Abstracting from the specific code I would like to highlight the endogeneity problem here - generalising statments such as the premise of this topic is exactly the cause for such topics 
Yes, there are (probably many) instances where Julia is faster than Matlab (or other languages) but the premise that it should “usually” be faster is misleading (espeically in the context of this topic it appears as if it should “always” be faster). I state that because I was also misled in the sense that when I started learning Julia I was expecting all those gains in speed, even for relatively simple tasks exactly beacuse of the paper of Villaverde or the comparisons on the website.
If you did DSGE models with VFI as in the mentioned paper AND it came out slower than comparable languages it would indeed be interesting and puzzling. However, once you start doing other things it becomes much more complicated to predict which is faster.
I work in time series, especially bayesian time series econometrics and spend a lot of time waiting for my code. I tried switching to Julia only to find that it is slower to Matlab also in my instances. I then looked around and asked experienced Julia programmers (in person, but I think I also had some discussions here) for help and in my case it turned out it cannot get faster and it all depends on the BLAS in the background.
Remember also, that code might be slower for two reasons:
- the language is slower
- the code is not written well for the language.
No. 2 is BIG for me, I have over 13 years experience with Matlab and at best half a year in Julia. I am sure I write suboptimal Julia code because I am plagued with doing stuff necessery in Matlab but unneeded here and because I don’t know what is the best way to do stuff in Julia.To compare 1) objectively you need to write the code in the best way possible for each language and account for the time of writing the code in that manner.
All I am saying is that it appears that you think that Julia should be faster per se, exactly as I did and I think we should watch out as this sets expectations for new users.