GSoC_2018_time_series_prediciton

Time series prediction

Time series problems could be found in various areas of our life:

• health care: ICU vital signs
• financial market
• sensor data

All of them requires a very different way of processing the input data and to train ML models, than in case of 'traditional' data set, even cross-validation is specific for this domain, see a quick intro into this domain.

Mentors

• Viktor (github: vigsterkr, IRC: wiking)

Difficulty & Requirements

• C++ as Shogun needs to be extended quite a lot
• time series indicators
• independent

Description

The application of the project could be from any of your favourite domain, e.g.:

• healthcare: ICU data eICU, MIMIC
• finance: CryptoCurrency History, BETA-IV, S&P 500 Stock Data

The main contribution of this project would be to develop methods within Shogun to support time series data processing and modelling. As there are already very good libraries that supports some of the required methods for time series analysis (e.g. Stan) the initial task would be to design how to integrate these tools into Shogun as implement an MVP that exposes the required method into Shogun itself.

Useful resources

• Stan
  • https://github.com/nwfsc-timeseries/atsar
  • http://www.unofficialgoogledatascience.com/2017/07/fitting-bayesian-structural-time-series.html
  • http://tharte.github.io/mbt/mbt.html
  • https://github.com/michaelweylandt/R-Finance-2016-Tutorial-Materials/raw/master/RFin2016.pdf
  • http://mc-stan.org/users/documentation/tutorials