Spectral divide-and-conquer MCMC for long stationary time series
arXiv:2609.23985
Abstract
Bayesian inference via Markov chain Monte Carlo (MCMC) methods can be prohibitively slow for long time series. Distributed Bayesian inference across multiple cores or machines is non-trivial since simple data splits in time perform badly under temporal dependence. We show that a systematic partitioning of the periodogram observations in the frequency domain gives a fast and accurate posterior approximation for stationary series that can also be used when the computations exceed the memory capacity of a single computational node. Our method uses the asymptotic independence of periodogram observations in the Whittle likelihood to develop a distributed fast Fourier transform combined with embarrassingly parallel Markov chain Monte Carlo (MCMC) algorithms in modern cluster-computing frameworks. We establish that the error of the spectral divide-and-conquer MCMC posterior approximation relative to the exact time-domain posterior converges to zero in probability in a shrinking neighbourhood of the full-data Whittle posterior mode. The proposed method is shown to be competitive with the current state-of-the-art distributed MCMC for time series when dependence is weak, and substantially more accurate in a highly persistent setting. The methodology is used to perform a rolling forecast evaluation of one-second market liquidity costs from histories of up to 24 million observations.
arXiv commentSubstantially revised version with a new empirical application and updated numerical results