1
$\begingroup$

I am making monthly forecast of sales volume, and have IT system in place to observe daily sales. I want to compare actual sales with forecasted sales every 10 days (twice in a month - on 10th and 20th day of every month), and then adjust previously forecasted value for the month if required. Idea is that actual sales volume for first 10 days should help in improving/adjusting forecast for remaining 20 days. A Naive approach will be to look at the ratio of actual vs forecasted sales for first 10 days and then use this ratio to adjust forecast for remaining 20 days. Is there any statistical methodology to do it in better way? Any pointer will be helpful? Please note that actual sales is not uniform across 10 days bucket, I observe that sales generally pick up in last 10 days of the month.

Example : Suppose forecast for December 2017 is 300 units (that will be 10 units per day assuming uniform distribution of demand). On December 10th, total sales till date comes to be 120 units, question is that how can we revise forecast for remaining 20 days?

Note : I have historical sales data available for last 3 years, and monthly forecast is time series forecast using this historical data. In general sales follow seasonal trend.

$\endgroup$
  • $\begingroup$ Why not use all available information at all times to obtain the best forecast? $\endgroup$ – AdamO Nov 7 '17 at 13:58
1
$\begingroup$

Your proposed approach is ad hoc, but it makes sense.

One question is whether higher sales during the first 10 days of a month actually predict higher sales during the rest of the month ("short term trend"). After all, it might well be possible that customers who were going to buy during the second half of the month decided to pull their purchase forward in time - and in this situation, higher sales at the beginning of the month might actually predict lower sales during the rest of the month.

Many customers of ours want forecasts to be "highly reactive" in the way you are describing. It usually leads to higher volatility and lower accuracy. I strongly recommend that whatever you do, you examine carefully whether your "high adaptivity" makes matters better or actually worse.

The alternative would be to directly model and forecast on daily granularity. If you have (intra-yearly and intra-monthly, with higher sales during the last third of a month), you can use appropriate models like or .

$\endgroup$
  • $\begingroup$ Thanks Stephan! Customers preponing their purchase from second half to first half of the month is not much concern for now (it should be taken care when we aggregate at month level). Problem comes when there is level shift because of some recent development. For ex, if a competitor reduces its price then our sales is likely to go down, and these recent developments are not included in time series forecast. Assumption is that last 10 days of sales should give us signal if there is level shift. I am starting with ad hoc methodology, comparing previous forecast with revised (as you suggest). $\endgroup$ – user3697157 Nov 8 '17 at 11:32

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.