|
|
Parameter estimation using forecast skill
|
When
|
2.00pm on Friday 23rd November
|
|
Where
|
B617, Leverhulme Library, Columbia House
|
|
Presentations
|
|
|
Speaker
|
Hai Liang Du
|
|
From
|
London School of Economics & Political Science
|
|
Abstract
|
Traditional approaches to parameter estimation are not optimal when applied to nonlinear models. Even in the perfect model class scenario, methods based on least squares (and total least squares) are fundamentally biased (see McSharry and Smith, 1999). A simple approach to nonlinear modeling is presented which, by looking at the ensemble forecast performance, avoids the shortcoming of both least squares and total least squares parameter estimation. The ensemble initial condition is constructed with simple inverse noise and probabilistic forecast skill is used to rank parameter values. Better constructed ensemble of initial conditions would allow future improvements as well looking at the forecast performance in different lead times, and comparing shadowing times. Our approach will be illustrated on chaotic maps; applications in more complicated settings (even up to that of operational weather models) are also underway.
|
|
For further information
|
Thomas Hewlett (Postgraduate Administrator) Ext. 6879
Department of Statistics, Columbia House
|
|
|