Forecasting for Corbion in Corona Times

In these unprecedented times, forecasting requires innovative approaches. Where a knee-jerk reaction to source more information is not necessarily the best approach, the application of more advanced analytics, machine learning and good pilots can offer answers. The need for taking the right decisions has never been more important, there is literally no time to waste.

Can predictive analysis help in discovering latency and what data is actually available? From business needs to data driven solutions, from data engineering to valuable insights in S-I-R epidemiological models to understand how infection in the population will change over time. Using the latest information per country to forecast scenarios for COVID development, the forecast model can be fed with this data to transform this information into insights for business. Short-term forecasts can utilize the order book, based on initial and re-estimated forecasts and, in doing so, propose enrichment.

Employing any available tool or technology to help you

For short-term forecasting, impact analysis and Bayesian enrichment can be deployed.
The “bathtub” curve can be utilized for mid-term planning and combined with the “rubber duck” curve to produce a reliable forecast in these corona times. It might be necessary to consider several developments in preparing future scenarios based on the consequences of COVID for market/industry, bearing in mind that the impact will vary per country. Machine learning techniques could be utilized to aggregate levels, both from pre-COVID and during-COVID periods. The combination of scenarios and driver-based forecasting could be useful for predicting long-term expectations.

It all starts with insights

Machine learning using additional data sources can predict patterns and improve your ability to forecast. This data could include booked shipments, macro-economic indicators, and collaborations. Employing feature generation to develop model features which contribute to forecast modelling. Identifying possible drivers of demand and which data sources may influence this demand will dictate data collection and the first models. These data sources could include historical sales, order book, stock levels, commodity prices, sell-through information, and macro-economic indicators.


In a final report-out session, results are reviewed and assessed using advanced forecasting models and additional datasets for generating a forecast. This can be supported by a broad suite of machine learning algorithms and visual machine learning for evaluation & training models. To be followed by a pilot.

Selecting a good pilot

What is the definition of a good pilot? A good pilot limits scope, it is big enough to be realistic and yet small enough to allow thorough evaluation. If the required data is not available, this can be time consuming to collate and everything hinges on this in the modelling process. Although this is the domain of data science, it is imperative that people from the business are involved to establish the correct objectives and provide market knowledge essential for modelling. The pilot should try to solve a relevant business problem, especially where significant unexplainable swings occur, and where there is a potential explanatory variable such as commodity price futures.


For the duration of this global crisis, no matter what business or market you are in, a selection from or all of the aforementioned approaches can assist in determining the intelligent course of action for your organisation. Delay is not an option, foresight lies within your grasp.