As reported by most of the blogs and newsletters on technology and analytics, marquee banks achieve competitive advantage by leveraging analytics to improve decisions, enhance marketing, and influence consumer behavior. The availability of data and low-cost storage technology is creating an imperative for financial services companies to fully leverage and centralize analytics across their enterprise.
That is the popular belief. We live in that particular era of technology, where technology is considered capable enough to predict, with sufficient precision and accuracy, that which customer would need which banking product and when would he or she ask for it. Not only this, predictive analytics is also regarded so supple and nimble, that it could prompt offers for customers in real time which are potential enough to influence consumer behavior with immediate effect. In my view, we place such a tremendous amount of trust in data analytics that we miss to attend the classic paradox of PAST vs. FUTURE. Let us take a look at what the paradox is.
The core of predictive analytics relies on capturing relationships between explanatory variables and the predicted variables from past occurrences, and exploiting it to predict future outcomes. Putting it simply, one’s past data works as proxy for the expected outcomes of similar events in future. At this conjuncture, I ask a question: We toss an unbiased coin ten times, of which nine times the coin shows up its Head. The frequentist approach says that it is 90% plausible that the coin would fall on its tail when tossed next time. Does it mean we are 90% sure that the next toss would show up Heads? Does it mean that if a hundred coins are tossed, 90 of them are going to turn heads up? The answer is NO. Even though in the last ten tosses, the result was heads in 9 tosses, the eleventh toss is an independent event! Probability of getting a Head in a toss is still 50%. The reflection of probability in frequency of outcomes in equivalent events will be conspicuous only when the number of occurrences of the event approaches infinity.
Predictive analytics works analogous to the case discussed above. The input, more or less, to a predictive analytics system is the past data of all customers who enjoy banking services of debit cards, credit cards, deposits, loans etc. With application of predictive analytics on this past data, banks devise their customer relationship management strategies, including advertising and marketing campaigns; monitor customer buying habits, up sell and cross-sell initiatives; and attempt to build long-term customer loyalty, retention, customer screening and rewards programs.
The Cat and the Washing Machine
I am buying a concept of ‘The Cat and the Washing Machine’ from a book – Antifragile, authored by renowned economist Nassim Nicholas Taleb. The concept posits boldly that based on behavior, systems could be classified into two categories. One is the Cat, comprising of animate, organic objects; other is the washing machine – the inanimate ones. Washing machines – Inanimate—that is, nonliving—material, typically, when subjected to stress, either undergoes material fatigue or breaks. They are fragile. The house, the food processor, and the computer desk eventually wear down and don’t self-repair. However, typically, Cat – the natural—the biological—is antifragile, depending on the source (and the range) of variation. Such systems elicit stressors to grow: rather they need stressors for their development. As an example, a human body benefits from stressors to get physically and mentally stronger.
To broaden the scope of analysis, let us redefine the distinction. More effective is the distinction between noncomplex and complex systems.
The washing machine, the torch, the lighter, the fan and the pen have a fixed way of working. You press the lighter, containing inflammable fuel, in a particular fashion and it lights a flame. Whether in summer or winter, whether in India or Greenland, whether by a man or a woman, by a toddler or an octogenarian, its behavior is same, and consequently, predictable. These are non-complex systems.
Social groups, societies, governments, institutions, organizations, businesses, markets and eventually the economy are complex systems. Though they are created by man, they are closer to the cat than to the washing machine. They may not be strictly biological, but they resemble the biological in a way, they multiply and replicate. And they are antifragile and robust as well! If we think of rumors, ideas, technologies, and businesses, we shall find them to be of the ‘Cat’ category. The noteworthy point is that such systems have a typical kind of complexity associated with it. The complexity of such systems could be understood with the help of an example: We take a specific animal away from its place and we disrupt a food chain: its predators will starve and its prey will grow unchecked, causing complications and series of cascading side effects.
In essence, ripples are manifold and unimaginable in complex systems, in case of a minor rearrangement of its components. Moreover the ripples are drastically different in every single case. In non-complex systems, ripples are minimal. The issue with the business case of predictive analytics is that though it works on all complex systems, i.e. social groups, businesses and economies, the analytical treatment given to those by predictive analytics systems is the one deserved by the non-complex systems. The Cat is mistakenly dealt with as if it were a washing machine! J
To make the concept visible, let us take an example. The likelihood of a prediction being correct, of how a lighter (non-complex) would work, is much more than that of how a savings account holder would want to use his savings. The reasons are simple.
- We know how does a lighter function, and
- There is hardly a chance of any deviation in its function.
However, if we make similar assumptions for consumer (complex) behavior: that we know how does (a certain variety of) credit card consumer(s) function and that there is hardly a chance of any deviation in his/her function, we get into a serious trouble. The likelihood of the prediction being correct drops. This brings us to an important question as business managers/ consultants: What do we choose to predict and what we do not? The non-complex can be predicted accurately, but no one needs their prediction. The complex is not predictable by a layman and so this needs expertise (and probably therefore outsourcing). But the fact remains that accuracy of prediction is substantially diminished in case of complex systems.
I have been discussing such problems with data modelers and would want to discuss with many more in different organizations in future. What comes out is that the predictive analytics systems themselves are complex systems. This compounds the problem! To have a look at specific issues, please refer the following link.
I would request the reader to be in the shoes of a customer while going through the article above. As a customer, reading through the list of challenges faced in predictive analytics, I have started having doubts over predictability of the systems targeted at the consumer: at least partially.
Let us park the second idea. As discussed before, everything converges in the end. J
We have defined the paradox in the previous section. We now know that the market demands us to deliver prediction of a complex entity, as accurate as possible. Let us see how the traditional method does that.
It is widely known that traditional predictive analytics tools are based on the assumption of repeatability: that by storing and modelling data of the past, we can forecast or predict the same data of the future. Of course there are tolerance limits and there are explanatory variables which impact the values of predicted variable. In short, it is past data of x to predict future of x.
The other way is to detect that ‘x’ is determined by ‘p’ and that ‘p’ is not as risky as ‘x’ to predict! So, better use analytics to predict ‘p’. Predicting ‘x’ is then a cakewalk. The focus of business shifts from the skill of prediction to the skill of detection of what all ‘p’ like factors could be, and then to the skill of calculation of x based on predicted values of p. It is clear and so sellable too.
As an example of this approach, let us see how the Baltic Dry Index is calculated. The Baltic Dry Index (BDI) is a measure of what it costs to ship raw materials—like iron ore, steel, cement, coal and so on—around the world. The Baltic Dry Index is compiled daily by The Baltic Exchange. To compile the index, members of the Baltic Exchange call dry bulk shippers around the world to see what their prices are for 22 different shipping routes around the globe. Once they have obtained these numbers, they compile them and find an average.
The Baltic Dry Index is a leading indicator that provides a clear view into the global demand for commodities and raw materials. The fact that the Baltic Dry Index focuses on raw materials is important because demand for raw materials provides a glimpse into the future. Producers buy raw materials when they want to start building more finished goods and infrastructure—like automobiles, heavy machinery, roads, buildings and so on. Producers stop buying raw materials when they have excess inventory and when they stop infrastructure projects.
Typically, demand for commodities and raw goods increases when global economies are growing. For investors, knowing when the global economy is growing is helpful because that means stock prices, commodity prices and the value of commodity-based currencies should be increasing. Conversely, demand for commodities and raw goods decreases when global economies are stalling or contracting. For investors, knowing when the global economy is contracting is helpful because that means stock prices, commodity prices and the value of commodity-based currencies should be decreasing.
Though both the approaches of prediction eventually predict about complex systems, the difference between the two is that the traditional method is using a complex systemic variable as the means while the second approach depends on a robust and non-complex systemic variable. The business trick here is to use non-complex variable(s) to predict a complex one. Such relationships are there and if not, could be created.
For instance, unlike inflation in economy and guidance estimates declared by public companies, Baltic Dry index is robust in not only short, but medium run as well. It could be altered only by manufacturing more ships which takes at least two to three years or by trading quantities of raw material which are more or less than the actual needed. Either of the two cases would hardly occur. Also, this index is based on what has happened. The trading is already done. The predictor is placing itself smartly as he knows that the Baltic Dry Index could be a significant determinant of the stock prices. The traditional predictor would have used a model where growth projections of the economy and the sector/ industry would be inputs to predict the stock prices.
If a bank is looking at customer’s demographics, stage of life and family size as a means to present pension funds, children education plans and retirement plans, it is using robust parameters to predict a complex system. However, if past year’s transactions of credit cards spent on purchase of furniture or gadgets make a bank infer that the customer would prefer discounts on similar purchases ahead, chances are that this would classify under the header of iatrogenic.
If this article is appearing to be anti-predictive analytics, it is not. The intention is not at all to discourage prediction. We need predictions and we need predictive systems. All we do not need is an iatrogenic. The name for a net loss, the (usually hidden or delayed) damage from treatment in excess of the benefits, is iatrogenics, literally, “caused by the healer,” iatros being a healer in Greek.
I would illustrate this with a classic incident. In the 1930s, 389 children were presented to New York City doctors; 174 of them were recommended tonsillectomies surgery. The remaining 215 children were again presented to doctors, and 99 were said to need the surgery. When the remaining 116 children were shown to yet a third set of doctors, 52 were recommended the surgery. This story allows us to witness probabilistic homicide at work. Every child who undergoes an unnecessary operation has a shortening of his/her life expectancy.
The proportions of children recommended with surgery in the three events were 45%, 46% and 45% respectively. Do we see how strikingly similar are these proportions even when the event is completely random and outcomes completely based on medical tests and their evidences! This is a classic example of how the Cat was slanderously assumed to be a Washing Machine.
Similarly, it is the ‘Need to do something’ which appears at work in the case of traditional predictive analysis. It is an established fact that Humans (and hypothetically other complex systems too) view data in form of a normal frequency distribution. What does not get appreciated is that the normal distribution is of the nominal data and not real values. For example, it is usual for banks to call up customers earning slightly above average incomes in the age group 23-35 years for high end credit cards. The nominal data is the income and age group. The real values could be the ones considered and discounted with purchase aspirations, affinity towards credit and past experiences about possession of money. How often we find banks interested in affinity towards (the credit card kind of) credit of individual customers and the reasons underlying the varying predilection? This paves the way for an iatrogenic mechanism.
Usually, such iatrogenic patterns are not visible and so are not pronounced as a danger until it is late. J