ARMA is obsolete - but whats next?

This is one of our recent post from LinkedIn. Follow us on LinkedIn to stay updated with regular insights and news.

It's a classic, around since the 70s, used it myself back in uni, but data science needs have clearly evolved.

With TSFresh, Blue Yonder's engineering team set the sails for brute-force feature engineering, combining a large set of 100+ predefined feature templates with traditional feature selection methods. This approach performs well, but their implementations limitations: computational complexity on real-world data and limited support for multivariate time series.

image

That is why we built getML’s FastProp to tackle these challenges head-on.

Our recipe? Optimized memory management in our getML C++ engine together with supervised learning.

In one of our Kaggle notebooks, we benchmarked getML alongside Featuretools and TSFresh on a textbook time-series problem: predict PM 2.5 concentration, a key air pollutant in the air of Beijing. The result? getML outperformed TSFresh by over 10 percentage points in R-squared.

We have two leading hypotheses for this:

  • Complex Feature Learning: getML uses feature learning to produce richer, more predictive features.
  • Efficient Memory Management: Our memory optimized data structures enable a deeper lookback, capturing more valuable historical data patterns.

The runtime improvement was even more dramatic. getML was approximately 400x faster than Featuretools and 20x faster than TSFresh in a head-to-head comparison.

Are ARMA-based models still in your toolbox? Have you worked with Featuretools or TSFresh before?

Curious to Learn More? Dive into our Jupyter notebooks for detailed benchmarks — link below!

Find the full Jupyter notebooks in our example section: