Occurrence detection, correlation and classification among large numbers of time series

10 June 2016
Alexander Denev

Markit is a leading global provider of financial information services. We provide products that enhance transparency, reduce risk and improve operational efficiency.

We wish to find ways to automatically detect and label ‘extreme’ occurrences in a time series such as structural breaks, nonlinearities, and spikes (i.e. outliers). We hope to detect these occurrences in the levels, returns and volatility of a time series or any other transformation of it (e.g. moving average).

We also want to look for the same types of occurrences in the multivariate case in a set of time series through measures such as e.g. correlations, eigenvalues of the covariance matrix etc. The number of time series involved is of the order 3x10^6.

We wish to explain the appearance of an ‘extreme’ occurrence or a cluster of occurrences endogenously, as an event conditional on the values of the time series in the set, both contemporaneously and/or as conditional on their time lags.

Furthermore, we would like to classify the events that caused the occurrence in some major categories, if found e.g. shock to oil supply, general risk aversion, migrations etc. both algorithmically and by allowing human corrective judgement (which could become the basis for supervised learning).

  • Industrial and Interdisciplinary Workshops