search

kellijohnson-NOAA / AR-perf-testing

Investing autocorrelated recruitment deviations in Stock Synthesis
0 stars 1 forks source link

Bias in forecast #1

Closed kellijohnson-NOAA closed 8 years ago

kellijohnson-NOAA commented 8 years ago

Bias in forecast even when AR is fixed at the true value. To investigate the error in the bias, we are currently fixing steepness at 1.0 and running a small set of results. According to @James-Thorson

One potential source of bias that occurs in nonlinear models estimated using Empirical Bayes (i.e., where we're treating random effects, eg.., rev-devs, via maximizing the penalized likelihood) is that the expectation of a nonlinear transformation (i.e., the true mean) is not equal to nonlinear transformation of an expectation (i.e., our estimator):

E( f(theta) ) != f( E(theta) )

where theta is our rec-devs, and f is the assessment model. If steepness = 1, it eliminates one type of nonlinearity from the forecasted dynamics (the other is individual growth). So if our bias is from this nonlinear transformation issue, fixing steepness at 1 might decrease or eliminate it.

The file to run this test can be found in the following commit: 5d5341adbcc3c8e4620a533c25666632ea2c393d

kellijohnson-NOAA commented 8 years ago

Currently running iterations which will be done in an hour or so, but I will have to sort through some plotting code to figure out how to view the results as it is not typical ss3sim output.