Inferring rapidly-varying big data time series–from deconvolution to a new form of time series model

Traditional time series models can only encapsulate slow variation in the underlying generative mechanism. However, in many scenarios, this is not a realistic assumption. There seems to be an unavoidable conflict between how rapidly the structured part of the model can change, versus how much we need to average in order to retrieve parameters stably. We here introduce a new class of nonstationary time series, and show how efficient and rapid inference is still possible in this scenario, despite the generating mechanism changing quickly. The methods are illustrated on drifter time series, from the global drifter programme. Computational efficiency becomes a key constraint when handling 20000 long time series to obtain a global understanding of circulation, making this a big data problem. Depending on the latitude of the observations, the underlying generative mechanism of the observed phenomenon is either slowly or rapidly changing, and we show how the newly introduced methodology can resolve both scenarios.

This is joint work with Arthur Guillaimin, Adam Sykulski, Jeffrey Early and Jonathan Lilly

• Speaker: Sofia Olhede, UCL
• Wednesday 30 November 2016, 14:00–15:00
• Venue: MR5 Centre for Mathematical Sciences.
• Series: CCIMI Seminars; organiser: Rachel Furner.