Lunch Meeting 2014 Nov 12

12 Nov 2014

Chris, Pablo, Nathan, Moe's 11:30 lunch meeting.

Items Covered

• Got an update on the number of people attending.
  • See updated excel sheet from Pablo.

A key point: the goal here is to explicitly create a toolkit for "bottom-up" heterogeneous-agent models. This is not a toolkit for any of the many many other thing that could exist for computational economics (well at least not initially) -- but rather, the 'bottom-up' approach to

Key thing: modules: we think we have a good first-pass idea for how to arange a decentralized, modularized, "bottom-up" approach to coding heterogeneous-agent models. Key questions to answer:

• How do we get people to volunteer to be a part of this?

• How do we figure out what the other modules need to be?

Here's an initial answer: "driven by implementation of specific problems -- both 'classic' models useful for current work, and new cutting-edge research in progress."

• We flesh out the initial modules we need by executing some basic examples. This is already underway with Chris' MicroDSOP notes and with the MPC Krusell-Smith model.

• We flesh out new modules by implementing specific problems; new modules will emerge from that process.

  • There are a number of models which would be great to have for policy and general community work

• Graduate students are a prime group for implementing many of these. The Q: how to encourage volunteers?

  • via advisors

  • via mentoring relationships (from people involved in the toolkit project):

    • Eg CGE dissertation
    • mentors would not be responsible for / in charge of the dissertation, but would assess the code that the student writes, with the goal of guiding code to the point of being admissable to the toolkit.

  • financial incentives: "bounties" for particular problems solved

    • possibly small "dissertation continuation" grants?

  • earlier ideas:

    • student contests, eg as hosted by SCE at CEF conferences.
    • something like the replication wiki, in way that students could get cited.

  • NEW IDEA -- an extremely simple, low-cost "ejournal" which consists entirely of IPython (or IJulia) notebooks, which are "vignette-length" and demonstrate the code itself. They can additionally be, eg., technical appendices to published papers, but they need not be -- especially if they are notable contributions to something like the "bottom-up heterogeneous-agent toolkit."

    • Have JEDC publish these as "computational appendices:" in the back of the journal, simply have a page with the title, perhaps an abstract, and a link to the notebook.
    • The notebooks are still peer-reviewed, and being included in JEDC means they are also cite-able.
    • Code practices can increase the ease in which code is peer-reviewed. For example, adding small template-style unit tests will give a reviewer an immediate testing framework (which can often immediately illustrate how the code is practically used, and reviewers could adjust and explore the tests as much or little as desired).

Observation: many of the libraries exist -- but what is missing is a common framework to glue these together in a modular fashion.

Use cases of this effort:

• peer implementations of some aspects that are needed

• "private libraries" put out in public by authors.

• It would be great to have an example of the toolkit up and in-progress by the workshop. The v 0.1 of Python code is doable; a v 0.1 of the MPC macro model is likely not doable by the workshop, but perhaps by the conference next year.

  • Other possible example (as target): the "bequest motives" paper

• TOPIC: Nature of next steps

  • eg connect with SCE

• TOPIC: making a toolkit universally available. Two options:

  • a virtual machine
    • "require stack (A), else can use virtual machine (B)
  • Python Conda and distribution management
    • TravisCI and continuous integration
  • [sebastian]
  • Who will prompt discussion / MC?

• TOPIC: IPython notebooks and use to disseminate research/code

  • Sylvain

• Top-down TOPIC: classification of problem-types

  • Dynare-style: only uses perturbation method
  • cont-state, cont-control
  • mixed-state, cont-control
  • perhaps a 4th

• TOPIC: list of models we'd like to produce with the toolkit

  • From CFPB [Chris elicits these]

  • From OFR

  • From Fed

  • From IMF

  • From others

  • Will want to get this list up quickly and passed around.

• TOPIC: Licenses?

  • Public domain for govt work
  • others?

• TOPIC: Encouraging citations and peer review of code (as incentives for uptake and use)

  • Example: IPython / IJulia computational appendix, or "computational letters"