I am building a simulation. I want to generate arrivals according to a negative binomial process. The data will show the minute of each "customer" arrival and look something like the following:

Arrival      Time
1            1/1/2015 0:12
2            1/1/2015 0:56
3            1/1/2015 1:27
...          ...
83465        1/10/2015 22:12

I have historical count data that I am modeling in hopes of estimate the parameters of the negative binomial distribution. It looks like this:

Hour                    Arrivals
10/1/2014 0:00          2
10/1/2014 1:00          3
10/1/2014 2:00          1
...                     ...
10/20/2014 22:00        1

I know that if I was using Poisson regression, I would have an estimate for a parameter $λ$ returned by the regression. This parameter would represent the expected number of arrivals during a given period (in my case an hour).

I could then simulate this process by from time $t=0$ by sampling from an exponential distribution with mean $μ=1/λ$. This would return the an interarrival time $t1$. I could then schedule an arrival to occur at $t0+t1$. I could repeat this process for the next arrival which would occur at $t0+t1 + t2$ and so on...

What is the analogous process for a negative binomial process? NB regression returns an estimate for the mean but also the dispersion $θ$.

  • $\begingroup$ It is unclear whether you are interested in simulation (all model parameters are then fixed and synthetic datasets are produced by a random generator) or in estimation (all model parameters are then unknown and the dataset used to estimate those parameters). $\endgroup$ – Xi'an Jan 6 '15 at 20:45
  • $\begingroup$ Simulation. All parameters are known, and I want to create a synthetic dataset of timestamps that represent "customer" arrivals. $\endgroup$ – DG1 Jan 6 '15 at 20:50
  • $\begingroup$ Specifically, the mean and dispersion will be known based on negative binomial regression of count data. This data was collected in the past. I am running a simulation of the future. $\endgroup$ – DG1 Jan 6 '15 at 20:51

This is answered in the accepted answer of this question. In short, the waiting time between events in a compound Poisson-Gamma distribution is a form of generalized Pareto distribution.

Another resource is Properties of the Negative Binomial Lévy Process (Kozubowski & Podgórski 2009) who go into more theoretical detail on generalized negative binomial processes, and give two methods for generating paths in section 3.2, although I must admit, I do not completely understand it.


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.