Time Series analysis support

[xieliaing]

time series has a wide range of real world applications and is used literally by all businesses. Currently there is no formal support for time series analysis. Is there any plan for this?

[cesarsouza]

Hi @xieliaing,

Many thanks for opening the issue!

I completely agree with your remarks. But answering your question: Yes, there are plans to add support for time series analysis in the near future. Currently the framework is already able to perform classification of time series through Support Vector Machines, Hidden Markov Model and Hidden Conditional Random Fields classes, but it is true that it may not still be able to perform regression and more general analysis like in ARIMA models.

As such, may I ask exactly which method you would be more interested in seeing the framework support in the future? If possible, please add some references for models you think would be the most useful in the real world so we could prioritize what should be implemented first.

Best regards,
Cesar

[cesarsouza]

Also, if you would like to contribute a method yourself, please do not hesitate to send in a pull request!

Best regards,
Cesar

[CatchemAL]

Ooh interesting. I'll follow this thread closely. As @cesarsouza says, a little clarity on what time series analysis would particularly interest you would be good to know.

If it's of any interest, I was planning to add EWMA and exp weighted moving variance-covariance calculations to the framework at some point in the next couple of months. If you'd really like it, I can probably add them sooner (or you're very welcome to contribute if you'd prefer as César also mentioned).

[xieliaing]

@cesarsouza @AlexJCross
Thanks for the response. Yes, time series analysis is a big topic and we can choose something small and most frequently used as starting point.

I see these as two categories:

1. Distribution and descriptive analysis. For example, various filters, stationarity test (adf, kpss, and etc), calculation of periodogram, and calculation of ACF, PACF, and etc. I have an example of ACF using Fourier Transformation from Accord.Audio. This class of methods will help analysts like me to conduct formal time series study much more easily.
2. Time series modeling, from EWMA to X13 to ARIMA or even State space model. A formal implemetation of MLE based ARIMA model will definitely help. currently I use an iterative linear regression process to approximate ARIMA (a method used in 80's due to limitation of computation power).

I would love to contribute but I am not professional programmer, so my code may not be as good as one from a professional one.

[CatchemAL]

I would love to contribute but I am not professional programmer, so my code may not be as good as one from a professional one.

Well most of my code is terrible but that doesn't stop me :-)

Seriously though, if you'd like to add something, please do! Even if it's just the guts of an algorithm, I'm sure it would be very useful to a lot of people. I don't have much knowledge on the categories you listed above (I'm no statistician); I will, however, add EWMA this week.

Thanks,
Alex

[xieliaing]

I am not good at class design, but I can follow the design of StatsModel in Python. A good start point is the tsa\stattools.py. I can translate that into C#. Many building blocks are already available in Accord.Audio which I still need to get familiar with first.

[cesarsouza]

Hi @xieliaing,

That could be a good start! In fact, the StatsModel library is under the 3-clause-BSD license, so it should be fine to base implementations on it. It could be possible to start with simple translations of the basic machinery to support hypothesis tests such as Ljung–Box, ADF, KPSS. Most of the methods that will be used in those tests would likely be implemented as static methods in some static class anyway, so the design would not differ very much from the Python file you mentioned.

For the tests themselves, I can wrap them into classes later, following the rest of the style of the framework, adding them to a new TimeSeries namespace under under Accord.Statistics.Testing.

Please don't worry about design at this stage! It would be better to start with working implementations first so we could write enough unit tests to be able to change the design at will later, without risking introducing bugs.

Best regards,
Cesar

[xieliaing]

Sounds like a plan. What I can do now is providing C# implementation of these functions first, test them. Later these functions can be incorporated into a well designed TimeSeries class.

[xieliaing]

@cesarsouza Any guidelines on pull requests and check in code? Code style?

[cesarsouza]

@xieliaing Yes, the contributing guidelines are here. For code style, I would say it would be preferable to stick to the original formatting guidelines provided by Visual Studio (i.e. the formatting applied when hitting Ctrl+E, D) since it is probably the most common format out there.