1
$\begingroup$

I'm pretty new to including time in any sort of modeling, so forgive me if any of my questions are basic.

I'm trying to predict a final sales number over a certain period of days using the incremental sales and how far out from the final date I am. Here's a sample of my data:

      EVENT.DATE       SALES.DATE SALES.TO.DATE DAYS.OUT FINAL.QUANTITY
1       4/2/2014        3/18/2014             0       15             42
2       4/2/2014        3/19/2014             2       14             42
3       4/2/2014        3/20/2014             4       13             42 
4       4/2/2014        3/21/2014             4       12             42
5       4/2/2014        3/22/2014             4       11             42
6       4/2/2014        3/23/2014             4       10             42
7       4/2/2014        3/24/2014             4        9             42
8       4/2/2014        3/25/2014             6        8             42
9       4/2/2014        3/26/2014             6        7             42
10      4/2/2014        3/27/2014             6        6             42
11      4/2/2014        3/28/2014             8        5             42
12      4/2/2014        3/29/2014             8        4             42
13      4/2/2014        3/30/2014            13        3             42
14      4/2/2014        3/31/2014            13        2             42
15      4/2/2014         4/1/2014            15        1             42
16      4/2/2014         4/2/2014            26        0             42
17      5/3/2014        4/18/2014            10       15            412
18      5/3/2014        4/19/2014            12       14            412
19      5/3/2014        4/20/2014            20       13            412
20      5/3/2014        4/21/2014            36       12            412
21      5/3/2014        4/22/2014            44       11            412
22      5/3/2014        4/23/2014            47       10            412
23      5/3/2014        4/24/2014            60        9            412
24      5/3/2014        4/25/2014            66        8            412
25      5/3/2014        4/26/2014            75        7            412
26      5/3/2014        4/27/2014            95        6            412
27      5/3/2014        4/28/2014           111        5            412
28      5/3/2014        4/29/2014           120        4            412
29      5/3/2014        4/30/2014           128        3            412
30      5/3/2014         5/1/2014           174        2            412
31      5/3/2014         5/2/2014           207        1            412
32      5/3/2014         5/3/2014           263        0            412

I'm assuming that I can't just use OLS because the 16 observations for each event date are not completely unrelated from one another. I've been looking into modeling that involves panel data, but I'm not sure if that is applicable. I've looked into time series as well, but I can't figure out if I can use that because my dates are often irregular. I'm willing to do the research myself, I'd really just like some jumping off points.

If anything is unclear, please let me know.

$\endgroup$
4
$\begingroup$

This is identical to a time series forecasting problem that Procter and Gamble asked us to solve where they had been unable to timely detect changes in expectations as new daily data was recorded. They gave us sales data by day for a three year period and wanted AUTOBOX to predict the month end number as time lapsed i.e. from different origins within the month. They wanted to be able to compute the probability of making/hitting/exceeding a month-end target after each day's sales had been received. After developing a useful daily model taking into account the usual suspects (memory/day-of-the-week/holiday effects, level shifts/local time trends/week-in-month-effects/particular days-of-the-month etc.) we convoluted the forecast uncertainty for future periods to be able to place probability limits on the forecasted sum for the remaining days in the month. If you have such data I will be more than happy to demonstrate this to you on your data as I believe that this problem is quite popular and our solution is very important as it illustrates sound/innovative statistical practice. For more on this see http://autobox.com/cms/index.php/blog/entry/will-we-make-the-month-qtargetq-number .

$\endgroup$
  • 1
    $\begingroup$ Perhaps they should pay you back for all that!... Getting the same problem solved by free is not a very good deal!... P&G trying to save some bucks is so bad!!... $\endgroup$ – Brethlosze May 30 '15 at 0:34

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.