Inaccurate demand forecasts? Why your COVID-era transactional data is important in 2021
Few people would envy the challenge facing demand forecasting teams right now. Gone are they days where you could rely on historical data to form rolling estimates to build forecasts on. Demand forecasting teams need new approaches and external data to get a standardized, manageable view of the chaos of 2021. But you shouldn’t just throw out your 2020 data - there is so much you can learn from it.
PredictHQ’s data is used by some of the smartest demand forecasting teams and models in the world across, so we want to surface some of the lessons they are implementing into their demand forecasting.
Preparing to make the most of economic recovery in 2021
Industries such as travel and hospitality were hit hard by the pandemic. Others, such as restaurant chains and retail fared better but not without enormous changes. Here are some ways businesses in these verticals can make the most of their eventual recovery period:
If your business was previously focused on long-term forecasting, consider updating your models to be optimized for more short-term forecasting. This will enable your company to navigate the ongoing volatile external environment in 2021.
Upgrade your models to include external data sources to identify factors drive the notable changes in your demand. By tracking catalysts such as health warnings, restrictions easing, severe weather, and other unscheduled events, you can have data-driven strategies ready to go when disruptive circumstances occur.
Keep in mind as attended events have begun to return, there will be a release of pent-up demand. Certain industries are going to be more susceptible to what some are calling “revenge spending.” As people have not been able to do things like travel and dine out, it’s likely that we’ll see a large uptick in spending in these sectors once normal day-to-day life resumes in more communities. We’re already seeing this in New Zealand, where events such as farmer’s markets and football games have become much busier than before COVID lockdowns.
Tracking all the factors impacting demand at each location will be challenging, as there will be many. Even if the events don’t seem directly related to your business, it’s critical that you keep a pulse check on the status of events in your area as an indicator of easing restrictions, rebounding confidence and footfall surges. Being aware of the rate at which events are picking up near your stores or locations will give you insight into where you need to start updating things like your staffing and supply chain plans.
What Data Scientists need to keep in mind when updating their 2021 forecasts
So 2021 is going to be chaotic, but data scientists still need to forecast demand as accurately as possible. The key fact for data scientists in 2021 is that short and long term 2020 trends are not because of your business. It’s much more likely that they’re attributable to external factors, such as vaccination rates impacting confidence or restrictions easing.
You will need to identify these factors and their impact to make accurate plans for their impact on your demand.
You need to combine both 2020 data with historical data to understand the discrepancies between demand.
You will need to incorporate additional intelligence through external data into your models to understand the context of your stores/sites. Teams should consider adding new data resources to make the most of it.
How do you make the most of your 2020 data:
Retrain your models: You cannot rely on pre-COVID forecasting models to be accurate, and using your 2019 forecasting models for 2021 will definitely lead to inaccuracies.
Hypothesis testing: Make educated guesses about what has changed for your business, and then simulate these changes in your models. If you compare and find a difference between your assumptions and the reality of your 2020 data, you can be fairly sure that factor influences your business, and test your hypothesis with your forecasting models.
Quantify feature importance: You can understand the gravity of individual features by adding them into your models, and comparing them to historical periods where your business was not affected by them. The more context you have to accurately account for the impact of different demand causal factors, the more accurate your forecasts will be. Add these new features, retrain your models, and your result will tell you about the feature importance.
Lesson 1: 2020 was the year of displaced demand
Throughout 2020, attended events were wiped out by shelter-in-place orders and government mandated shutdowns. This distorted demand patterns significantly:
Events such as festivals and expos that drive foot traffic were cancelled or postponed.
People traveled less during school and public holidays, disrupting long established surges for holiday destinations.
Remote schooling disrupted the standard patterns for many retailers such as the post-drop-off coffee run or the pre-afternoon pick up grocery shop.
College students didn’t return to campus at the same rate, upending long established surges of demand.
Capacity limits in stores led to bottlenecks and long lines.
Yet even though normal consumer patterns were disrupted, demand didn’t disappear. Predicting demand just became near impossible for those relying on historical data. Demand patterns differed from 2019 patterns, just displaced into new channels.
As we close out the first quarter of 2021, demand continues to shift channels based on rules and attitudes that vary state by state. Food and retail businesses that shifted almost entirely to online, order-ahead or delivery are now working out how to manage in-store purchases (and staffing and stocking) again. Travel companies are trying to identify which domestic hubs are likely to thrive in the new normal. Businesses need to adapt their forecasting strategies and understanding the balance between new and old channels will be key for economic recovery.
Identifying how your demand evolved in response to different demand impacts, such as shelter-in-place orders, retail restrictions and schools closures in 2020 will equip your company with data-driven insight into how the reversing of these rules will impact your operations in 2021.
Lesson 2: Future forecasting will require more than historical data only
Because 2020 disrupted demand so significantly, every company had to expedite their digital transformation strategy. For many, this included shifting towards continuous and dynamic ways of forecasting. Gone are the days of set-and-forget rolling forecasts driving decisions worth millions of dollars every month.
But continuous forecasting requires contextual data. Many large companies switched their focus to solvable issues - fixing their data deficit to enable the machine-learning solutions to demand forecasting and labor optimization. A recent article by McKinsey noted that “the COVID-19 crisis provides an example of just how relevant external data can be. In a few short months, consumer purchasing habits, activities, and digital behavior changed dramatically, making pre existing consumer research, forecasts, and predictive models obsolete. Moreover, as organizations scrambled to understand these changing patterns, they discovered little of use in their internal data. Meanwhile, a wealth of external data could—and still can—help organizations plan and respond at a granular level.”
The key to this is seeking out high quality external data to ensure your more dynamic and resilient forecasts are accurate. It is through this data that you can understand what drove your demand before COVID, during 2020 and how your customers are responding to similar events in 2021. New data will help you better understand the context of each of your stores or focus areas because it will identify what is driving your demand spikes or dips. These could be government lockdowns, severe weather, the return of fans to sports stadiums or parents doing school drop off.
Lesson 3: Take these insights and turn them into targeted features in machine learning models
Uncovering what drives your demand with external forecast-grade data is the first step. Once you know your catalysts, it’s time to prepare for incoming impacts.
To create more accurate demand forecasts, data science teams will need to build out more robust features into your models that take new data sources into account. While it’s unlikely that we’ll see another global pandemic at the same scale as COVID-19 for quite some time, making your models smarter and more resilient is the one opportunity every business can grasp moving forward.
By leveraging your COVID-impacted data from 2020, you’ll be able to more accurately forecast during other disruptive events. You can accomplish this by decomposing your data from this period. The main things to understand are:
Downward shifts: your model needs to be able to recognize the drop in demand as both short term and long term trends, separate from things like seasonal trends.
Recovery rates: this is where external data will come in handy. Understanding when and where economies are reopening is key. The rate at which they’re heading back towards “normal” operations will be informed by things like government mandates and events returning.
Resuming events: as demand-driving events like concerts and festivals get rescheduled for 2021, demand will increase sharply. Knowing the magnitude of how these events will impact demand for your business will be crucial in the coming year.
Large attended events such as conferences, sports games and students returning to campus have already begun to be scheduled. But the recovery will vary state by state, and it’s likely to be a chaotic and fragmented period for at least 12 to 18 months. PredictHQ tracks thousands of events across 19 categories, calculates a predicted attendance, and checks and updates these frequently every day to ensure they are accurate. Get in touch with our team to get access to our reliable demand impact data, which can be easily inputted into your existing models and demand forecasting models.