COVID-19 is a 'black swan' event. The pandemic is world changing; old models and old datasets aren’t going to directly apply for quite a while. You will need to rely on human expert analysis of all the complicated legal and emotional factors that are influencing the different markets.
If you’re literally manufacturing toilet paper, the forecast is easy. Make and ship as much as you can fit on the trucks. What about planning sales for luxury items? We’re already seeing demand close to zero and will for a couple months or more. These things are visible in the data already and very clear to a human, but a machine learning model doesn’t know “this situation is still just getting started.”
Anything that relies on historical data in a traditional capacity is useless in these circumstances. So how do we plan, now? We’ve polled our experts, collected ideas, and assembled the following recommendations on how to shift your thinking and forecasting/planning process for these uncertain times.
Strategy #1: Shift to Expert Assumptions
Most forecasts are largely based on historical data, whether or not they leverage advanced machine learning algorithms. Analyzing historical trends and patterns is one of the most valuable techniques available to predict the future, but no model would have been able to foresee COVID. Even over a month into the crisis, classic seasonality/trend models will not be able to react quickly enough to anticipate the short-term COVID effect.
Subject matter experts, however, know the business and can stay aware of changing conditions in the market that may not be apparent from available data. Using heuristics and their knowledge of current events, a person can much more quickly adapt to daily changes in the global situation. In normal circumstances, we tend to avoid this because it can be a source of bias, but we're in desperate times.
To enable the business experts to take control over the forecast, we recommend the creation of a separate, user-controlled forecast to be managed along with the system-generated forecast. The historical-driven forecast may still be valuable as a baseline, but it is important to give your experts the control they need to manage the situation. Use targeted dashboards and reporting to quickly recognize and react to the different impacts COVID is having; identifying where the automated models may be struggling the most. This can pinpoint the areas that require the most attention.
Strategy #2: Speed Up
Having agile systems and processes to make more real-time forecasts is now a requirement, not a nice to have. Forget annual budgeting, even quarterly forecasts are dead. We still don’t know the full scope/impact of COVID, and the situation develops every single day.
Whether continuing to use a predictive model or relying on the expertise of SMEs, we need to react as quickly as possible and reforecast as soon as something changes. If a governor announces a statewide shutdown or reopening, we need to immediately reforecast as this can have a direct impact on supply/operations.
Strategy #3: Look at Other Disasters
COVID is a unique crisis, but not the only one we’ve ever experienced. It is possible that there are other events or recessions we have recorded that caused similar impacts on business. Through machine learning and analytics, we may be able to leverage specific points in history to predict/plan for the current situation. This can help inform a user-defined forecast or be fed into a machine learning model.
In these scenarios, reporting/analysis is extremely valuable to see how different segments of the business were impacted. For example, maybe in certain geographic areas we’ve seen similar store closures during extreme weather events. Maybe the 2008 recession mirrors certain economic events occurring now that can be leveraged.
Strategy #4: Bootstrapping & Scenario Analysis
Although limited, we do have some data available around the business’ reaction to the current COVID climate. This may be the most valuable data we have when trying to predict what’s going to happen in the short-term future – and the short-term future is more important now than ever before. Demand may shoot up following an economic-reboot or drop off a cliff if the situation worsens.
Using this recent subset of history, we may be able to either:
- Build a new short-term model based on current/recent trends
- Generate new inputs/predictors to be fed into existing models
- How far from ‘normal’ are we? How far off are our current models?
- This can allow for scenario analysis, tweaking predictors/model inputs and seeing the models’ response
- More granular models allow for more detailed analysis
- Keeping models as granular as possible is more important than ever before, particularly when tracking things like cash flow and sales.
- Add more flexibility through scenario modeling
- It’s going to be extremely important to plan for different scenarios, since the future is so up in the air.
- You may want to plan for a best/worst case scenario based on when the economy might restart or have an entirely new environment spun up to manage sweeping data changes.
Keeping the models granular is going to be extremely important here, since COVID is being handled so differently in different geographic areas.
Strategy #5: Prepare the Data for a Return to Machine Learning
Many retailers are seeing massive upswings in demand for specific products, but it's hard to tell whether these trends represent a 'new normal' level of demand vs. widespread panic purchases that are unlikely to repeat themselves (e.g., the toilet paper craze.) These lockdowns will get lifted at some point, which means demand curves will
- surge for some categories as people are excited to travel and live their normal lives or
- continue to be constrained by the recession
Accurate predictions are unlikely given how much purchase behavior is continuing to change throughout this lockout period, but we can help make better decisions by offering confidence intervals / high probability density intervals where we'd normally feel confident sending exact answers.
A few options of handling the models are:
- Ignore the COVID period, assume a return to normal trends
- Consider the COVID period when calculating trend, but assume a return to normal seasonality
- Integrate new predictors/inputs to which the models can attribute the extreme growth/decline
- Building new models or making major updates to existing models may be warranted.
- When planning expenses/payroll, the workforce may be vastly increased/decreased depending on the industry.
- Furlough modeling may be necessary, along with a plan to re-hire a substantial amount of the workforce.
Undoubtedly, the current situation is a challenge to any forecast, plan, or predictive model. By incorporating new approaches and modifications to existing forecasting models, you can evolve your forecasting and planning process to stay relevant. Be flexible, add detail, and speed up.
Missed the webinar? Watch it now!