Forecasting - Corona Virus Impact (UK)
General Impact
- Market volatility means that the predictability of sales across the industry is dramatically reduced and the accuracy of the forecast algorithm will suffer
- Additionally, periods of closures mean fewer data points for the algorithms to base forecasts on, which will further reduce accuracy
- This impact on accuracy will apply for an extended period during and after the coronavirus impact this year, as well as in the following year in the equivalent period before, during, and after the impact
- For new locations (with no sales last year), this will mean the system cannot forecast until we have the first weeks of sales. After an initial week of resumed sales, forecasts will be available
Approach – During the impacted period
-
A greater emphasis needs to be placed on managers to apply their local knowledge and observations as sales begin to resume. Adjusting the daily values to appropriate levels, using the system functionality to identify open and close times should yield acceptable accuracy and a realistic shape of the day
-
Recent sales (which includes data from the impacted period) will continue to be used as part of the forecast for weeks ahead during this time. The algorithm will automatically determine the optimum weighting between last year’s sales and recent sales as business returns to normal. Initially, it is likely that recent sales will be more representative (yield a higher accuracy) of current sales than last year’s data and will be given a higher weighting as a result. If and when sales return to previous levels, the optimum balance between recent sales and last year’s sales will change and the algorithm will apply greater weight to last year’s sales. Should sales continue to not follow the pattern from previous years, this will be automatically detected, and the weight given to last year’s sales will be reduced accordingly
Approach – After the impacted period
- The impacted periods and sales data from that time will be excluded from the algorithm going forward. This includes the entirety of the impacted period, both before and after lockdown(s)
-
The exclusion of this data will improve the accuracy overall (versus not excluding it), but the fact that there is less representative data means the accuracy of the algorithm, in general, will be worse than previously. The period of time during which the accuracy will be worse will extend up to 10 weeks before and after the impacted period next year, and to a lesser extent in subsequent years (diminishing every year)
-
Managers’ input will continue to be important in applying their local observations to the forecasts
Notable Days
- Notable days which do not fall within the impacted period will be unaffected and continue to work as before
- Notable days falling within the impacted period (such as Easter, and May Bank Holiday) will be problematic in the following year, due to our strong reliance on last year’s data to forecast these days
- Two different approaches are being evaluated:
- Tag notable days next year (2021) to the equivalent notable days last year (2019). This functionality has been utilised by some of our customers for Easter holidays previously, where the pattern of Easter was inconsistent from one year to the next
- We apply notable days as ‘Tags’ across the historical data set, which allows our algorithm to determine the historical impact of the notable day. For future forecasts, the day is forecasted as if it were a normal day, and the impact of the notable tag is overlaid
- We will benchmark both approaches and implement the version with the highest level of accuracy based on historical data
- The chosen approach will be managed centrally and communicated out to customers – customers will not need to take any action themselves
What do I need to do?
- Communicate operationally that the forecast accuracy will be reduced for an extended period of time so that managers can take this into account
- Emphasise the importance of manager forecasting and re-forecasting during the week – local observations will be critical during this time
- We will communicate any further developments on this as well as the outcome of our testing to find the optimum approach for 2021 notable days in due course
- For locations with no sales last year and a closed-down period of greater than 8 weeks, anticipate an initial week with no forecast sales in the system, as the algorithm will have no data on which to base this on. This will automatically resolve after a week of sales. This includes the shape of the day, so historical Rota Templates would be useful for the first week
Comments
0 comments
Please sign in to leave a comment.