Spectral divide-and-conquer MCMC for long stationary time series
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.