search

davidwhogg / Undetectable

the next generation of Measuring the Undetectable
0 stars 0 forks source link

add correlated noise #9

Open davidwhogg opened 11 years ago

davidwhogg commented 11 years ago

use a GP to generate it

eggplantbren commented 11 years ago

Cool idea, but if you then fit the data with an iid normal likelihood function you can get a misleadingly narrow posterior distribution. There are ways around that (apart from using a slower likelihood that allows correlations), I would like to discuss this when I visit.

davidwhogg commented 11 years ago

love to discuss -- this is the whole point -- we are often wrong about our noise, treating it as iid when it isn't, or treating it as having known variance when it doesn't; we should understand the failure modes.