WWW Improve Your Business

DFSS: Monte Carlo Simulation in Business Process Design

When designing or re-designing a business process it is often difficult to model changes based on deterministic factors (using simple math, or empirical formulas). Business processes are dynamic and so stochastic methods work better to model them. Stochastic models help to understand the probabilities of things happening (or not) and thus can quantify risks better in design.

Martha Gardner of GE Global Research recently spoke about this topic. You can find her paper here. Ms. Gardner talks about DFSS in an engineering environment, and her examples include problems like the stiffness of helical springs. However, these methods can be sucessfully implemented in solving business process problems.

Translating the initial business problem statement, usually based on the 'voice of the customer', into the final problem statement in very clear terms will require understanding the relationship between the process input and process outputs. This will help us define the process "Y's" and "X's". The process "Y's" are the metrics or indicators that we will want to optimize or improve. These "Y's" will be related to process "X's" - the process measures which you would be able to capture data around.

For example - if you are trying to improve the service level at your call center, you may have identified 'Abandon Rate" - the rate at which callers hang up - as one of your process Y's. In this case, by reducing Abandon Rate you would be improving service level. Abandon rate would depend on numbers of calls at a time, the length of the hold time, and the number of call center agents. So these would be your process "X's".

You could capture data on all of these "x's" - number and time of calls; hold length of calls; and of course you know how many call center agents are taking calls. Empirically, you would be able to use all of this data to see very easily when you have your highest abandon rates, and thus take measures to adress these (adding additional agents, adding call options, etc.) However, your solution is limited to the data that you have. You are not exploiting the distribution of the data, and what this tells you.

Stochastic Methods like Monte Carlo will do this. All of the process "x's' will have different distributions of data. Monte Carlo simulation will use these distributions to generate data for these "x's'', and over thousands of iterations, simulate your best "Y" for your solution.


No comments: