0
Time-series :
Frequency = Monthly.
Seasonality = Yearly.
Trend = No-Trend.
Aggregated at country level.
Time-line : 2011 through 2021(most recent)
Forecasting metric = MAPE (Mean Absolute Percentage Error)

I am facing increase in MAPE with my ensemble(Auto-ARIMA + ETS + Prophet) model. The issue is due to recent COVID-19 shocks. The Ensemble model has an MAPE of around 5% when back-test until Jan-2020. i.e use data till Dec-2019 and forecast for Jan-2020. It basically is (m+1) forecasting approach, i.e use data till month m and forecast for next month i.e m+1. Back in mid-2020 the sales dropped drastically, which caused the increase in MAPE, and now recently the sales are higher than usual and the MAPE is high again. I am looking for ideas to handle the COVID shock on my time-series.

Things I have tried :

  1. Remove the COVID period(start of 2020 to mid-2021) completely, and back fill the values using the ensemble model.
  2. Selectively remove the points which has high error during my back-testing, and only fill those gaps using the model.
  3. Use COVID cases as a regressor, but this does not add much value.

The MAPE is as high as 20%, because the conditions have changed drastically. We are in the new environment in terms of sales and demand, and it feels like a cold start issue. What are some of the common techniques to address such use-cases.

Stephan Kolassa
  • 95,027
  • 13
  • 197
  • 357
Amandeep Jiddewar
  • 111
  • 5
  • 1
    An assumption that post-COVID data would look exactly like the pre-COVID one is a big one. Obviously, making a forecast from the lockdown time data would not be a good idea, but using only pre-COVID data can also be misleading since there were many social and economic changes in the meantime, it's also still unclear how the COVID thing would be developing in the future. – Tim Sep 23 '21 at 06:37
  • 1
    If your performance is measured by MAPE, you may be interested in [What are the shortcomings of the Mean Absolute Percentage Error (MAPE)?](https://stats.stackexchange.com/q/299712/1352) – Stephan Kolassa Sep 23 '21 at 07:05

0 Answers0