Repository files navigation

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 andAUTO.ARIMA (UsingR throughrpy2)
• ARIMA, ARIMAX, SARIMAX (Usingstatsmodels)
• Prophet (using Facebook's Python Library)
• tsfresh from Blue Yonder for automated feature extraction from time-series data.

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 anyoverall trend orseasonality
  • 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
  • Determineorder 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

• Notes on Regression and TS Analysis - Duke Univ
• Rules for identifying ARIMA models
• Sean Abu's SARIMAX tutorial
• Modern Pandas', handling TS + SARIMAX
• ML Mastery' ARIMA Tutorial
• Statsmodels Documentation
• Hyndman Book
• SO auto.arima equivalent in Python
• DO Tutorial - GridSearch for (p,d,q)