How to start contributing - GSoC

[Harshdeep Harshdeep]

Hi,

I am Harshdeep. I am planning to contribute to statsmodel and would like to know how to start contributing to it? I was interested in the project idea - Automatic Forecasting for GSoC 2018.

Thanks.

[josef...@gmail.com]

On Tue, Feb 13, 2018 at 7:08 PM, Harshdeep Harshdeep <harshc...@gmail.com> wrote:

Hi,

I am Harshdeep. I am planning to contribute to statsmodel and would like to know how to start contributing to it? I was interested in the project idea - Automatic Forecasting for GSoC 2018.

Hi Harshdeep,

See the second thread today for general getting started with statsmodels.

about the forecasting proposal:

Chad will be the main mentor for this topic and will know more.

The topic is a bit a collection of heterogeneous functions trying to find the "best" forecasting model. We have some helper functions just for the order selection in ARMA type models but not much more.

There is auto.arima in R and various additional diagnostic functions to identify outliers, transformations like box-cox, identifying trend and seasonal patterns, ...

Many of the forecasting facilities in R are based on work by Hyndman which should have more explanations.

What can be done here in a GSOC project will need to be defined in more details when writing the proposal. I think it will be mostly on top of the current models and it will not be necessary to understand all the details of the models.

One example that includes several model variations and options is https://github.com/statsmodels/statsmodels/pull/2892 which is missing unit tests.

Also Holt-Winters and exponential smoothing is in master. This and the unobserved components statespace model allow for a large number of options and variations, but there is currently no support to choose the model specification automatically or semi-automatically.

Aside: on stats.stackexchange https://stats.stackexchange.com/users/3382/irishstat looks at many examples and advertises his forecasting package. His comments and answers might give some ideas about potential model choice problems.

Josef

 

Thanks.

[josef...@gmail.com]

Here is a related issue, that I just found again, with a link to a stats.stackexchange question with answers by IrishStat and Hyndman

https://github.com/statsmodels/statsmodels/issues/2571

Josef

 

Josef

 

Thanks.

[Chad Fulton]

Hi Harshdeep,

I'm glad to hear you're interested in a GSOC with Statsmodels, and the automatic forecasting project in particular. I will try to post more details soon about my thoughts for how this project might go.

Since the proposal must be created by the student, though, it's best if you read some of the background about the topic and start to think about what parts you think are important and feasible to complete in a summer. My guess is that we will already have most of the required models (e.g. sarimax and exponential smoothing) and so the project would be about constructing the automatic forecasting infrastructure.

Best,

Chad