First annual distinguished lecture by David Spiegelhalter
Don't know, can't know: Communicating risk and deeper uncertainty
David Spiegelhalter, Winton Professor of the Public Understanding of Risk, University of Cambridge
David Spiegelhalter, Winton Professor of the Public Understanding of Risk, University of Cambridge
Statistical or machine learning involves predicting future outcomes from past observations. Many present day applications involve large numbers of predictor variables, sometimes much larger than the number of cases or observations available to train the learning algorithm. In such situations traditional statistical methods fail.
We prove that the only tail risk measure that satisfies a set of economic axioms proposed by Schmeidler (1989, Econometrica) and a statistical requirement called elicitability (i.e. there exists an objective function such that a reasonable estimator must be a solution of minimizing the expected objective function) is the median shortfall, which is the median of the tail loss distribution and is also the VaR at a high confidence level.
Likelihood methods provide one of the most versatile and effective ways to handle data. They give us tests and confidence intervals with very strong optimality measures. But the cost for using them is usually that we have to know a family of distributions generating our data.
This lecture provides an overview of the real options approach to valuation mainly from the point of view of the author who has worked in this area for over 30 years. After a general introduction to the subject, numerical procedures to value real options are discussed.
Some statistical issues related to the monitoring of surgical quality will be reviewed in this presentation. The important role of risk-adjustment in healthcare, used to account for variations in the condition of patients, will be described. Some of the methods for monitoring quality over time, including a new one, will be outlined and illustrated with examples.
Statistical models in which the ambient dimension is of the same order
or larger than the sample size arise frequently in different areas of
science and engineering. Examples include sparse regression in
genomics; graph selection in social network analysis; and low-rank
matrix estimation in video segmentation. Although high-dimensional
models of this type date back to seminal work of Kolmogorov and