BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Local scoring rules - Dr Matthew Parry\, Dept of Plant Sciences an
 d Statisitcal Laboratory\, DPMMS
DTSTART:20100505T131500Z
DTEND:20100505T141500Z
UID:TALK24773@talks.cam.ac.uk
CONTACT:Rachel Fogg
DESCRIPTION:Suppose you publicly express your uncertainty about an unobser
 ved quantity by quoting a\ndistribution for it. A scoring rule is a specia
 l kind of loss function intended to measure the quality\nof your quoted di
 stribution when an outcome is actually observed. In Bayesian decision theo
 ry\, you\nseek to minimise your expected loss. A scoring rule is said to b
 e proper if the expected loss under\nyour quoted distribution is minimised
  by quoting that distribution. In other words\, you cannot\ngame the syste
 m!\nIn addition to having a rich theoretical structure – for example\, a
 ssociated with every scoring\nrule is an entropy and a divergence function
  – scoring rules can be tailored to the problem at\nhand and consequentl
 y have a wide range of application. They are used in statistical inference
 \,\nfor evaluating and ranking weather and macroeconomic forecasters\, for
  assessing the quality of\npredictive distributions\, and in student exami
 nations.\nI will discuss a class of scoring rules with the rather attracti
 ve property that the quoted distribution\nneed only be known up to normali
 sation. This property is a consequence of requiring\nthe scoring rule to b
 e “local”\, i.e. the score may depend on the quoted distribution at un
 realised\noutcomes but only if those unrealised outcomes are “close” t
 o the actual outcome. After completely\nspecifying local scoring rules for
  both discrete and continuous outcome spaces\, I will consider application
 s\nto missing data problems\, the connection to the pseudolikelihood metho
 d\, and the recent\nwork of Hyvärinen et al. in machine learning.\nThis i
 s joint work with Philip Dawid and Steffen Lauritzen.
LOCATION:LR6\, Engineering\, Department of
END:VEVENT
END:VCALENDAR