What is prescriptive analytics? – in short it is concerned with the optimal use of resources (people, transport, machines etc.) given a set of constraints and a well defined objective. One of the classical use cases is found in logistics – which loads should be batched together, and which routes used so that the overall cost of transport is minimised? This type of optimisation was once synonymous with operations research, although the term is not as popular as it was in the last decades of the last century.

The role of prescriptive analytics is complementary with business intelligence (descriptive analytics) and predictive analytics. Descriptive analytics tell us what has happened, predictive analytics tell us what might happen, and prescriptive analytics tell us **how** it should happen. Some suppliers of advanced analytics platforms (FICO and Angoss for example) are combining predictive and prescriptive analytics to determine which analytical models should be used under which circumstances. A set of predictive models that determine who should be given a bank loan, and under what circumstances (most bank loans are issued using predictive models), can be optimised for use through prescriptive analytics.

Real prescriptive analytic projects involve thousands of variables and constraints. The complex optimisation problems associated with airline operators and retail chains are good examples. Airlines have to optimise the segmentation of seats into various classes, optimise crew deployment, fuel purchasing (it is cheaper in some countries), routing – and so on. The aim is to optimise profitability, while maintaining safety, adhering to labor laws, satisfying customers and so on.

In practical every day terms products like Excel include a solver – a fairly primitive add-on for solving optimisation problems. Another, and more sophisticated set of solvers comes from Frontline Solvers, capable of handling much greater complexity problems. For very large enterprise optimisation problems IBM and FICO lead the way.

The evolution of mathematical algorithms and commodity hardware means that real-time optimisation is a possibility for all but the largest optimisation problems. So a logistics company may re-optimise its whole schedule given that some of the trucks have broken down and there is bad weather in one part of the territory. This evolution of the algorithms means that several distinct classes of problem can be addressed:

- Linear optimisation assumes a linear relationship between the variables in an optimisation problem and the objective – if the price of fuel increases for a transport company, then the profit diminishes in a linear manner. This was the first, and simplest form of optimisation, and in the early years it was quite common for these optimisations to take days and maybe weeks to complete.
- Non-linear optimisation assumes that the relationship between variables and the objective is non-linear (obviously). This may be a quadratic, or some form of exponential relationship – price demand curves being a good example. As the price of a product increases, the demand for the good does not decrease linearly. It does not have to rise much above competitor products before demand is zero. Non-linear optimisation is much more complex, and often requires much greater computing resources.
- Stochastic optimisation uses probabilities for inputs instead of hard numbers. So we might be trying to optimise a production facility based on a sales forecast. Most businesses will use hard numbers for the forecast, and if the actual demand is significantly greater or less, so the business will realise less profit and/or lost opportunity. If the input is not a hard number, but a probability distribution, the optimiser can work out the best production plan while taking this uncertainty into account. This is very powerful, and has direct links with predictive analytics, where the output from predictive models is very often probabilistic in nature.
- Integer (or discrete) optimisation only considers variables that are whole numbers. We cannot have 2.60 airplanes scheduled to fly from Chicago to New York every day, to meet anticipated demand. We have to settle for 2 or 3 since airplanes come in discrete units!

Until recently optimisation (or prescriptive analytics) has been an activity isolated from the live production environment, and form other forms of analytics. Today there are businesses using optimisation in near real-time to optimise offers to customers, to dynamically change the allocation of business seats on an airplane, to re-route delivery trucks – and so on. More advanced users are combining predictive and prescriptive analytics. The output from probabilistic predictive models can be used in stochastic optimisation, and a set of predictive models can be optimised for deployment by using optimising them.

If your organisation isn’t using prescriptive analytics, then it should be. This applies to medium size businesses as much as large ones. There is no excuse – send your bosses to this “What is Prescriptive Analytics” article. Tools such as those provided by Frontline are fairly easy to use, powerful, and require minimal training. Optimisation usually pays big dividends and we are on the cusp of an explosion in its use.