BIMSA >
BIMSA Digital Economy Lab Seminar
BIMSA Digital Economy Lab Seminar
Effective Filtering of Multiscale Stochastic Volatility from Realized Measures: homogenization, dimensional reduction, and optimal Sampling
Effective Filtering of Multiscale Stochastic Volatility from Realized Measures: homogenization, dimensional reduction, and optimal Sampling
Organizers
Johansson Anders
,
Manyao Deng
,
Ruize Gao
,
Liyan Han
,
Zhen Li
,
Jin Liu
,
Fei Long
,
Dongbo Shi
,
Ke Tang
,
Xing Yan
,
Qi Zhang
Speaker
Time
Friday, September 4, 2026 3:00 PM - 4:00 PM
Venue
A3-2-303
Online
Zoom 242 742 6089
(BIMSA)
Abstract
Multiscale stochastic volatility models capture financial market dynamics through coupled slow persistent trends and fast mean-reverting fluctuations, with volatility observed through discrete-block realized measures of sampling length $\Delta$ in high-frequency financial econometrics.
This multiscale structure creates a fundamental filtering dilemma: because block averaging smooths out rapid movements, estimating the fast state directly is impossible and renders full high-dimensional computational cost. Conversely, simply ignoring fast noise creates persistent systematic bias in the slow trend due to nonlinear averaging.
To resolve these obstacles, we develop a rigorous homogenization framework in the observed length. We construct an asymptotically nonlinear filter on the slow state, preserving a closed-form corrector to eliminate the bias from fast fluctuations. By establishing dual Sobolev energy bounds, we prove a error rate and derive the optimal block scale for estimating the stochastic volatility.
Extensive numerical simulations and empirical validations on high-frequency market data are also shown to verify the theoretical convergence rates, bias elimination, and order-of-magnitude computational gains of the proposed filter.
This multiscale structure creates a fundamental filtering dilemma: because block averaging smooths out rapid movements, estimating the fast state directly is impossible and renders full high-dimensional computational cost. Conversely, simply ignoring fast noise creates persistent systematic bias in the slow trend due to nonlinear averaging.
To resolve these obstacles, we develop a rigorous homogenization framework in the observed length. We construct an asymptotically nonlinear filter on the slow state, preserving a closed-form corrector to eliminate the bias from fast fluctuations. By establishing dual Sobolev energy bounds, we prove a error rate and derive the optimal block scale for estimating the stochastic volatility.
Extensive numerical simulations and empirical validations on high-frequency market data are also shown to verify the theoretical convergence rates, bias elimination, and order-of-magnitude computational gains of the proposed filter.