Much has happened since the introduction of Business Analytics (BA) as a framework for data-driven decisions. The amount of data has exploded and technologies like Internet of Things and Machine Learning are only creating more data.
Even though the potential for making better decisions based on data improves, it hasn’t necessarily become easier. But the amount of data not only grows, it also accelerates. It arrives faster and is faster to become irrelevant again. So in order to compete on data-driven decisions, one must be able to make better decisions, faster.
You may ask yourself, what are you really going to use all that extra data for?
To put it a little bluntly: If you do not manage to turn your data investments into better decisions, the investment is wasted, and you risk falling behind in the race with the competitors who have already figured it out.
So, what's the next step after Machine Learning? And how do you translate the extra insight that data and Machine Learning provide into better decisions? To give an answer to this, let’s take a look at data and Machine Learning in the context of Business Analytics, and see how optimization can be used to compliment the BA portfolio you might already have.
Business Analytics in three levels
When we talk about data-driven decisions, it might be beneficial to see it in a specific framework. Let’s decide to use Business Analytics, which deals with the use of mathematical and statistical methods on data to make better decisions. A lot of literature has been written on the subject, e.g. "Competing on Analytics: The New Science of Winning" (Davenport and Harris 2007), which provides an introduction to the field and argues why analytics can be used to provide competitive advantages.
Business Analytics is a broad concept that covers a large number of methods. They can be divided into three categories, according to their complexity and what needs they address. Or what questions they answer.
The three levels are:
Descriptive Analytics - What happened? What's going on right now?
Predictive Analytics. Why did it happen? What will happen in the future?
Prescriptive Analytics. What should we do about it?
The first level, ’Descriptive Analytics’, deals with the past, up to and including the present. Methods here cover reporting, dashboards, and the like. These are essential tools that every business need, and it is easy to see that there is something to be gained for the business that has not yet mastered Descriptive Analytics. A company that is not able to invoice correctly could, for example, miss out on large revenues.
The next level, ‘Predictive Analytics’, takes it a step further and uses existing data to create new knowledge. This could be to uncover coherence correlations between seemingly unrelated data or to use existing data to predict how data will evolve in the future, sometimes with great accuracy. This is where Machine Learning comes into the picture.
Third level, ‘Prescriptive Analytics’. This is where optimization comes into the picture. Here, mathematical models are used not only to create insight, but to actually make (or at least suggest) decisions based on data.
The Coffee House - An example
As an example, let’s consider a hypothetical company ‘The Coffee House’, a company that services coffee machines. In our example, the company has a team of service technicians who drives around to customers every day and service their coffee machines. This is both planned service visits, according to the machines' specifications, repair visits due to machines that are broken, as well as installation of new machines at new or existing customers. So basically, a simple service company that helps other companies with the most important thing - good coffee.
You could argue that there is hardly any need for planning. Each technician may have a fixed program with regular customers, or he will grab a stack of work cards from the office every morning and spend his day visiting customers - with occasional detours for emergency visits. This is a way of running the company and it might have worked this way for years, but it also might be a really unstructured way of running a company like The Coffee House. How do we know if the company is really doing? And could the resources be used better? You may have an overall idea, mostly based on intuition, that it’s all in all "fine", but how do you know if the company has been run in the most efficient way? The basic operational decisions, which customers should be visited when and by whom, are at best made on an ad hoc basis. Can we be sure that those decisions have been updated over time? By analyzing the company's data (Descriptive Analytics), we might pick up some signs that not everything is as it should be:
More and more employees report overtime hours
Several planned customer visits are postponed, some so the scheduled service is delayed or canceled altogether
The number of visits per technician decreases
The number of emergency repair visits is increasing
Employees spend relatively more of their time on the road, rather than with customers
So, with the help of Descriptive Analytics, we gain insight into how the company actually runs.
But with Predictive Analytics we gain even more insight: Based on data, it seems like an obvious hypothesis that more emergency tasks have arrived because there is no time for the planned service visits. So a vicious circle. But is there really a connection between the two? Using methods from Predictive Analytics, e.g. Machine Learning, we can ask why it happened and investigate whether there really is a correlation. Maybe there is. But perhaps some of the customers have grown and have hired more employees without acquiring more coffee machines. The machines are used way more than the set service intervals.
Another option for Predictive Analytics is the ability to predict breakdowns. You can imagine that advanced machines can transmit sensor data, which then can be collected and used to predict the probability of an imminent breakdown. Then we have the opportunity to catch the problems in advance. But so far we have only set up more data to consider. Decisions still have to be made, but there is no easy way to make the decisions. This leads to the last step in Business Analytics, Prescriptive Analytics: What should we do now? By formalizing decision-making, we have the opportunity to evaluate different decisions up against each other before they are carried out in real life and perhaps even choose the best one.
Some of the decisions to be made include:
If customers over-utilize the capacity of their machines, we must increase the service frequency or recommend customers to acquire new machines. Maybe we can earn more on more service visits, but do we have the capacity to do so?
If technicians spend a lot of their time in their vehicles, they may have an incorrect distribution between technicians and customers. By using methods from route planning, the technicians would optimize their distribution of customers and the routes between them.
If the technicians often need to return to the warehouse to pick up extra spare parts, you may need to look into the possibility of constructing one or more smaller warehouses. But where should they be built, and how many should be built?
If we know that a machine is in danger of breaking down, should we plan an extra visit? If a technician is nearby anyway, it may be a good idea, but if it happens at the expense of another visit, then what should be the priority?
There are several decisions to be made. For some it might be easy to see the consequences, but for others it can be difficult.
Optimization is an important part of the Prescriptive Analytics toolbox. In optimization, you take your decision problem within planning, investment or something else, and formulate it as a mathematical model. A model consists of three parts:
Decision variables, which describe the decisions to be made, and a domain for each variable, which describes which values the variable can take.
Constraints that describe the possible interactions between variables.
An objective (or goal) function that describes the overall utility of the decisions.
The result of an optimization assigns a value to all decision variables so that all constraints are met. If we look at the examples from The Coffee House, we can describe the decision problems as mathematical problems. We can solve them with specialized software, or even implement an algorithm that solves the problem. Maybe even a combination. But it may not be entirely obvious to see why optimization problems are so complex compared to the other types of Business Analytics. The answer is probably that they all contain complexity which manifests itself differently.
For optimization, the complexity comes from the inter-dependency between the decisions. The number of possible solutions can be all combinations of all variables, which is a number that increases incredible fast as the problems occur. It is the same thing that makes it relatively easy for people to make good decisions for small problems, but quickly becomes unmanageable as the problems get bigger. As part of the final step in Business Analytics, optimization naturally belongs at the more complex end of the scale, along with the more advanced parts of predictive analytics. Here we move away from data as the basis for decision-making and consider decision-making directly.
We have looked at how the previous steps in Business Analytics can be used to identify problems and opportunities. Optimization is about taking advantage of the uncovered opportunities. Therefore, optimization is also suitable for improving the use of resources in areas where you might not discover big problems - it is rather about exploiting the potentials that the other levels have identified (see for example "The Optimization Edge: Reinventing Decision Making to Maximize All Your Company's Assets", Sashihara, 2011). The somewhat cautious guide on when to get started with optimization would therefore be: Use the lower levels in Business Analytics to identify problems and opportunities and use optimization to get the best out of the opportunities that your data uncovers.
In this blog post, I have discussed the importance of companies for making high quality decisions, quickly, as well as how this ability provides a clear competitive advantage over competitors who are unable to keep up with development.
With the rapid growth of data and methods of data analysis, the ability to make good decisions has increased significantly. But so have the requirements for both the quality of the decisions and the speed with which they must be made, as the competitive landscape, the volume and speed of the data, as well as the complexity of associated software products change.
Using a service company like The Coffee House as an example, we see how even a relatively low-tech company has an increasing amount of data and complexity to deal with. Using Business Analytics as a framework, we can identify how the company performs on specific parameters and how data can be used to make the company even more competitive.
Fortunately, data-driven decisions are already part of the business strategy at most companies and thereby they are already climbing up the analytical path. So what's the next step, after the company has implemented reports, dashboards, and maybe even methods from Machine Learning?
As you might have guessed the answer is optimization or using math and software to make decisions. This way, you make better decisions faster and avoid many of the biased assumptions or guesses that are normally part of human nature.
Starting using optimization in business decisions is in itself an optimization problem but be certain that a competitive advantage awaits the first who masters it.