The Datapred blog

On machine learning, time series and how to use them.

The problem with deep learning and time series

Posted by Datapred | Aug 1, 2018 8:16:05 PM

Industry is finally catching up on the potential of machine learning for predictive analysis. But data scientists faced with a prediction challenge (demand prediction, predictive maintenance, churn prediction…) usually favor one of two equally inadequate approaches:

  • Conventional modeling: « If I train, validate and test my Random Forest like I did in school, I will be OK ».
  • Misguided sophistication: « LSTM networks are natively sequential! Let’s kick ass with the latest advances in deep learning ».

We have explained elsewhere why conventional modeling doesn’t work with time series. The goal of this article is to explain why the second approach doesn’t work either - why deep learning doesn't work for most industrial prediction challenges.

There are three main reasons for that.

  1. Deep learning models are not sequential. Their internal parameters don't update sequentially.
  2. Training deep learning models requires much more data than most industrial processes generate.
  3. Some neural networks (such as LSTM networks) can theoretically handle time series, but their predictions are hard to explain. And in industry, interpretability is often as important as accuracy.

As a consequence, deep learning models applied to industrial predictive challenges often underperform... simple lagged values! We have seen that over and over again - in contexts as diverse as consumer goods, e-commerce and the automobile industry.

So what should you do?

In our experience, three modeling tactics contribute to prediction accuracy:

  1. Sequential learning. Incrementally training and testing your solution ("backtesting") will usually give you immediate 10-15% accuracy gains.
  2. Online aggregation. Instead of using a single super-sophisticated model, or switching from one model to another based on data configuration, use a diversified and adaptive combination of predictive models. This methodology not only improves accuracy but also enhances interpretation.
  3. Industry expertise. Use the human expertise around you for feature engineering, model selection, cost function definition... It will pay a hundred times more that endlessly tweeking your deep learning model.

Datapred automates most tricky aspects of machine learning for time series. Contact us to discuss how we could help with your predictive analysis challenges.

You can also check this page for a list of resources on machine learning for time series.

Topics: time series, machine learning, sequential data, backtesting, neural networks, deep learning, LSTM

Written by Datapred