/ts-forecasting-ensemble

CentOS based Docker container for Time Series Analysis and Modeling.

Primary LanguageJupyter Notebook

Build Time Series Forecasting Ensembles

This CentOS-based container running on Python3 has the tools necessary to build an ensemble of time-series forecacasting models. Inside the /home folder, there are sample data and notebooks with examples on building the following models

  • ETS, TBATS and AUTO.ARIMA (Using R through rpy2)
  • ARIMA, ARIMAX, SARIMAX (Using statsmodels)
  • Prophet (using Facebook's Python Library)
  • tsfresh from Blue Yonder for automated feature extraction from time-series data.

Forecasting tl;dr

These are the steps

  • Explore

    • Plot the data
    • Clean outliers, Impute missing values if needed

  • Transform

    • Take the natural log if needed

  • Decompose

    • Check if the time-series has any overall trend or seasonality
    • Plot the decomposed series

  • Check for Stationarity and find $d$

    • Is the series stationary?
    • Run the Augmented Dickey Fuller test,
    • Check ACF & PACF plots to
    • Determine order of differencing needed to stationarize the series

  • Check for Autocorrelations and find $p, q$

    • Examine ACF and PACF plots

  • Fit ARIMA/SARIMAX model over a grid

    • Use (p, d, q) and set up a grid search
    • Find the best model using
      • AIC/BIC
      • Out of Sample Prediction Error
    • Check your Residuals, they should be ~$N(0, 1)$ and look like white noise

  • Make predictions

PS: that ARIMA models assume non-seasonal series, so you'll need to de-seasonalize the series before modeling

Recommended Reading