ar() method for fitting an autoregressive time series model

[papamarkou]

Fitting an autoregressive time series model to data (by default selecting the complexity using AIC) is needed in many different occasions beyond sole time series analysis. For instance it is used for computing MCMC convergence diagnostics and in other cases. R considers the function so important that it ships with its "base" code rather than being part of a package. My view would be that we need this function. Before embarking on coding it, I wanted to discuss it and get your views - do you also find it worth having it somewhere in JuliaStats? If yes, would you want it in Stats or in a separate JuliaStats package called TimeSeries?

[andreasnoack]

I think we should start a TimeSeries package under JuliaStats and see if we can agree on some core functionality there. I am not very satisfied with the time series functionalities in R's Stats and therefore I don't think we should see that as our goal, but I am fine with an ar function similar to the one in R.

[StefanKarpinski]

My rather limited understanding is that Pandas has good time series support, so perhaps that can serve as more of a design guide than R.

[nfoti]

The python module statsmodels has good time series support (it's another Wes McKinney project) that would be good to look at.

Sent from my iPhone

On Oct 13, 2013, at 11:11 AM, Stefan Karpinski notifications@github.com wrote:

My rather limited understanding is that Pandas has good time series support, so perhaps that can serve as more of a design guide than R.

— Reply to this email directly or view it on GitHub.

[papamarkou]

I agree that a standalone TimeSeries package would be a petter place than Stats. I haven't used Pandas or Python's statsmodels package before, but they look a good start at a first look. TISEAN has several routines for non-linear time series analysis (which is a somewhat more specialized toolbox). The functionality I have so far in mind is:

• ARMA models,
• ARIMA models,
• some elementary operations such as differencing and other tools for removing trend or seasonality,
• volatility related functions,
• Lyapunov exponents and similar functions for exploring non-linear dynamics,
• maybe the partial-, auto-, cross- correlations could be migrated from Stats to TimeSeries.

This list is just a start, any help to extend it would be great. Shall we create a TimeSeries template under JuliaStats in the near future as a placeholder and add code incrementally depending on what functionality we agree to support?

[papamarkou]

Thank you for your replies @AndreyKolev, @milktrader. I thought I may reference the older thread JuliaLang/METADATA.jl#336 here, so that we all join the discussion. @andreasnoackjensen, @StefanKarpinski, @nfoti, @johnmyleswhite, @dmbates, @simonster, shall we start forming an idea of what would initially go in a TimeSeries JuliaStats package, and shall we open a project placeholder under JuliaStats to slowly start making some progress? Please include in the discussion anyone I may had missed to add.

[StefanKarpinski]

That sounds like an excellent idea to me.

[milktrader]

I'm happy to move the TimeSeries package over any time. Presumably I need agreement to change the ownership? Also, happy to do the same with TimeModels, which can be a good placeholder for modeling time series until we decide whether it's one package (along with TimeSeries) or two separate ones.

[StefanKarpinski]

Someone in the JuliaStats organization needs to give you admin privilege while you do this.

[papamarkou]

My feeling is that the consensus will be to have a single package TimeSeries, let's wait to see what others will say :+1:

[milktrader]

TimeSeries now is just a handful of utility methods and the title is a bit regal for such a package, so I'm on board with placing time modeling into it.

[milktrader]

Here is a link to the non-trivial code in TimeModels. https://github.com/milktrader/TimeModels.jl/pull/1

[andreasnoack]

I think it would be a good idea to move the TimeModels/TimeSeries repo over to JuliaStats. Maybe it is an idea to keep the two packages separated. The problem of handling time series data is really separate from setting up the statistical modeling functionality.

[papamarkou]

It's true that these are two different functionalities. If a time series type would be designed to be really fundamental (as fundamental as handling even date-time data currently in Datetime, which would be used nearly as broadly as data arrays), it makes sense. Otherwise, handling time series would probably be rather trivial to be justified as a separate entity. How shall we proceed to move the repos over to julia stats?

[milktrader]

I just need an invitation and I'l do it. Let's start with them being separate packages. I'm currently hacking around a TimeFrame type in the TimeFrame branch of TimeSeries. https://github.com/milktrader/TimeSeries.jl/issues/26

[papamarkou]

I am ok with the suggestion to keep them separate. I don't know how to invite you - @andreasnoackjensen, @StefanKarpinski, if you agree, would you like to authorize @milktrader to make the transfer?

[milktrader]

If you have admin privileges, this should do it https://help.github.com/articles/how-do-i-add-a-collaborator

But I've not done it before so not sure.

[papamarkou]

To be honest, because I was added to JuliaStats about 4 months ago, I don't know if I would be allowed to take the liberty and add someone without the consent of one of the older members - Stefan are you ok with this?

[johnmyleswhite]

Please just make a specialized team that has access to the time series stuff. In that case, there's literally no possibility of harm.

[StefanKarpinski]

I don't even have push access to this repo, let alone admin rights in JuliaStats as a whole. @johnmyleswhite or someone else who's an admin here will have to do it. But, yes, I'm ok with this (as a spectator).

[johnmyleswhite]

I've created a team for TimeSeries. What else needs to be done? Transferring some repos to JuliaStats?

[papamarkou]

I followed these instructions https://help.github.com/articles/how-do-i-set-up-a- and the team management link didn't appear. @johnmyleswhite, would you like to do it as I think I don't have the privileges to complete it myself.

[papamarkou]

Yes, the TimeSeries and TimeModels packages need to be transfered to JuliaStats, thanks John.

[johnmyleswhite]

Ok. To do that, I'll give you and @milktrader full control over JuliaStats temporarily. After that's done, we'll go back to teams with more granular control.

[papamarkou]

Thanks a lot John. @milktrader would you like to transfer your 2 repos so that @johnmyleswhite can revert the admin rights afterwards?

[milktrader]

Sure.

[milktrader]

Okay, done. The only privies I'm interested in is with TimeSeries and TimeModels. I presume that's what the teams within JuliaStats organization will enable.

[papamarkou]

Good, I will close this issue now. ar() can be added in TimeModels.

[johnmyleswhite]

All the admin work is done now. @milktrader and @scidom have full control of TimeSeries and TimeModels, both now under JuliaStats.

[papamarkou]

Thanks John, only to double check, do I keep have admin access to MCMC?

[johnmyleswhite]

Yes, none of that should have changed. Let me know if something went wrong.

[papamarkou]

Cheers!