Seminar: Modelling timescales and efficient inference for Lagrangian time series

SpeakerAdam Sykulski
DateThursday, 01 Nov 2012
Time16:00 - 17:00
Location Room 102, Department of Statistical Science, 1-19 Torrington Place (1st floor)
Event seriesStatistical Science Seminars

Modern acquisition of real-world time-varying processes results in large heterogeneous datasets whose features operate at different and constantly changing timescales. It becomes very important in such settings to propose appropriate statistical models and inference methods that scale, such that meaningful summary statistics can be determined which characterise the key structure of the data. In this talk we propose applying such modelling and inference procedures to Lagrangian time series obtained from an extensive global dataset of surface drifter observations. These heterogeneous time series require stochastic models that are flexible enough to capture heterogeneity, but simple enough such that time-variability can be inferred from relatively few observations. For this reason our proposed model does not fully describe the generation of the data, but is a second-order model that accounts for the key variability in the time series.

Inference is addressed using local Whittle estimation in the Fourier domain, where we adopt a semi-parametric approach to correctly account for the fact that parameters of continuous models are being inferred from irregularly sampled discrete data. We present numerous compelling examples demonstrating that the temporal variability of drifter time series can be neatly captured using our methodology, which could potentially impact on the use of oceanic data in global climate models.

iCalendar csml_id_120.ics