Probably the millionth time asked here but I'm curious what people today think.
I need to use one of the two languages for my grad macro course next semester. Professor recommends Julia since he uses it, but I know Python is more widely used.
What do you think is more worthwhile to learn? I'm at a LRM university so there's a good chance I will go on to tech/industry. Is Julia valuable in industry/do you foresee it becoming valuable 5+ years from now?
The 20-something code-chimpanzees who lack a high school education will say Julia.
The PhD researchers will so the language doesn't matter. Use the one that already has robust, bug-free, reliable libraries that have been used and tested by thousands of others, that does what you need so you can spend all of your time on your research rather than reinventing the wheel, and writing and debugging code.
Fair comments above but it also depends on what you are doing.
If (eg) something structural where you mostly can’t count on prewritten libraries and speed matters then julia is worth a strong look.
But it is true that the language doesn’t matter that much. I find both languages easy to use, but prefer Julia despite the absence of certain libraries.
Python for sure if you are going to industry:
https://www.tiobe.com/tiobe-index/
If you are going to work on large scale macro models and/or Bayesian inference (MCMC sampling), it may also be beneficial to learn C down the road to write high performance code. But do try to exhaust Python options first.
they are not so different, in fact julia was designed to look similar to python
in fact, learning a language for the majority of people involves learning how to do loops, conditionals, creating functions, maybe, MAYBE building classes of more complex data types
everything else boils down to using the libraries to do what you need to do, so, pick the one that does the job. Since your professor is goint to use julia, you use julia
If you are going to work on large scale macro models and/or Bayesian inference (MCMC sampling), it may also be beneficial to learn C down the road to write high performance code.
You might be able to write your own Metropolis-Hastings sampler in a few lines, but you'll need a PhD in differential geometry to code an efficient HMC sampler. Nobody does that.
If you are going to work on large scale macro models and/or Bayesian inference (MCMC sampling), it may also be beneficial to learn C down the road to write high performance code.You might be able to write your own Metropolis-Hastings sampler in a few lines, but you'll need a PhD in differential geometry to code an efficient HMC sampler. Nobody does that.
I have never looked into HMC but the common variety of Bayesian problems can be tackled with the Gibbs sampler, possibly with embedded MH steps. For multimodal problems in macro SMC (sequential Monte Carlo) is becoming popular and that is easily implemented in any language. But for problems where the likelihood is expensive to evaluate there are massive speedups in C relative to Python, Matlab or even Julia.
1. R or Stata for reduced form econometrics and time series
2. Julia for structural and macro
3. Python for almost everything else.
A good general approach, but if you work overlaps several areas then it is preferable to do most of the work in one, maybe two languages. Since Python + various numerical/statistical libraries covers a huge amount of use cases it is the natural choice. Also keep in mind Stata is closed source and licensed, and R has taken an awkward direction with part of the community pushing tidyverse and related packages, which many people with traditional computer background find extremely awkward.
Julia suffers from the problem faced by all young languages, i.e. the comparative lack of long-established, trusted libraries for linear algebra, optimization, etc. Naively written Julia is also not as fast as often touted; to achieve high speed one needs to write "unpythonic" code adding type declarations, loops, etc. In other words, one needs to write C-like code to obtain C-like speeds. But even then there remains various overheads vs. native code. This is why where performance really matters (video games) you won't ever see Julia used.
For industry, Python is a no brainer. But even the academic community has largely pivoted towards Python. For instance the first edition of this ML book
https://probml.github.io/pml-book/book1.html
used Matlab 10 years ago. The newer edition replaced all Matlab with Python.
I have never looked into HMC but the common variety of Bayesian problems can be tackled with the Gibbs sampler, possibly with embedded MH steps. For multimodal problems in macro SMC (sequential Monte Carlo) is becoming popular and that is easily implemented in any language. But for problems where the likelihood is expensive to evaluate there are massive speedups in C relative to Python, Matlab or even Julia.
Nobody cares if your time-to-run is 10% less if your time-to-code increases by 10x.