Wednesday, 19.12.2007, 10:30
Room 527 Bloomfield Bld., Technion
We study the problem of testing an expert who is required to submit
forecasts that have a learnable parametric representations. The class of
stochastic processes with such representations, introduced by Jackson,
Kalai and Smorodinsky (1999), include exchangeable and Markov process as
special cases, and encompasses all processes used in Bayesian statistics.
We define a simple test that screens experts whose forecasts belong to this
class. The test stipulates an initial phase during which the expert may
choose to learn from data. At the end of this phase, the expert's
conditional forecasts are tested according to a simple frequentist
criterion. We show that this test passes an informed expert, but that it
cannot be strategically manipulated by an uninformed one.