No new normal: Demand forecasting after COVID-19

COVID-19 is not just a shock to the system for the world’s supply chains. Markets have changed fundamentally and permanently. According to a study by Supply Chain Insights, referenced in this excellent video, demand volatility was one of the greatest perceived risks among supply chain leaders even before COVID. 

But permanent change is not the biggest challenge. As Lora Cecere explains so well in this article, we should not be looking for a new normal – because it’s not there anymore.It’s only a matter of time before the next big shock and system change happens. It could be related to climate change, political upheaval, you name it – but it’s going to happen. 

“We should not be looking for a new normal - because it’s not there anymore.”

Therefore we need to adapt our forecasting so that it can handle not just a different world, but a constantly changing one.

How to make forecasts more responsive

Let’s imagine a constantly changing market. To keep it as simple, this market oscillates between two states: 1 and 2.

Consider the figure. 

In market state 1, a particular product has the relationship to air temperature shown by the blue observations. You can imagine the product being icecream if you like (who doesn’t like ice cream, right?)

The blue curve is our model to explain this relationship. As long as this state of the market persists, the blue model is helpful for demand forecasting and planning. 

Now imagine that the market changes abruptly to state 2, as represented by the green observations. In this new state, the green model would be helpful. 

But what happens if we naively update our models during this transition period? Well, we get the red curve. This model is trained on data from the old and the new world, so we get some kind of compromise. This model, while arguably better than the old one, predicts demand that is never going to happen in the real world. 

If the market oscillates between these two states we may end up stuck in the middle,  training our model on a mixture of data from two different market states. In reality, the market is going to treat us even worse – not just switching between two fixed states but ever onwards to new states. 

In effect, with this naive approach we will always be under or over predicting demand. That’s going to cost us a lot of cash in inventory or lost sales and permanently lost customer satisfaction.

What can we do to remedy this awful situation? That’s a big subject, but here are three steps we can take to get started. 

Step 1: Monitor forecast performance

First and foremost we need constant, low-latency monitoring of forecast quality. 

If our forecasts aren’t keeping up with the market, forecasting error will increase. Whether you’re using MAPE or RMS or other accuracy measures, keep track of these and set up alerts if they exceed appropriate thresholds.

Consider smoothing out errors, as shown by the blue curve here, to reduce false positives. But don’t use a lagging smoother like moving average for this – low latency is key, remember?

Also make sure you put in place effective escalation processes so that the time from an alert to corrective action is minimized. 

Step 2: Improve forecast responsiveness

This can be done in many ways. A good first step is to consider the length of the training data period as a hyperparameter (see box) to be optimized for our forecasting model. For example, do we train on the last week, month, or year of data?

As the speed of market change accelerates, it is likely that data that is several months old does not contain useful information anymore. Rather, it can mislead our model. We should optimize all hyperparameters of course, but this is a crucial one.

Further, even after COVID-19 market effects – hopefully – subside, our world will keep getting ever more complex and dynamic. And our supply chains – or value networks – will naturally follow along.

The way to succeed long term, therefore, is to change your processes and technology so that they become agile, proactive, and demand driven. Your system – people, processes and technology – must be constantly looking for, adapting to, and exploiting change. 

Step 3: Sense demand

Demand sensing seems to be the new black in forecasting. But it is not (just) a buzzword. It is a common-sense approach to the new reality of rapid and accelerating change and complexity. The goal is to capture information as quickly as possible to guide and correct our forecasts. 

Demand sensing means we think in terms of hours or days, not weeks or months. 

We must incorporate all the internal and, especially, external signals that contain information about demand. For example, social media can be processed with new, powerful natural-language machine learning. For example, you can apply deep learning to build social media demand sensors that are specifically adapted to our brand(s) and products. 

Weather forecasts are available at low cost – if you are willing to put in the work to process them. Otherwise there are plenty of tools and companies to do that for you. The weather has a surprisingly large effect on demand for many products.  

There are many other signals you can exploit, including macroeconomic ones like local unemployment rates. The first step here is to start thinking creatively and more importantly, just making the decision to start sensing demand.

Seize the opportunity

This new, ever changing world is a threat to the status quo. But it is also a huge opportunity, if you act faster than the competition. 

If you’d like my help with improving your demand forecasting, get in touch with me. I would love to help you.

Leave a Reply

Your email address will not be published.