Edit I've rephrased this question severely.

Suppose I have a fitness center, and instead of a monthly fee, the people are paying for courses they are taking. I've got the monthly total revenue of all participants who started in January, for seven months (up to July). I've also collected the monthly revenue generated by participants who started in February and from those who started in March, also up to July. These cohorts don't need to have the same size. The data looks like this:

        Jan     Feb     Mar     Apr     May     Jun     Jul 
Coh1    6970    9890    8610    7200    6270    4680    4860
Coh2    NA      5970    7850    6720    5720    5020    3680
Coh3    NA      NA      4350    5420    4570    4010    3460

Note that the data above is not tab-separated - the head and the NA have excessive tabs for a nicer display here.

I'd like to predict the revenue generated by the second cohort in August, and for the third cohort in August and September, assuming that their training habits over the time are the same as these of the first cohort.

Plotting the first row ($x_{1, {\rm Jan}}, \dots, x_{1, {\rm Jun}}$) over the second (shifted by 1 to the left, $x_{2, {\rm Feb}}, \dots, x_{2, {\rm Jul}}$), and the third (shifted by 2 to the left) looks pretty linear, so linear regression seems to be suitable here.

Employing this shift, it seems reasonable to model $x_{2, {\rm Feb}}, \dots, x_{2, {\rm Jul}}$ as observations , and $x_{1, {\rm Jan}}, \dots, x_{1, {\rm Jun}}$ as independent variables to predict the August value for Coh2, $\hat x_{2, {\rm Aug}}= \beta_0 + \beta_1 x_{1, {\rm Jul}}$.

However, how does one proceed with the next cohort? Intuitively, it seems like a better idea to predict the August values for Coh3 using the shifted values of Coh2 ($x_{2, {\rm Jan}}, \dots, x_{2, {\rm Jul}}$) than to go back to the less recent observations of Coh1 ($x_{1, {\rm Jan}}, \dots, x_{1, {\rm Jun}}$). While I am fine with estimating $\hat x_{3,{\rm Aug}}$ by plugging $x_{2, {\rm Jul}}$ into its corresponding equation, I am really reluctant to use $\hat x_{2,{\rm Aug}}$ to get $\hat x_{3,{\rm Sep}}$ - since I'd use a prediction (which has errors) to predict something else. Doesn't this create even more errors?

TL; DR: What happens when I use a prediction instead of an actually measured value to get a prediction?


Are you allowed to? Sure. An example would be simulations using past predictions to predict the next iteration. Just be aware of your assumption and know that errors compound. So if there's some initial error in your prediction, the next predicted value will occur "on top" of that error.

  • $\begingroup$ Is there some way to quantify the error propagation? Is there some way which allows me to compare the results of $\hat x_{3,{\rm Sep}}$, using either $x_{2, {\rm Jul}}$ or $\hat x_{2,{\rm Aug}}$? $\endgroup$ – Roland Jan 31 '14 at 19:28
  • $\begingroup$ Measuring it directly with a simple test/train split where your model predicts data it hasn't seen before is probably the best idea. Then you could test the strategy of including the predicted variable against excluding it and compare predicted values against actual. $\endgroup$ – eric chiang Jan 31 '14 at 19:45

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.