I'm not sure how to look up and read about this problem, and I'm hoping you might be able to point me in the right direction. In essence, this question is about whether or not I should purchase another unit this year:

Given that I have sold 3 of item x this year, 
what is the probability that I will sell a 4th?

I think this is related to a manufacturing question of this variety:

Given that the machine has produced 3 units without failure
what is the probability that it will produce a 4th unit without failure?

In a less simplified version of this question, you might take into account that sales is a time series (that unit 1 was sold at time 0, unit 2 was sold at time 2 and unit 3 was sold at time 2.5), and attempt to predict the probability that unit 4 will be sold within time 5 - but I'm not sure I have the statistical foundation to understand how to do that quite yet.

What is this type of problem called, and what should I focus on learning to be able to answer it?

  • $\begingroup$ Your manufacturing example is from reliability study area. Your sales is not very clear. You seem to be trying to apply manufacturing models to it. I doubt this would work. $\endgroup$
    – Aksakal
    Dec 10, 2014 at 23:11
  • $\begingroup$ I'll look into Reliability more -- but I'm not sure why you think it won't work. Imagine it's June, and we have sold 1000 rubber bands. Should I order another rubber band? Probably - I'm just wondering how I might calculate how sure I will be. $\endgroup$
    – Robert
    Dec 10, 2014 at 23:16
  • $\begingroup$ Try time series, such as AR or ARIMA. The reliability deals well rare events, hopefully, while sales deals with huge volumes, again hopefully. $\endgroup$
    – Aksakal
    Dec 10, 2014 at 23:17
  • 2
    $\begingroup$ To answer such a question about sales would require some strong assumptions; to apply an argument from reliability, the assumptions you'd need are unlikely to be reasonable for sales. $\endgroup$
    – Glen_b
    Dec 10, 2014 at 23:29
  • 1
    $\begingroup$ It would take some Procrustean effort to view these data as a time series. It is certainly a realization of a stochastic (point) process; recognizing that will open up a large number of approaches for you to consider. A simple one, for instance, would model the distribution of durations between sales. You could incorporate factors related to the presence of weekends, holidays, seasons, and so on, if that were necessary. To succeed with this (or any other approach) you will need data that are much more extensive than the three observations mentioned in the question! $\endgroup$
    – whuber
    Dec 11, 2014 at 15:22

2 Answers 2


I'm sure there are more in-depth answers to be given for this general problem, but for the specific problem you mentioned, this sounds like it would be straightforwardly modeled with a Poisson distribution.

A Poisson distribution models independently occuring events in a fixed time period. In your example, you had one event in the $[0,1)$ interval, no events in the $[1,2)$ interval, and two events in the $[2,3)$ interval, you can represent the events like this: $k = \{1, 0, 2\}$. The MLE for $\lambda$, the Poisson parameter is $\frac{1}{n} \sum k_i = 1$.

To find the probability that at least one event occurs within the interval $[3,4)$ is equivalent to finding $1-Pr(X=0)$ (one minus the probability that no events occur). So we can use the Poisson function's probability distribution function: $Pr(X=i) = \frac{e^{-\lambda} \lambda^i}{i!}$ to calculate this.

$Pr(X=0) \approx 0.37$. So the probability that at least one event occurs in the interval $[3,4)$ is about $0.63$.


I think that sales forecasting is too big of a topic. Without being specific, you will not get a lot of useful help here.

Having said that I'd suggest starting with autoregressive time series models, such as AR(P) or ARIMA(P,1,Q). Sales are usually persistent, i.e. next month sales are similar to this month. They are also often highly seasonal, so you may need to add seasonality to your model. In time series it's often done through multiplicative seasonal time series.

If the forecast is short, less than a year, and the firm is not in explosive growth mode, then ARMA(p,q) maybe enough.

Sales may depend on external variables, such as commodity prices or marketing spending, you can add them in ARMAX or ARIMAX framework. Be careful with potential endogeneity issues with variables like marketing spend though.


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.