Forecasting after Covid

COVID-19 has impacted the demand for most companies. For some the situation has already been normalized, some have experienced a pent up of demand during the lockdowns, which is now catching-up. For others, demand remains significantly below 2019 levels. We all share the same experience: COVID made forecasting harder than it already was... But do not be afraid, there is something that we can do about it! Just hear us out and get that forecast back on track!

Internal data forecasting

Internal data forecasting is simply believing that historical demand is representative for the future. For the statisticians among us: extrapolating trend and seasonality of the timeseries. First remark: will this estimated trend be representative now COVID hit us? What can we do to make sure that we have good quality input for our forecasting models?

Well, we can do 3 things:

1. Look at the input
   Adjust the historical data so it is relatively in line with the ‘normal situation’. Different techniques can be used to smooth data patterns. You can use prior years to smooth datapoints, use an exponential smoothing model or prune the data (bring outlier values back into the ‘normal range’). Which triggers of course the question: What is the new normal?
   
   When cleaning patterns, be aware that seasonality can be missing!
   Capture the real demand history in a separate line and use it as a benchmark!
   
   
2.  Look at the model
   Validate and manage the forecasting methods that are used. This scenario is only valid in case you have some experience and knowledge about statistical forecast models.
   
   Let us list some questions that can be used when validating the forecasting methods:
   - Is it still valid to use 4 periods as testing or should we increase or decrease this holdout period?
   - Are the parameters of my forecasting method still correct?
   - Do we need to include or exclude some models from the list?
   - Should we change our average (3month) to an average of the last 6 months?
   - Do we want to forecast this combination/product/customer?
   - Garbage in = Garbage out, investigate the quality of the historical data. In case there is uncleaned historical data or a lot of uncertainty, investigate the use of short-term models.
   
   
3.  Look at the output
   Adjust the output of the forecasting engine. Keep the historical data as is, let the model run and validate, change, or adjust the output if necessary. Just keep 1 thing in mind; as every department has its own bias, make sure there is transparency and cross functional communication to evaluate and monitor the changes!

External data forecasting

More than ever, it is important to look beyond the company walls and to enrich the baseline forecasting process with external information. External information can originate from different sources: customer forecast, questionnaires, macro-economic indicators, …

Allow me to go a bit more in detail on macro-economic indicators. Find indicators that are correlated or leading with the market or demand pattern. Those leading indicators will capturing turning points that internal methods would simply have extrapolated. Let us explain the use of internal and external forecasts with a real time example.

Internal data forecasting will extrapolate the negative trend caused by COVID. When adding external data insights, turning point detection becomes possible. In our example, the external indicator predicts a recovery in 3 months which is not yet visible in the demand pattern of the customers.

Why should we do both (internal and external)? Good question! External data makes it possible to forecast higher level trends, to detect turning points, to get strategic insights in the market. However, we should still be able to translate that higher level forecast into an operational (lower level) forecast. By combining internal and external data, this becomes possible!