How can your SKU time series improve forecasting accuracy with meta-learning?

SKU time series improve forecasting accuracyStatistical forecasting is known to offer low-bias historical inferencing for a diverse range of time series, resulting in good forecasting accuracy across multiple industries. For all the many distinct demand pattern variations that exist, there are unique statistical models that aim to describe these different patterns. How to decide which forecasting model performs optimal for which demand pattern? This question is already relevant if the demand patterns in question are the top 3 SKUs of your company, but how to deal with this question if you need to forecast tens of thousands of SKUs every month?

Meta-learning will better capture your demand pattern

A straightforward and robust way of matching a statistical forecasting method to a series of historical demand values is to simply try them all and select the statistical forecasting model which achieved the best forecast accuracy during a rolling forecast. However, the main downside of this approach is that this requires a significant number of computations. In addition, the information used to select the statistical model is confined to the data of that single demand pattern.

What if we could somehow store the decision for one statistical model over all the others for every demand pattern? Meta-learning offers a way to do just that. Meta-learning in demand forecasting is a method that uses descriptive statistics of a demand pattern to match the way the pattern ‘looks like’, to a statistical model that offers a high accuracy and low bias forecast. Which comes down to a multi-class classification problem within the field of machine learning.

Assigning statistical forecasting model to a time series

SKU time series improve forecasting accuracy through meta-learningWe can utilize a two-step machine learning classification approach of assigning a statistical forecasting model to a time series. By first encoding a time series into context-independent descriptive statistics, we can train a decision tree classifier to assign a statistical forecasting model. In the second step, a multitask neural network handles the conditional question of which parameters the assigned model should use. Once a training set of observations has been labeled, both classifiers can be trained. Assigning a statistical forecasting model to a new time series comes down to calculating its descriptive statistics and allowing the classifiers to recognize the best-fitting model plus parameters.

One of the benefits of meta-learning is the reduced number of computations to be performed in exhaustively searching the performance during a rolling forecast test run each period. In addition, an increase in forecast accuracy is anticipated on very variable demand patterns, such as strongly intermittent demand. These two benefits are especially relevant today, where the ability to accumulate data and have access to an increasing number of demand patterns, increases the number of forecasts to be generated.

Staying ahead in the demand forecasting game

To stay ahead in the practice of demand forecasting, EyeOn continuously explores new techniques to further increase the value we can offer our customers. Recently our intern Robert Duijfjes explored the possibilities and added value that meta-learning can offer in demand forecasting across industries. He concluded that the added value is mostly visible in extremely volatile demand patterns, which is exactly the part of the portfolio where customers often encounter challenges. Applying meta-learning to these cases results in more focus and therefore a major time-save for the customer because the demand planners only need to perform forecast enrichments on a smaller part of the portfolio.

Next to meta-learning, we are exploring other options to improve your forecast performance. With these innovative methods we will be able to achieve improved forecast accuracy and bias, whilst handling larger portfolios.

To make sure our clients can benefit most from these new innovations, we’ve developed the Fast Forecast Scan: a quick tool that provides you with rapid insights into the demand characteristics and forecastability of your business. As a first step, we perform a thorough deep-dive into your data and provide actionable data quality insights. With improved data quality, the Fast Forecast Scan provides you, within a few days, with data-backed insights on the highest possible forecast accuracy that can be reached and identifies the main opportunities for improvement. Learn how the Fast Scan can help you reveal your forecast optimization potential here.

Curious to hear more about meta-learning or other forecasting innovations? Reach out to our experts Rijk van der Meulen or Edward Versteijnen!

Search for