Before we completely transition from analytic methods to all the things we need to think about after our analysis is complete, it's useful to take a step back and remind ourselves what types of problems we're trying to solve and what the business drivers are behind them. Our objective is start shifting our focus back from the technical aspects of our analytical process to the contextual nature of underlying problem. We also want to reinforce the connection between some of the methods you may have seen in other courses and the real world problems they try to address. We obviously can't cover every possible type of analysis you might come across in your organization. But we'll try to present a cross section of problems that occur frequently in many industries. In this video, we'll look at what we might call classic problems that have a significant history in business functions. Specifically, we'll discuss segmentation, direct marketing and customer lifetime value, statistical process control, routing and inventory management, queuing and staffing, and credit risk and fraud. In a separate video, we'll focus on some emergent problems and analytical approaches associated with technologies like the web, mobile and social media. For this overview, we'll keep our discussion at a pretty high level. We'll take each idea, describe what it is, talk about why we actually do it, then list a few of the analytical methods that we might apply to make it happen. For some of our examples, we'll look simple visualizations that might accompany the analysis. This should help get you thinking about some of the topics coming later in the course. Let's start with segmentation. The idea behind segmentation is pretty simple. We're basically trying to find groups of customers or events or items with similar characteristics and grouping them together. Generally, we want elements in our groups to be as similar as possible to each other. But at the same time, we want the groups themselves to be different enough that it makes sense to treat them differently. When elements within a group are similar, we call that homogeneity within groups. And when the groups themselves are different from each other, we call that heterogeneity between groups. A good segmentation has both properties. There are many uses of segmentation. A business may develop different products to serve the needs of different customer segments. It may use different communication methods or channels to sell the same products to different customers. Some businesses even model their organizational structures after the segments they intend to serve. In most cases, the business is trying to optimize the effectiveness and efficiency of their activities by customizing them in some way to the market. Segmentation helps to inform how that should be done. Segmentation can also be used as a first step in a larger analysis. For example, I might use customer segments as inputs to a statistical model, like a regression, to predict some type of behavior, whether or not I intend to act on that prediction by segment. There are quite a few analytic methods that can be applied to execute a segmentation analysis, including factor analysis and principal components analysis, clustering, and decision trees or other propensity modeling techniques. For example, we might use a clustering technique that yields an output like this. Here we have three natural groupings of elements on two dimensions. We could consider treating these groups as separate segments in our business activities. Let's move on to some applications related to marketing and customer lifetime value. There's obviously a lot of different things that a marketing organisation might do, but one common need is the ability to predict whether a specific customers will exhibit some behavior of interest. Such as purchasing a product, taking an offer, changing products or leaving to go somewhere else. Usually, the reason we want to understand these behaviors is so we can interact with customers in a more efficient and effective way. It can be expensive to reach out to customers, and we'd rather spend our limited resources reaching out to those who are most likely to respond in the way we want them to. Even when communication is relatively cheap, like email or text messages, marketers want to avoid overloading customers with too many messages, and therefore seek to only provide customers with the most relevant information. Like segmentation, we can also use predictions as inputs to larger analyses. One example of this is customer lifetime value. A customer lifetime value calculation basically just adds up all the revenues and costs associated with the customer over time, and expresses them in today's dollars. It's usually represented by a discounted cashflow equation that looks something like this, where t is the time period and r is a discount rate. We can use event predictions to drive certain revenue or cost producing events. But one of the most common uses is to make predictions on how long a customer's lifetime will actually be, or what the value of n should be in this equation. Methods for even prediction can include propensity modeling techniques like logistic regression, decision trees or neural nets, survival analysis, or even simple historical observation. The use of deep analytics in the practice of marketing is almost a given in today's organizations. However, some of the earliest widespread uses of data analytics were in operations and manufacturing, where techniques were applied to help businesses improve efficiency and effectiveness. Statistical process control, for example, has been in use since the 1920s. Originally applied to manufacturing processes, the basic idea around statistical process control is that we establish a process, measure the natural variability in that process. Then set up a monitoring mechanism that alerts us if the process changes for some reason. Often this technique is used as part of a larger objective of process improvement where we understand a process, reduce the variation in the process. Then work to improve the overall level of performance in that process. For example, lets say we wanted to measure the percentage of defects at particular stage in an assembly line. And we wanted to be alerted if there is a spike in defects or if the average percentage moved up or down over time. We could use statistical process control to help detect these changes in pursuit of quality control. Similarly, we could use the same technique on quantity measures, like the amount of material required to accomplish a job. Or even time, like how long it takes to accomplish a task or to complete a process as a whole. Again, while this technique originated in manufacturing, it can really be applied to any repeatable process. And there are applications in just about every industry in functional area. As far as methods go, those used in statistical process control are actually pretty straightforward. Normally it involves application of classical statistics and sampling methods, coupled with some specific ways of looking at the data, like this control chart. Here we just have the value of some measurement over time. The lines at top and bottom are called upper and lower control limits, and are calculated using basic statistics derived from historical measurements. Early in the series, there’s some natural variation in the measure around the center line. But there's also one point that exceeds the upper limit. Later in the series, it appears that there has been a shift of some sort of in a process and where the points consistently lie below the center line. A second class of operational problems involve the management and movement of physical items, like raw materials, inventories, and finished products. Generally speaking, businesses want to make sure they have all the right items in the right place at the right time. But they also want to minimize cost and the certain constraints on production, storage, and transportation that restrict what they're able to do. Examples of common problems in this area include the newsvendor problem, where we need to choose how much inventory to buy based on expected patterns in sales. The transportation problem, where we must decide how to route product from one set of production points to a set of consumption points in the most optimal way. And the assignment problem, where we need to allocate materials or tasks to machines or other resources to allow the most efficient processing. The methods used to solve these problems, broadly fall into prescriptive analytics category. And include optimization and simulation techniques like linear programming, Monte Carlo analysis, queuing analysis, as well as a variety of specific mathematical methods and procedures. A closely related class of problems involves optimizing the movement, activities, and positioning of people. There are all sorts of problems that fall into this category. Here are some examples, how many people to staff at a call center or retail store. How to set up the order line at a fast food restaurant. Where to deploy sales personnel across regions. What level and mix of healthcare professionals should be present at a hospital. How to balance visitor traffic across an amusement park over the course of the day. The business objectives around these type of problems can be similar to those involving objects, including maximizing efficiency and throughput and minimizing cost. However, they also tend to evolve a customer or employee satisfaction component, as well. Most companies want to have happy customers and happy employees. And there are variety of ways in which staffing and movement can impact just how happy those people are. For customers, we may want to minimize wait times or ensure that the right resources are available to help the right customers. On the employee side, we may seek to avoid task monotony or best ensure that people's jobs are well-matched to their skills or other career objectives. As you might imagine, there are applications for this type of problem solving in many organizational functions, including operations, customer care, sales, and human resources. The methods we use to solve these problems are also similar to those we apply to the movement of objects, although the data elements we optimize or the constraints that apply might be somewhat different. Again, optimization in simulation techniques and application of queuing theory are at the core of the analytics we apply these cases. There are also some pretty innovative ways of visualizing these type of analysis, including be able to watch how costumers move through a 3D rendered environment over time. The last set of common problems we'll discus in this video are around credit risk and fraud detection. These problems are dominant in industries like financial services, but are used in pretty much any company that extends credit by performing services in advance of payment. Or where there is opportunity for people to financially take advantage of an organization via fraud. Some prominent examples are in telecommunications, utilities, insurance and healthcare. The business objectives here can be pretty straightforward. First, we're trying to balance the way we provide products and services to the market with the risk that we might not get paid for fully for those products and services. Based on the credit worthiness of an individual, an organization might structure its products differently. For example, a utility might require a deposit, or a mortgage company might offer lower or higher interest rate to different borrowers. We might also want to try to detect when something fishy is going on. This can mean different things in different organizations. One familiar example is with credit cards, where mechanisms are put in place to detect abnormal purchasing behavior, and may even put an automatic hold on the account when certain activity is observed. Typically, a company extending credit will develop their own risk models based on both internal data and external credit bureau data. The methods involved are generally predictive analytics methods like logistic regression, decision trees, or neural networks. For fraud detection, we might employ predictive models, but we may also use things like descriptive statistics, statistical process control or other data driven heuristics to trigger interventions based on observed activity. So that's a quick spin through some of the more well established type of analytic problems you might see in the world of business. Again, they only represent a fraction of all the applications of analytics that are out there. However, this is a pretty good starting set to get us thinking about how we visualize and interpret analysis, how we develop strategies and tactics to address the underlying business problems, and how we communicate both the results of our analysis and our proposal for how to take action.
