One of the things that the COVID-19 Pandemic has confirmed, is that our world is subject to change and our seemingly steady world trends can be drastically changed overnight. When working in the field of Predictive Analytics, we rely on these trends to make accurate predictions in which many business processes are dependent on. At Mediaan, we could not predict the global impact that COVID-19 would trigger. However, we have experienced the benefit of using data science algorithms that are quick to adapt and were able to take these new trends into account for forecasting.
Forecasting models for different business processes
Forecasting models appear in different business processes impacting and automating decision making. Any industry can benefit from a better estimate of what will happen next. At Mediaan, we have implemented various models over the years for our clients, these include:
- Call prediction for call center agent scheduling.
- Orders prediction for logistic planning.
- Stock depletion prediction for warehouse stocking.
- And many others…
Recently, we also started creating COVID-19 patients forecasting models for aiding hospitals to plan on-duty staff and reserve isolated rooms. We will discuss more on this topic in an upcoming new blog.
Adaptive data science models
Traditionally, decision making was done with rule-based systems or statistical interpolation. The problem with these methods is that they have to be implemented with as many “what if” scenarios in mind as possible. “What if” we release a new product? “What if” our biggest client switches their orders from Monday to Friday? Then, when an unplanned situation occurs, these methods either cannot make predictions any more, or their output starts to become very unstable, unusable, and unexplainable.
However, multiple tools within data science allow us to have a more robust system in place. They are quick to adapt to new situations. When unplanned situations occur, they either continue making accurate forecasts or at least have a stable output that can continue to be used.
Two examples of such tools are:
- An adaptive kernel (like the locally periodic kernel) for a Gaussian Process.
- An autoregressive neural network model.
The locally Periodic Kernel
The locally periodic kernel allows us to model trends that change over time. This is important even during normal situations since trends usually do change over long periods. However, now, with many countries and businesses undergoing quarantine, change is rapid. Learning is not immediate, instead, there usually is a delay in adapting because a robust system needs to take into consideration that a rapid change could be an anomaly, only when the anomaly is sustained for a longer period, it can be considered a new trend.
The image below shows an example of a Gaussian Process forecasting model. On normal conditions, this B2B ordering process had an oscillating 20-days cycle, however, around mid-April, the process volume dropped and switched to a 10-days cycle. After the change in orders trend, the forecasting model (represented by the orange line) took a few days to adjust its predictions, however, thanks to the Locally Period Kernel, it managed to adjust itself to the new volume magnitude and cycle period.
Autoregressive deep neural networks
Autoregressive deep neural networks are also an option for dealing with unforeseen trends. The basic idea behind these systems is to use a deep neural network model to learn long and short term business patterns and embeds these in its weight. Then, each day, feed the model with the predictions and the real observations of the previous days.
The better the forecasting model is, the closer the predictions and the real observations should be. However, when quarantine or another trend impacting external influence hits our environment, the predictions start to deviate from the real observations. By feeding both the model’s predictions and the real observations back into the model, it can adapt and compensate for the mistakes it made in the previous days for the new predictions of today.
Next year predictions based on quarantine data
There is still the question about what will happen next year when we try to create forecasting models that need to learn seasonality based on this year. In 2021, we will feed training data from 2020 into our forecasting models, but we are still unsure of what can be learned from 2020. Quarantine has completely changed our behavior patterns. So, will the models learn that March, April, and May are the best months for e-commerce sales even though they are far from Christmas? Will they learn that in these months people are less likely to drive even though they aren’t vacation months? Will models learn the Dutch people have turned their backs to King’s Day?
Developing forecasting models in 2021 will be harder than ever. We will need to pull data from earlier years when life was “normal”. Data from quarantine time will need to be smoothened so that our training data is continuous. Many processes have been completely halted, so even data simulation might be required to fill in gaps from this year.
Make models more robust to cope with changes
No solution solves every problem and drastic pattern changes can affect predictive models hard. However, data science offers a variety of tools to make models more robust to cope with changes. We are looking forward to helping our customers monitor the performance of their predictive models as quarantine measures get lifted and we try to go back as much as possible to normal.
This blog is written by Bruno Lubascher – Team Lead Data Science at Mediaan.