begin{equation} x_t \sim N\left(\theta, \frac{\sigma^2}{2\kappa}\right)\ \ t\rightarrow\infty \end{equation}\]

from quantflow.sp.ou import Vasicek
pr = Vasicek()
pr

Vasicek(rate=1.0, kappa=1.0, bdlp=WeinerProcess(sigma=1), theta=1.0)

pr.sample(20, time_horizon=1, time_steps=1000).plot().update_traces(
    line_width=0.5
).update_layout(
    title="Mean reverting paths of Vasicek model"
)

m = pr.marginal(1)
m.mean(), m.std()

(1.0, 0.6575198539828996)

m.mean_from_characteristic(), m.std_from_characteristic()

(0.9999996171672217, 0.6575201729021828)

Non-gaussian OU process

Non-Gaussian OU processes offer the possibility of capturing significant distributional deviations from Gaussianity and for flexible modeling of dependence structure.

Following the seminal paper of [BN01], we look at a model based on this SDEs

(54)\[\begin{equation} dx_t = -\kappa x_t dt + dz_{\kappa t} \end{equation}\]

The unusual timing \(dz_{\kappa t}\) is deliberately chosen so that it will turn out that whatever the value of of \(\kappa\), the marginal distribution of of \(x_t\) will be unchanged. Hence we separately parameterize the distribution of the volatility and the dynamic structure.

The \(z_t\) has positive increments and no drift. This type of process is called a subordinator [Ber96].

Integration

When the subordinator is a Compound Poisson process, then the integration takes the form

(55)\[\begin{equation} x_t = x_0 e^{-\kappa t} + \sum_{n=0}^{N_{\kappa t}} e^{-\kappa t-m_n} j_n \end{equation}\]

where \(m_n\) are the jump times of the Poisson process \(N_{\kappa t} and \)j_n$ are the jump sizes drawn from the jump distribution.

Integrated Intensity

One of the advantages of these OU processes is that they offer a great deal of analytical tractability. For example, the integrated value of the process, which can be used as a time change for Lévy processes, is given by

(56)\[\begin{align} \int_0^t x_s ds &= \epsilon_t x_0 + \int_0^t \epsilon_{t-s} d z_{\kappa s} = \frac{z_{\kappa t} - x_t + x_0}{\kappa}\\ \epsilon_t &= \frac{1 - e^{-\kappa t}}{\kappa} \end{align}\]

Lévy density

It is possible to show, see [BN01], that given the Lévy density \(w\) of \(z\), in other words, the density of the Lévy measure of the Lévy-Khintchine representation of the BDLP \(z_1\), than it is possible to obtain the density \(u\) of \(x\) via

(57)\[\begin{equation} w_y = -u_y - y \frac{d u_y}{d y} \end{equation}\]

Gamma OU Process

The library provides an implementation of the non-gaussian OU process in the form of a Gamma OU process, where the invariant distribution of \(x_t\) is a gamma distribution \(\Gamma\left(\lambda, \beta\right)\).

In this case, the BDLP is an exponential compound Poisson process with Lévy density \(\lambda\beta e^{-\beta x}\), in other words, the exponential compound Poisson process with intensity \(\lambda\) and decay \(\beta\).

from quantflow.sp.ou import GammaOU

pr = GammaOU.create(decay=10, kappa=5)
pr

GammaOU(rate=1.0, kappa=5.0, bdlp=CompoundPoissonProcess[Exponential](intensity=10.0, jumps=Exponential(decay=10.0)))

Characteristic Function

The charatecristic exponent of the \(\Gamma\)-OU process is given by, see [Sab21])

(58)\[\begin{equation} \phi_{u, t} = -x_{0} i u e^{-\kappa t} - \lambda\ln\left(\frac{\beta-iue^{-\kappa t}}{\beta -iu}\right) \end{equation}\]

pr.marginal(1).mean(), pr.marginal(1).std()

(1.0, 0.31622776601683794)

import numpy as np
from quantflow.utils import plot

m = pr.marginal(5)
plot.plot_marginal_pdf(m, 128)

from quantflow.utils.plot import plot_characteristic
plot_characteristic(m)

Sampling Gamma OU

from quantflow.sp.ou import GammaOU
pr = GammaOU.create(decay=10, kappa=5)

pr.sample(50, time_horizon=1, time_steps=1000).plot().update_traces(line_width=0.5)

MC testing

Test the simulated meand and stadard deviation against the values from the invariant gamma distribution.

import pandas as pd
from quantflow.utils import plot

paths = pr.sample(1000, time_horizon=1, time_steps=1000)
mean = dict(mean=pr.marginal(paths.time).mean(), simulated=paths.mean())
df = pd.DataFrame(mean, index=paths.time)
plot.plot_lines(df)

std = dict(std=pr.marginal(paths.time).std(), simulated=paths.std())
df = pd.DataFrame(std, index=paths.time)
plot.plot_lines(df)

Appendix

The integration of the OU process can be achieved by multiplying both sides of the equation by \(e^{\kappa t}\) and performing simple steps as indicated below

(59)\[\begin{align} e^{\kappa t} d x_t &= -e^{\kappa t} \kappa x_t dt + e^{\kappa t} d z_t \\ d\left(e^{\kappa t} x\right) - \kappa e^{\kappa t} x_t dt &= -e^{\kappa t} \kappa x_t dt + e^{\kappa t} d z_t \\ d\left(e^{\kappa t} x\right) &= e^{\kappa t} d z_t \\ e^{\kappa t} x_t - x_0 &= \int_0^t e^{\kappa s} d z_s \\ x_t &= x_0 e^{-\kappa t} + \int_0^t e^{-\kappa\left(t - s\right)} d z_s \end{align}\]