Bayesian Estimation of Time-Varying Regression with Changing Time-Volatility for Detection of Hidden Events in Non-Stationary Signals

O. Krasotkina, A. Kopylov, V. Mottl (Russia), and M. Markov (USA)


Nonstationary regression, parameterized a priori model, quadratic dynamic programming


Problems of signal analysis may be practically always considered as those of recovering some hidden dependences, which are time-varying in the general case. In many situations, the nostationarity mode of the dependence should be expected to change in the observation interval, maybe with some jumps or spikes. In this work, we consider a statistical framework and a family of respective algorithm for time varying linear regression estimation which preserve essential peculiarities in basically smoothly changing regression coefficients. The method being proposed is simple in tuning and has linear computational complexity with respect to the signal length. In particular, we show how this technique allows to watch the dynamics of the hidden asset composition of an investment portfolio from publicly available data with the purpose of detecting sharp changes in its investment strategy.

Important Links:

Go Back