Forecasting with Augur

[elisachisari]

Forecasting with Augur

Augur is our official DESC forecasting code. We plan to make some advances in infrastructure, features and docs.

Contacts: Elisa Chisari Time: Morning Main communication channel: #desc-mcp-forecast GitHub repo: https://github.com/LSSTDESC/augur In-person/Virtual/Hybrid: Hybrid Zoom room (if applicable): Zoom room 1

Goals and deliverable

-Begin implementing link between new firecrown and augur. -Revisit bias implementation in current augur (can we make it work with the old version?). -Go over draft docs and check feasibility of implementing forecasting of model selection criteria as well.

Resources and skills needed

Python, CCL, Firecrown, general MCP knowledge. Newcomers welcome!

Detailed description

[add additional details about the project]

[rmandelb]

Hi Augur sprinters -

One thing I was wondering about was whether any existing Fisher codes and implementations of the Fisher bias formulae may be of use for your sprint? My (former) student Husni developed a Fisher code for a DESC project (which has a paper in draft form) that has gone into another submitted DESC paper led by @ztq1996 , and I believe @nikosarcevic and @c-d-leonard may be using it as well. He did a lot of tests of this code, both internally and against the DESC SRD Fisher, and against MCMC. If there's interest in looking at his Fisher or Fisher bias formula implementation, and potentially integrating some of it or using it as a benchmark/comparison, I'd be happy to discuss!

[rmandelb]

Pinging also @fjaviersanchez - sorry for not including you previously, and please check out the above message ^

[fjaviersanchez]

Thanks @rmandelb, no worries! It would be great if we could chat for a bit. I'd be more than happy to check the code and implementation, or use it as benchmarking. Currently, there are two Fisher options in augur: one that calls cosmosis, and one that computes the numerical derivative of the likelihood via numdifftools. The former was benchmarked against the DESC SRD, the latter, seemed to sometimes show numerical convergence issues. As for bias, I started the implementation of @biancasersante's derivation, but I haven't benchmarked it, so having a reference would be ideal! In any case, happy to talk more via slack or during the sprint :)