Particle filtering is a technique used for solving filtering problems involving partially observable Markov decision processes (POMDPs) or, more generally, problems with nonlinear, non-Gaussian state-space models. In Stochastic processes and quantitative finance, it is commonly used for state estimation and model calibration, where the target is to determine the systemβs state and parameters from noisy observations. We illustrate the concept using the following key steps.
1. **Model Formulation**: In a stochastic process, a system evolves over discrete time steps tβ=β1,β2,ββ¦,βT. Let xt denote the hidden state of the system at each time step, and yt be the observable measurement at time t. The state-space model can be expressed by two equations:
State Transition Model: xtβ=βf(xtβ ββ 1,βut,βwt),
Observation Model: ytβ=βg(xt,βvt),
where wt and vt are the state and observation noise respectively, and ut is any external control input to the system.
2. **Sequential Bayesβ Theorem**: The estimation problem involves finding the posterior distribution of the system state xt given the history of measurements up to time t: p(xt|y1β:βt). This can be computed recursively using Sequential Bayesβ theorem, which involves two steps:
* Time Update (Prediction): estimate the prior by integrating over the state transition model
p(xt|y1β:βtβ
ββ
1)β=ββ«p(xt|xtβ
ββ
1)p(xtβ
ββ
1|y1β:βtβ
ββ
1)dxtβ
ββ
1
* Measurement Update (Correction): update the prior using the current observation yt
$$p(x_t | y_{1:t}) = \frac{p(y_t | x_t) p(x_t | y_{1:t-1})}{p(y_t | y_{1:t-1})}$$
3. **Particle Filtering**: The challenge in the above sequential Bayesian estimation is to compute complex integrals and posterior distributions. The Particle Filtering technique (also known as Sequential Monte Carlo methods) offers an approximate numerical solution to this problem. The main idea is to represent the posterior distribution of xt using a set of N random particles or samples (also known as the empirical distribution), denoted as xt(i), each carrying a weight wt(i). The algorithm involves the following steps:
* Initialization: Sample particles from the initial state distribution p(x0) and assign initial weights $w_0^{(i)} = \frac{1}{N}$.
* Prediction: Sample particles from the state transition model xt(i)ββΌβf(xtβ ββ 1(i),βut,βwt).
* Update: Compute the importance weight for each particle based on the observation model yt
wt(i)βββg(yt|xt(i))
* Normalization: Normalize the importance weights to sum up to 1.
$$w_t^{(i)} = \frac{w_t^{(i)}}{\sum_{j=1}^N w_t^{(j)}}$$
* Resampling: If the effective sample size $\sum_{i=1}^N \left(w_t^{(i)}\right)^2$ falls below a predefined threshold, resample the particle set by selecting particles proportional to their normalized weights. This helps avoid particle "degeneration" and maintains a diverse set of particles.
In the context of quantitative finance, particle filtering techniques can be used for various tasks, such as asset allocation, model calibration, and risk management. For example, consider a stochastic volatility model in which the return rt and the latent volatility Οt are governed by the following dynamics:
$$\begin{aligned}
r_t &= \sigma_t \epsilon_t, \quad \epsilon_t \sim N(0, 1) \\
\sigma_t &= \phi \sigma_{t-1} + \eta_t, \quad \eta_t \sim N(0, \beta^2)
\end{aligned}$$
Here, Ο and Ξ² are model parameters that control the volatility dynamics. Using stock market data as observations, one could estimate both the latent volatility state Οt and model parameters by incorporating them into the framework of particle filtering.
In conclusion, particle filtering is a widely used technique for state estimation and model calibration in the context of stochastic processes and quantitative finance. It provides an approximate solution to Bayesian estimation problems with a set of random particles, allowing convenient implementation and adaptation to various models in the financial domain.