At its core, all business is about making calculated projections on human behavior—attempting to anticipate the future.
Companies that make the most accurate projections and operate on these projections in the most efficient manner, excel; the others, not so much. Every day, business men and women make decisions and solve problems. That’s the nature of the game. The success of these decisions, these solutions, however specific, however small they may be, ultimately add up—or not—to the success of the firm.
So, the trick is to make the right decisions and address the problems quickly and efficiently. Usually, we consider the decision, the problem, through the lens of experience and filter that experience through our definition of success. The greater our experience, the more specific our definition of success, the more likely we will make the right decision and find the best solution to the problem. Sometimes a gut feeling is really the voice of accumulated experience.
In fact, you could look on experience as the aggregate of bits and pieces of information that we’ve acquired over time and our definition of success as the most cost-effective, time-effective, service and market-effective contribution to the bottom line.
Today, our ability to collect and interpret bits and pieces of quantifiable information, data, about the market, our customers, and our performance, can augment, if not supplant our experience. Certainly, in today’s data conscious world, we should check our experience and base our decisions against what hard data may tell us.
There are three basic types of data analysis: descriptive, predictive, and prescriptive.
Descriptive analytics are exactly that. They describe what has happened. Did we meet our sales quota, keep our inventory under control, reduce delivery time, increase our profit quarter to quarter?
Predictive analytics make an attempt to look into the future.
All data is, by definition, in the past; but looking at that data over time, whether measured in years, months, or nanoseconds, can reveal trends that may indicate the likelihood of a future outcome. By keeping careful track of historical variations in demand, for example, a company may be able to forecast what will probably happen in the future, and thus optimize sales and operation planning. A history of missed or slow loan payments may indicate a credit risk. Bear in mind that as our ability to measure, collect, analyze, and report expands, so does our ability to predict.
Prescriptive analysis goes a step further and recommends a possible course of action. Given data that indicates an observable trend, what are the possible courses of action and, again, what is the likelihood of a particular outcome? With this awareness, firms can potentially rehearse the future. If we know when demand will peak, we can run and tweak alternative scenarios and determine which is likely to produce the most desirable outcome.
The data alone doesn’t really tell you anything. You need models and algorithms—written as mathematical expressions—that a computer can process.
It’s not as complicated as it sounds. We process data through our own models and algorithms every day. You walk into the kitchen to cook a meal. You’ve got tools—pots and pans and ovens—and resources—food in the fridge and pantry, spices in the cabinet. That’s a data set. We may have a recipe or two—models—that are possible. Now we have a process, the way we actually combine the ingredients according to the model, and finally we have a finished meal—output—that we can test against the tastes of our family.
But let’s say that it’s not just the refrigerator and the pantry, and not one or two recipes, but every recipe, every grocery store, and every market in the world. When the data set gets that big and the models that numerous, then we humans have a hard time sorting through the possibilities.
At one point in the past, we built the algorithms, but now we build algorithms that build and teach other algorithms. At first, none of these algorithms, these “bots,” are particularly smart, but just like we learn from experience (aka failure and success) so can bots. Sort of.
Say our bot is trying to learn the difference between a mushroom and a potato. We have a teacher bot give our learner bot a bunch of photos of mushrooms and potatoes and tell it which is which. The teacher bot can’t really teach, but it can test. The bots that test well are collected and the rest tossed. These higher-performing bots are examined, recombined, “re-schooled” and the testing, collecting, rebuilding process repeated again. And again. And again. And we’re not talking about three or four repetitions and three or four bots, but thousands and thousands of bots and thousands and thousands of repetitions. Of course, the whole process takes place in a computer that can run these thousands of iterations in a few minutes, if not a few seconds.
It’s called machine learning, which is something of a subset of artificial intelligence. Essentially, you feed the machine millions upon millions of examples of success and failure and tell it when it succeeds and fails. The machine then “learns” what works and what doesn’t. But the bot is only good at one narrow task—maybe the difference between a button mushroom and an Idaho potato. A Portobello and a new potato? Not a chance. So, the remedy is to make the test longer and feed the bot even more examples. And that, by the way, is why big data sets are so important. More data, more tests, better bots.
Part of the Alphabet group of companies (the largest of which is Google) is reCAPTCHA. You’ve seen their program. It’s the one that ensures that a human is logging into a website, and not a bot, and it does so by generating tests that humans can pass, but computers can’t. If you notice, lately, a lot of Captcha images have to do with traffic, road signs, lights and crosswalks. Tech pundits have suggested that every time we pick out the road signs or the stoplights, we’re helping teach and build the bots for self-driving cars.
Hard Data. Data Analytics. Data-based decision-making. It all sounds so factual. And with billions of brontobytes of data (yes, brontobyte—1,000,000,000,000,000,000,000,000,000) filtered through ingenious algorithms, how could the descriptions not be dead on, the predictions precise, and the prescriptions perfect? Well, because somewhere along the line humans collected the data, or determined what data would be collected and humans wrote the algorithms that filtered the data. And humans are biased.
Those biases come in a variety of forms and even the tiniest skew can create huge errors. Our experiences, over a lifetime, have told us what works and what doesn’t. It’s tough to set aside those experiences because the sum of them is who we are. So, when we cull data and define a measure of success, when we write an algorithm, it’s natural to look at the data, look at the success through the perspective of our own experience, accepting what our experience tells us is correct and rejecting all else. That’s confirmation bias.
Elizabeth Loftus at the University of California showed photos of car crashes to two separate groups of volunteers. Of one group she asked, “What was the speed of the cars when they hit?” To the other, “What was the speed of the cars when they smashed?” Not hard to guess what group said the higher speed. Based on the phrasing of the question asked, she got two different interpretations of the photos. Interpretation bias.
Predictive bias refers to how accurately a collection of data—say education, race, gender, income, age, zip code, etc.—can predict performance. There’s been a lot of controversy over Predictive Policing, which uses historical data to predict the likelihood of a crime or the likelihood of a previous offender re-offending. From Scientific American:
In what is the most widely cited piece on bias in predictive policing, ProPublica reports the nationally used COMPAS model (Correctional Offender Management Profiling for Alternative Sanctions) falsely flags white defendants at a rate of 23.5 percent and black defendants at 44.9 percent. In other words, black defendants who don’t deserve it are erroneously flagged almost twice as much as undeserving whites.
The data reflects the racially imbalanced world in which we live. The bias is baked in.
For the most part, the biases are known and the careful, considered analysis of data can provide extraordinary insight, particularly into the quantitative side of business operations. Moreover, as British philosopher Carveth Read said, “it is better to be vaguely right than exactly wrong.” However, it’s prudent to look at the results of big data analysis with a critical eye and acknowledge that an exclusive reliance on data analysis may insulate a business from the qualitative aspects of customers’ lives. Thus, “Thick Data.”
Thick data is a bit old school. It’s the product of ethnographers, anthropologists, and sociologists actively observing human behavior. Its strength lies in its ability to discern the qualitative “whys” around human behavior. Big Data is vastly superior at determining the quantitative aspects of human behavior and performance, the “what” and the “how much,” but Thick Data can point to the why. It can tease out what a customer or market might want or need but cannot easily define.
Faced with declining sales, a European grocery chain, in addition to its standard data analysis, undertook a major survey/questionnaire—6,000 customers, 80 questions. The responses, however, offered no insight. In roughly equal numbers, customers said price was the key factor, but they would be willing to pay more for quality. The company commissioned a team of researchers to look a little deeper. After two months of literally living with customers as they shopped, planned, and cooked, the researchers uncovered a major shift in consumer behavior. Apparently, families were not sitting down to eat at the same time every day. Sit-down dinners had become a thing of the past. The dining table was a place to collect mail and plug in computers.
The research also uncovered a new insight on shopping behavior. No more planning meals and shopping once or twice for the week. Nine trips a week was the average. There was no shopping for price. Customers needed fast, convenient, and to some degree distinctive.
The research team checked their findings against data the chain routinely collected about their stores and their competitors’ stores. On the one hand, it was no surprise that the best performing groceries were located in high traffic areas. On the other hand, an interesting insight emerged: these high performing stores had a style and a “mood” uniquely suited to the demographic of the particular locale.
Not to set aside issues of price and quality, but when the chain focused on tailoring the shopping experience to the lives of its customers, sales increased. The integration of Big Data and Thick Data produced an understanding that improved the bottom line.
Here’s a thought: Maybe decision-making based exclusively on data is incomplete, particularly when human behavior is involved. Perhaps, data is only a component in a process which—to be complete—must involve observation and critical thinking. Viewed in this way, analytics are just one instrument in the decision making, problem solving toolbox. And perhaps among the first questions we should ask before we pick up any tool is whether we are asking the right question.
Thomas Wedell-Wedellsborg, a contributor to the Harvard Business Review, notes that “85% of C-suite executives felt that their organizations were bad at problem diagnosis . . . and agreed that this flaw carried significant costs.” Analytics may point you emphatically in a particular direction, but if you’ve asked the wrong question, you’re emphatically wrong.
He goes on to describe the “slow elevator problem.” A large office building manager gets repeated complaints that the elevators are too slow, especially during peak hours. The first reaction is an analysis of elevator operation. Can the motors be re-tooled, or the response algorithms rewritten? Can elevator function be improved—how it reacts to call buttons, or how long doors stay open and how quickly they close? Does the building allow for the elevators to be enlarged, or another added? The second reaction might be an analysis of peak demand. How exactly does elevator demand fluctuate during the day? Can staggered work schedules and lunch breaks smooth demand? Can we depict, predict, and prescribe? Yes, we can; but just because we can generate a mountain of data to support a particular solution, is it the best, or even a better solution?
The solution in question turned out to be far simpler, far cheaper, and far more elegant. Mirrors. The building manager put up mirrors next to the elevators. Turns out that people, looking at themselves and the others who are waiting for the elevator, lose track of time. Complaints disappeared. When dealing with people (when are we not?), the manager who can step back from a situation and think critically about identifying the real problem is the one who will be successful.
Wedell-Wedellsborg remarks that the solution is particularly interesting in that it is not a solution to the stated problem. He notes that the initial framing of the problem is not necessarily wrong nor are the solutions data analysis suggests. However, “the point of reframing the problem is not to find the ‘real’ problem, but to find a better one to solve.” Here’s where critical thinking—and to some degree creativity—comes in, as do observational skills and an awareness of human behavior. If we think about this problem only in quantitative terms and turn to data analysis too quickly, we may miss the optimal solution. Sometimes, the ability to process terabytes of data in nanoseconds just gives us the ability to make poor decisions faster.
Big Data, Analytics, Machine Learning, and (soon) Artificial Intelligence are a permanent part of our lives. We benefit daily from the efficiencies and conveniences these technologies bring. And every day we rely on them more and more. After all, we need to understand what’s going on around us and we’ve always wanted to get a glimpse into the future. One step ahead of the competition is the secret to business success.
Let’s use the data and the analytics and the machine learning and the artificial intelligence to get that extra step, but let’s not lose sight of the importance of our own critical analysis, our own creativity. That’s what created all these technologies to begin with.