No free lunch when estimating simulation parameters
We estimate the parameters of 21 simulation models using 9 estimation algorithms to discover which one is better at matching simulations with data. Unfortunately no single algorithm minimizes estimation error for all or even most the estimation tasks; instead algorithm varies for each simulation, and sometimes for each parameter of each simulation. Fortunately it is straightforward to use cross-validation to match to each simulation the best estimation algorithm for it. In terms of confidence intervals the results are more clear: bootstrap generates more precise prediction intervals than either quantiles or Approximate Bayesian Computation.
9 May 2019, 15:00 (Thursday, 2nd week, Trinity 2019)
Manor Road Building, Manor Road OX1 3UQ
Seminar room A
Ernesto Carella (Oxford Martin School, Oxford University)
Organiser contact email address:
Members of the University only