Poisson Processes

In this section, we look at the family of pure jump processes which are Lévy provcesses. The most common process is the Poisson process.

Poisson Process

The Poisson Process \(N_t\) with intensity parameter \(\lambda > 0\) is a Lévy process with values in \(N\) such that each \(N_t\) has a Poisson distribution with parameter \(\lambda t\), that is

(34)\[\begin{equation} P\left(N_t=n\right) = \frac{\left(\lambda t\right)^n}{n!}e^{-\lambda t} \end{equation}\]

The characteristic exponent is given by

(35)\[\begin{equation} \phi_{N_t, u} = t \lambda \left(1 - e^{iu}\right) \end{equation}\]

from quantflow.sp.poisson import PoissonProcess
pr = PoissonProcess(intensity=1)
pr

PoissonProcess(intensity=1.0)

m = pr.marginal(0.1)
import numpy as np
cdf = m.cdf(np.arange(5))
cdf

array([0.90483742, 0.99532116, 0.99984535, 0.99999615, 0.99999992])

n = 128*8
m.cdf_from_characteristic(5, frequency_n=n).y

array([0.90403788, 0.99499564, 0.99839477, 0.99901958, 0.99901958])

cdf1 = m.cdf_from_characteristic(5, frequency_n=n).y
cdf2 = m.cdf_from_characteristic(5, frequency_n=n, simpson_rule=False).y
10000*np.max(np.abs(cdf-cdf1)), 10000*np.max(np.abs(cdf-cdf2))

(14.505795170645097, 17.76098399924986)

Marginal

import numpy as np
from quantflow.utils import plot

m = pr.marginal(1)
plot.plot_marginal_pdf(m, frequency_n=128*8)

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

Sampling Poisson

p = pr.sample(10, time_horizon=10, time_steps=1000)
p.plot().update_traces(line_width=1)

Compound Poisson Process

The compound poisson process is a jump process, where the arrival of jumps \(N_t\) follows the same dynamic as the Poisson process but the size of jumps is no longer constant and equal to 1, instead, they are i.i.d. random variables independent from \(N_t\).

(36)\[\begin{align} x_t = \sum_{k=1}^{N_t} j_t \end{align}\]

The characteristic exponent of a compound Poisson process is given by

(37)\[\begin{align} \phi_{x_t,u} = t\int_0^\infty \left(e^{iuy} - 1\right) f(y) dy = t\lambda \left(1 - \Phi_{j,u}\right) \end{align}\]

where \(\Phi_{j,u}\) is the characteristic function of the jump distribution.

As long as we have a closed-form solution for the characteristic function of the jump distribution, then we have a closed-form solution for the characteristic exponent of the compound Poisson process.

The mean and variance of the compund Poisson is given by

(38)\[\begin{align} {\mathbb E}\left[x_t\right] &= \lambda t {\mathbb E}\left[j\right]\\ {\mathbb Var}\left[x_t^2\right] &= \lambda t \left({\mathbb Var}\left[j\right] + {\mathbb E}\left[j\right]^2\right) \end{align}\]

Exponential Compound Poisson Process

The library includes the Exponential Poisson Process, a compound Poisson process where the jump sizes are sampled from an exponential distribution.

from quantflow.sp.poisson import CompoundPoissonProcess
from quantflow.utils.distributions import Exponential

pr = CompoundPoissonProcess(intensity=1, jumps=Exponential(decay=1))
pr

CompoundPoissonProcess(intensity=1.0, jumps=Exponential(decay=1.0))

from quantflow.utils.plot import plot_characteristic
m = pr.marginal(1)
plot_characteristic(m)

m.mean(), m.mean_from_characteristic()

(1.0, 0.9999978333375087)

m.variance(), m.variance_from_characteristic()

(2.0, 1.999998249717447)

pr.sample(10, time_horizon=10, time_steps=1000).plot().update_traces(line_width=1)

MC simulations

Here we test the simulated mean and standard deviation against the analytical values.

import pandas as pd
from quantflow.utils import plot

paths = pr.sample(100, time_horizon=10, 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)

Normal Compound Poisson

A compound Poisson process with a normal jump distribution

from quantflow.utils.distributions import Normal
from quantflow.sp.poisson import CompoundPoissonProcess
pr = CompoundPoissonProcess(intensity=10, jumps=Normal(mu=0.01, sigma=0.1))
pr

CompoundPoissonProcess(intensity=10.0, jumps=Normal(mu=0.01, sigma=0.1))

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

(0.1, 0.3178049716414141)

m.mean_from_characteristic(), m.std_from_characteristic()

(0.09999999428166687, 0.3178049681523679)

Doubly Stochastic Poisson Process

The aim is to identify a stochastic process for simulating arrivals which fulfills the following properties

• Capture overdispersion

• Analytically tractable

• Capture the inherent randomness of the Poisson intensity

• Intuitive

The DSP process presented in [Unk17] has an intensity process which belongs to a class of affine diffusion and it can treated analytically.

Additional links

• Doubly Stochastic Poisson Processes with Affine Intensities

• Closed-form formulas for the distribution of the jumps of doubly-stochastic Poisson processes

• On the characteristic functional of a doubly stochastic Poisson process

• Time Change

DSP process

The DSP is defined as a time-changed Poisson process

(39)\[\begin{equation} D_t = N_{\tau_t} \end{equation}\]

where \(\tau_t\) is the cumulative intensity, or the hazard process, for the intensity process \(\lambda_t\). The Characteristic function of \(D_t\) can therefore be written as

(40)\[\begin{equation} \Phi_{D_t, u} = {\mathbb E}\left[e^{-\tau_t \left(e^{iu}-1\right)}\right] \end{equation}\]

The doubly stochastic Poisson process (DSP process) with intensity process \(\lambda_t\) is a point process \(y_t = p_{\Lambda_t}\) satisfying the following expression for the conditional distribution of the n-th jump

(41)\[\begin{equation} {\mathbb P}\left(\tau_n > T\right) = {\mathbb E}_t\left[e^{-\Lambda_{t,T}} \sum_{j=0}^{n-1}\frac{1}{j!} \Lambda_{t, T}^j\right] \end{equation}\]

The intensity function of a DSPP is given by:

(42)\[\begin{equation} {\mathbb P}\left(N_T - N_t = n\right) = {\mathbb E}_t\left[e^{-\Lambda_{t,T}} \frac{\Lambda_{t, T}^n}{n!}\right] = \frac{1}{n!} \end{equation}\]

from quantflow.sp.dsp import DSP, PoissonProcess, CIR
pr = DSP(intensity=CIR(sigma=2, kappa=1), poisson=PoissonProcess(intensity=2))
pr2 = DSP(intensity=CIR(rate=2, sigma=4, kappa=2, theta=2), poisson=PoissonProcess(intensity=1))
pr, pr2

(DSP(intensity=CIR(rate=1.0, kappa=1.0, sigma=2.0, theta=1.0, sample_algo=<SamplingAlgorithm.implicit: 'implicit'>), poisson=PoissonProcess(intensity=2.0)),
 DSP(intensity=CIR(rate=2.0, kappa=2.0, sigma=4.0, theta=2.0, sample_algo=<SamplingAlgorithm.implicit: 'implicit'>), poisson=PoissonProcess(intensity=1.0)))

import numpy as np
from quantflow.utils import plot
import plotly.graph_objects as go

n=16
m = pr.marginal(1)
pdf = m.pdf_from_characteristic(n)
fig = plot.plot_marginal_pdf(m, n, analytical=False, label=f"rate={pr.intensity.rate}")
plot.plot_marginal_pdf(pr2.marginal(1), n, analytical=False, fig=fig, marker_color="yellow", label=f"rate={pr2.intensity.rate}")
fig.add_trace(go.Scatter(x=pdf.x, y=pr.poisson.marginal(1).pdf(pdf.x), name="Poisson", mode="markers", marker_color="blue"))

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

(1.9999909680081247, 4.689459199192303)

pr2.marginal(1).mean(), pr2.marginal(1).variance()

(1.999989818527503, 5.046045401891716)

from quantflow.utils.plot import plot_characteristic
m = pr.marginal(2)
plot_characteristic(m)

pr.sample(10, time_horizon=10, time_steps=1000).plot().update_traces(line_width=1)

m.characteristic(2)

(0.000708775281497072+0.013619153217117782j)

m.characteristic(-2).conj()

(0.000708775281497072+0.013619153217117782j)