Data Analytics

Big data is often characterized by three (sometimes four or even five) V’s (e.g. Wedel and Kannan, 2016, Marr, 2016): Volume, Velocity, and Variety. More data was created in the past two years than in the entire previous history of mankind. At the same time, data is coming in at a much higher speed, often close to real-time. Furthermore, data nowadays is much more diverse, including not only numeric data but also text, images or video data to mention a few. The first two V’s are important from a storage and computational point of view, whereas the last V is important from an analytics point of view. 

On the other hand, several people have argued that big data is just a hype that will go away. When we analyze the popularity of the search term “big data” on Google, we find that the usage of “big data” as search term has grown explosively since 2008, but has stabilized since about 2015 (figure 1). Marr (2016) states that the hype around “big data” and the name may disappear, but the phenomenon will stay and only gather momentum. He predicts that data will simply become the “new normal” in a few years’ time when all organizations use data to improve what they do and how they do it. We could not agree more with him.

But, understanding and acting on an increasing volume and variety of data is not obvious. As Dan Ariely of Duke University once put it, “big data is like teenage sex: they all talk about it, they all want to do it, but no one really knows how to do it”. Wedel and Kanan (2016) put it more formally and argue that companies have invested too much in capturing and storing data and too little in the analytics part of it. While big data is on the top of many companies’ agenda’s, few of them are getting value out of it today. Therefore, in this special issue we did not only want to highlight “big data”, but also the “analytics” part of it, as state-of-the-art analytics is necessary to get results from big data. 

We believe that a very basic lesson from “old-fashioned” marketing analytics also applies to the “new world” of big data, but is too often ignored: begin with an end in mind (Andreasen, 1985). If we do not know what decision we are trying to make, big data is not going to solve the problem: we are searching for the needle in a haystack without a needle. As Wedel and Kannan (2016, p115) note, the primary pre-condition for successful implementation of big data analytics is a culture of evidence-based decision making in an organization. Companies that are successful with big data analytics often have a C-level executive who oversees a data analytics center of excellence within the company. In such companies there prevails a culture of evidence-based decision making: instead of asking “what do we think?”, managers in such companies ask “what do we know?” or “what do the data say?” before making an important business decision. The big data analytics movement will also affect us at HEC Paris. We want to highlight three aspects, which are further discussed in this HEC Knowledge special issue. 

First, it will affect the education and training of our students. In this special issue, the article by Daniel Brown discusses several initiatives at HEC Paris regarding the education of our students. Particularly, it highlights a new joint master with Polytechnique on big data integrating topics in statistics/econometrics, computer science, and substantive business areas such as marketing or finance. This new initiative will further extent the course offerings in business analytics that HEC Paris already has.

Second, it will increase the breath of our research. In this special issue, we highlight four recent studies to make that point. The first two studies by Peter Ebbes and Valeria Stourm discuss new developments in the context of marketing analytics. Their studies show how combining a variety of data sources benefits customer relationship management. The next two studies by Mitali Banerjee and Gilles Stoltz’ develop new machine learning algorithms to analyze pictures and to generate forecasts. These two studies show how such algorithms can be used to judge creativity and to aggregate forecasts to help businesses make better decisions. Lastly, David Restrepo Amariles and Xitong Li discuss several issues regarding big data legislation and policies. David shows that big data practices and algorithms are not always compliant with the rule of law, whereas Xitong argues that certain big data requirements in firm financial reporting can make financial reports actually more complex and harder to read. 

Third, it will bring along opportunities for collaboration on problems of big data analytics between companies and researchers at HEC Paris. For empirical researchers at HEC Paris it is of increasing importance to be exposed to current business problems and data. At the same time, companies benefit from a close collaboration with researchers by getting access to state-of-the-art solutions and learning about the latest business analytic approaches in a field that is moving very rapidly. In fact, in several of our own past research studies, we have already successfully collaborated with companies on substantive data problems. Our collaborations have let to actionable insights for the company AND academic publications for the researcher(s), a clear WIN-WIN.

With this Knowledge@HEC special issue we highlight some of the exciting initiatives and research studies that are going on at HEC Paris. We hope that it inspires companies and alumni in the HEC Paris network to reach out to any of us with new data opportunities or big data analytics challenges, to further help to expand HEC Paris as a research institution!

These Masters build on a series of courses already in place which have marked out HEC Paris as a leader in teaching the interdisciplinary world of data science. This so-called “fourth paradigm” in the scientific world is an expansion of the data analysis fields of statistics, predictive analytics and data mining. The likes of Peter Ebbes and Gilles Stoltz have been teaching these topics at HEC Paris for years. The two research professors and members of the GREGHEC Research Group continue to spearhead efforts to provide insight into data culled from both structured and unstructured fields.

Bringing Statistics Alive

“In teaching data science and statistics you should tailor your courses to suit both your personality and those of your students,” exclaims Gilles Stoltz, reclining in his new office in V building. “Because it’s important to galvanize them and convey your enthusiasm. Let’s face it, when they set the cursor, it brings alive a subject-matter that on paper doesn’t seem very inspiring at first.” The co-author of Statistique en Action  draws on exactly a decade of teaching at HEC Paris to explain his pedagogic approach. And, even if his colleague Peter Ebbes has been on the HEC Paris campus half those years, they both share similar beliefs on transmitting data science: “My classes are never one-way monologues,” explains Ebbes. “I quiz the students, we explore case studies together, the classes become conversations which keep them engaged. Okay, it’s sometimes tricky when I have 70 in the core classes, but I remain very keen on a participatory approach and we have some lively exchanges.”

Gilles Stoltz has been teaching his L3 statistics class since 2007 in a course with the logo “for today’s citizens and tomorrow’s leaders”. In the space of 20 hours he and his team of five-fellow statisticians present the fundamental theories to five classes of international students and eleven classes of French students. “I love this kind of teaching at HEC. It’s like doing theater! In theory they’re not too enthusiastic about statistics but by basing my teaching on a multitude of examples, I manage to gain their enthusiasm.” Stoltz shares with us one of the most popular practical cases amongst students. “Surprisingly, one focuses on the Iranian presidential elections of 2009,” he explains, “that’s when the Americans thought they had found a smoking gun proving the showdown between the incumbent leader Mahmoud Ahmadinejad and opposition candidate Mir Hussein Mousavi had been rigged. The Washington Post published an article, “The Devil Is in the Digits”, which analyzed the results using uniform distribution for the last digits of the vote but with an incorrect methodology. “They should have used the Chi-square test of goodness of fit! They would have seen that the P value is 7%, which is a typical value. So, as several statisticians pointed out at the time, the data is in no way a blatant proof of manipulation. The students like this example, it brings to life a situation which has an impact on the geopolitical world.”

A Multi-disciplinary Field Invites a Multiplicity of Students

Peter Ebbes focuses his teaching on business analytics, marketing research and marketing models. For the former professor at the prestigious Penn State University, there are three skills the students should aim at acquiring: “Advanced knowledge of statistics and math, IT and a substantive field like marketing”.“Many graduates master two of these three skills, very few manage to acquire all three,” he explains from his office in W1 building. “Added to this, it’s a real plus nowadays to acquire all-round knowledge on computer storage, moving data from one place to another. Acquiring all these is the big challenge we now face as we are challenged by this quantum increase in data.” The multidisciplinary nature of data science is reflected in the diversity of participants in Ebbes’ courses. “The catchment is really wide,” he points out, “We have some interesting characters in the MBA courses, there’s one from Google analytics, another with a PhD in philosophy, for example. The latter has provided some brilliant and unexpected insights into the case studies we’ve been working. This diversity really enriches our work.”

Important Backing From HEC Foundation

Could this variagated horizon be why there is an increased call for such data science and analytics experts like Peter Ebbes? In the United States alone, it is believed the gap in data scientists could be at 190,000. “There is a clear demand,” explains Ebbes, winner of the 2016 HEC Foundation best researcher award. “Businesses are hungry for specialists who can sift through their data. And, meanwhile, our research benefits from access to their raw data. It’s a win-win situation.”

The decision by HEC to purchase an eye-tracker has also helped researchers at the school better study consumers’ purchases online. “For example, what do people do with online reviews? Do they read them? How many? How does it influence their decisions?” asks Peter Ebbes. “With a PhD student, we’ve been overlaying the data with the eye-tracker which helps us know how and what to read.” Ebbes insists the help the HEC Foundation has given in funding such programs has been vital “It’s one of the most supportive in Europe, on a par to the support that institutions receive in the US,” he insists. The Dutch academic also stresses the positive impact HEC’s business connections has on research. “The school has a long history of collaboration with top multinational companies, it’s very important to work with them. I only wish French businesses were more aware of the opportunities there are to collaborate with us on research.”

Paris-Saclay University on the Data Horizon

The proximity of major companies are a clear asset in HEC’s teaching of data science, according to Gilles Stoltz. The teacher of Sequential learning and sequential optimization at M2 level believes the best way to transform the courses’ theories and field work into practical applications is through internships. ““Some of my PhD students in math have been hired by private companies under the CIFRE agreement,” explains Stoltz. “They have access to field data and have to develop theories based on their academic trainings and on-the-field experience. When they elaborate an idea for an algorithm they needs to code it, sometimes using all the big data framework and environment. The private sector provides the ideal setting for this. Not to mention the fact that their salaries are heavily subsidized for the companies!”

With this in mind, Gilles Stoltz believes the rapprochement between HEC Paris and the other establishments in Paris-Saclay University can only be positive.  But he is also enthusiastic about the developments developments at HEC Paris. “HEC wants to recruit a senior academic in business analytics, and we’re in a good state of mind. Statistical tools are being used in a multiplicity of classes, especially involving marketing and finance.” In the early 2010s, French affiliate professors of marketing had only a qualitative approach, Stoltz maintains. “Nowadays, we also have new younger academics like Daniel Halbheer, Cathy Yang and Ludovic & Valeria Stourm who are teaching marketing in a very quantitative way. Offering the two approaches is a real progress.” And the dynamic researcher concludes the department’s approach has made teaching data science a less dense and impersonal challenge than mirror courses on the other side of the Atlantic.

Law is changing. Scientific, Mathematical, Algorithmic, Risk and Technology driven law (SMART-Law) is becoming a reality. The driving force behind this approach are advances in learning algorithms and the big data that they create. “We are currently experiencing a paradigm shift in the way that regulation and law functions in our society,” explains David Restrepo Amariles. “We are moving away from the familiar democratic system where laws were created and enacted by parliament and then enforced by judges and police. Here, rules and facts were separate. With SMART law, rules are drawn from facts and execution of the law is automatic.”

What are concrete applications of SMART law? 

Restrepo Amariles notes that SMART law is one part of the journey towards SMART cities, the first of which is Songdo in South Korea. “In such cities, everything is connected to the city system, and that system determines the rules and enforces regulation. For example, it is not possible to take the garbage out if it is not the right day, cars cannot drive faster than the speed limit, and in-car breathalysers prevent drunk-driving. This automated law enforcement means the law cannot be broken.” Even outside of SMART cities, there are many SMART devices in use that produce lots of data and are connected to the internet.

This data is then processed by algorithms, which are used to enforce legal rules. He notes, “In the UK, Her Majesty’s Revenue and Customs (HMRC) developed Connect, a software device used to help investigate tax fraud. This uses big data, collecting structured and unstructured data (such as that from social media and auction sites) to assess how well an individual’s lifestyle compares with their tax declaration. Today, approximately 90% of UK tax investigations are opened as an outcome of this software.” 

Lawyers need to apprehend the problem posed by smart law from a technical perspective in terms of computer science.

How do learning algorithms function?

Some of these algorithms can learn. “The Connect software was programmed by humans to collect data and confront certain legal rules. As it uses learning algorithms, it is able to evaluate, for instance, tax payers' response to HMRC interventions through the Analytics for Debtor Profiling and Targeting (ADEPT) and use this new information to recalculate behavior models and risk profiles. In the short term it might also able to identify that people who regularly buy goods on certain auction sites are less likely to declare VAT and flag them for investigation,” he explains. The key here is that the algorithm is not given this information but instead uncovers it all by itself.

Do learning algorithms abide by the rule of law?

Restrepo Amariles wanted to find out how learning algorithms can be used to improve regulation. He also investigated the effects they might have on our civil liberties. Looking at smart devices in five different fields — intellectual property, tax law, commercial and financial law, security and law enforcement, and corporate compliance — he focuses on issues that might arise related to data protection, privacy, discrimination and due processes. He points to the Uber algorithm as an example of how learning can lead an algorithm to become discriminatory: “It allocates car rides based on proximity and driver rating.

However, if a driver of a certain ethnicity is consistently poorly rated, the algorithm may learn this discriminatory conduct of consumers and exacerbate its effects, instead of helping to revert them. In the end, the learning algorithm has learnt to behave in a discriminatory manner, affecting its supposed neutrality.” This will mean that such an algorithm may no longer comply with anti-discriminatory laws.

He then asks the question, “How can we hold the algorithm accountable for this discrimination?” In truth, once it has learnt, it is not only difficult to hold an algorithm to account, but also to assess if its actions are indeed discriminatory. It may be possible to find out through reverse engineering or undertaking a ‘black-box’ experiment whereby the algorithm is supplied with information and you see how it responds.  

Certified: Rule of law compliant

 “We need to understand how to ensure algorithms are compliant with the rule of law, and also help make law more efficient,” Restrepo Amariles concludes. As this is a non-trivial task, he suggests that there should be a technical standard created that can be used across markets. Here, other algorithms could be created to test algorithm compliance with the rule of law. If they pass the test, they will be certified rule-of-law compliant. “Banning SMART devices to protect civil liberties would be dangerous and stunt innovation. Instead, a trusted system of certification should be created to protect individuals and industry.”

Creativity is a crucial aspect of human culture, yet it is hard to define and harder yet to measure. “The essence behind the term is being able to come up with new ways of seeing and doing,” HEC researcher Mitali Banerjee explains. Creativity not only defines the work of artistic pioneers or visionary scientists, but its different forms also animate the activities of businesses in industries ranging from technology to entertainment. “In some instances, such as the iPhone, creativity can involve recombining existing technologies in a new design; in other instances, creativity can involve designing new organizational processes,” says Mitali Banerjee. Apart from the difficulties of measuring creativity, little is understood about how creativity is valued in our society. Does everything creative become successful? “It is possible that some very creative works fail to rise to attention because of biases against such work or because of market structures,” the researcher points out. 

Measuring creativity and creative success on a large scale

Given that creativity is a key performance outcome in many fields, how does one go about evaluating the creative merit of an individual work? Past measures of creativity have often relied on expert evaluations, but, according to Mitali Banerjee, such evaluations tend to give an incomplete picture of the relationship between creative merit and success: “Asking experts to evaluate the creativity of even 100 artists entails a significant drain on their cognitive resources; having multiple experts might address the problem, but that involves introducing the experts' own subjective aesthetics and hence more noise in the measures.”

The need for a big-data approach is about more than overcoming the simple biases and limitations of experts: “Prior research suggests that creative work becomes famous or canonized. However, because of data limitations, such work has focused on a small set of creators, who often happen to be already the most celebrated ones,” says Mitali Banerjee. “With a computational measure of creativity, we can start examining if such results hold for a larger sample of innovators. Research in psychology and sociology suggests that even though we like the notion of creativity in principle, we are actually biased against creativity in practice. More often than not, new ideas and technologies encounter resistance rather than a red carpet.” Professor Banerjee decided to work with computer scientists to measure the creativity of paintings on a large scale and thereby gain a better understanding of how creativity is understood and valued in our society. 

There is a strong correlation between the expert and machine evaluations of creativity--so in that respect the computational measure reflects how we conceive of creativity.

A computational model for artworks

This novel measure of creativity relies on recent advances in statistics and computer science, which enable the exploitation of what is known as big data. The computation model was applied to visual features of more than 4,000 images of 55 early 20th-century artists. They include Natalia Goncharova, Pablo Picasso, Paul Klee and Robert Delaunay, who are associated with a range of abstract art movements including Cubism, Futurism, Orphism, German Expressionism and Rayonism. The researcher chose this period because it represents a paradigm shift in the history of modern art: “Until the 1900s, representational art had dominated much of the Western fine art world,” she comments. “Around then, we witness the emergence of several innovations in artistic style, broadly grouped as abstract art, which represent a distinct and radical break from the representational paradigm.”

Mitali Banerjee worked with several different algorithms with different computer scientists to extract features (some recognizable to human beings, like “ruggedness,” others recognizable only to computers without any such human interpretation), which in turn provided numerical representations known as feature vectors. (For the more enlightened, these approaches include unsupervised learning methods such as k-means clustering and artificial neural networks which are used to extract features vectors to represent a painting. Other methods such as k-nearest neighbors are used to classify the style and genre of the paintings). In parallel, Mitali Banerjee asked art historians to rate the artists' creativity, broken down along several dimensions: originality, stylistic diversity, abstraction, uniqueness and innovativeness. She then combined these scores into a single value (factor scores) for each artist, and compared the human measure and the machine measure of creativity.  

The complementarity of human and machine evaluations

Mitali Banerjee observed a strong correlation between the expert and machine evaluations of creativity. This correlation provides some validation of the computational measure, “so in that respect the computational measure reflects how we conceive of creativity,” she comments. The correlation between expert and machine evaluations was strongest when the machine was evaluating the most creative works of an artist (according to the machine). This suggests that human experts are thinking of the most creative work of the artist, even though Professor Banerjee had asked them to evaluate the average creativity of each artist between 1910-25.

Given the strong correlation between both types of measures, she plans to apply this approach to visual art on a larger scale, in particular to understand how creativity and fame evolve together. It may sound shocking to have a machine-based measure of something so ineffable as creativity, but Professor Banerjee likes to see it as a tool to complement human skills: “We humans are great at recognizing and evaluating complex products. We can learn from a relatively small set of examples whereas a machine needs more input, but machines can reveal what we have ignored because of our inherent biases. They can illuminate new ways of seeing. In order for these tools to give us meaningful insights it is important to take an interdisciplinary approach which integrates theories and tools from art history, computer science, psychology and sociology.”

Concerning the impact of her work, Mitali Banerjee speculates, “given how much our society already values creativity, and the demonstrable impact that creative ideas have on the economy and culture, insight into how creativity functions can impact how our institutions incentivize, recognize and support creativity.”

Our wide social networks are increasingly connected through social media and telecommunications. Technology enables us to create more lines of communication and our networks can be tracked. Peter Ebbes explains how these social networks could be of interest to marketing executives. “Traditionally, customers have a ‘customer lifetime value’ that predicts how profitable a company’s future relationship with them might be,” he says. “However, so far, marketers do not consider the value of a customer’s social network as a component of their lifetime value.” Ebbes and co-workers wondered what the value of social connections might be: “Could the network of an individual make them a more or less valuable customer?”

Collaboration between business and academia: a win-win 

The team embarked on a successful collaboration with the Seattle-based company Amplero, a big data and analytics firm, which gave the researchers unprecedented access to customer data. In their field experiment, they used telecommunications information from over a thousand individuals. “We staged a traditional marketing campaign whereby a randomised selection of prepay telephone customers received a message designed to increase their phone usage,” Ebbes explains. “Normally, we would only be interested in how successful the campaign was in relation to a control group who did not receive a message, but here we wanted to see if there was a ‘trickle down’ effect of that campaign through the target customers’ networks of social connections.”

Firms and industries are in a great position to take advantage of the increased profits that a network can provide.

Marketing affects the social network

Only the target customers’ primary connections who use the same telephone provider were tracked. To be considered a connection, two calls had to have been made, to or from the target, in the four weeks prior to the experiment. Ebbes discusses the team’s findings: “The connections of those that received the campaign were found to be less likely to suspend their contracts and more likely to increase their activity, relative to the control group,” says Ebbes, who goes on to add: “This is interesting as the connections themselves were not targeted and did not receive a marketing message. As the experiment was randomised, we can say that the campaign caused an effect on both the targeted customer and also their connections. So, for the first time, we demonstrate that traditional marketing campaigns have a social spill-over effect.”

A social multiplier

As a consequence of the marketing campaign, targeted customers increased their overall phone usage by 35%. Individuals connected to targeted customers increased their usage by 10%. This increase was not seen in individuals connected to the control group. From this data, Ebbes estimated a social multiplier of 1.28. He explains, “We see that the campaign has a spill over effect of 28% on the usage of the target customer’s connections. This is something that could be utilised by businesses to increase the profitability of targeted customer relationship marketing.” The team then altered their analysis approach and showed that increased activity among the non-targeted but connected customers is driven by the increase in communication between the targeted customers and their connections. This means that the marketing campaign propagated through the network to affect the non-targeted customers. Consequently, these customers also became more valuable to the company even though the company did not spend a penny on targeting them. 

Networks increase potential profits for firms 

Ebbes emphasizes that this study is applicable to all businesses with clear network externalities, which have the ability to target individual customers and observe their social connections and activity. “Firms and industries that offer products using networks or whose services require underlying networks are in a great position to take advantage of the increased profits that a network can provide,” he says. “At the moment, these include companies with stakes in telecommunications, online multiplayer games or file-sharing services like Dropbox, but there are also potential gains for traditional businesses."

In 2009, the U.S. Securities and Exchange Commission began imposing a new financial regulation on public firms. After a three-year, three-step phase-in, all firms listed on U.S. stock exchanges must now submit their financial reports in not only HTML but also the newer eXtensible Business Reporting Language (XBRL). “Public firms are required to disclose their financial situation and make this information freely available,” Xitong Li explains. “Financial reports were traditionally submitted and stored in HTML format that can be opened on an internet browser and is easy for humans to read and understand.” The Securities and Exchange Commission adopted the new ‘language’ to adapt to the demands of an increasingly digital world: reports are now filed in a format that can be easily processed by computers, which are expected to be more accurate, faster and lower cost than their human counterparts. 

A counterproductive regulation?

Li and co-authors were concerned that this new regulation might affect the overall readability of HTML-formatted human-readable financial reports and wondered what the consequences of this might be. “Using XBRL means that information will be easier and cheaper to process,” he explains. “Firms might anticipate this and add more information to reports, making them longer, more complicated and in the end, they actually become harder to read.” If this scenario is the reality, Li suggests that the regulation could be acting contrary to its intended purpose. “Now, firms must file two reports instead of one, which is time consuming and costly,” he says. “Some think that this might mean individual reports are rushed and less carefully prepared, which also reduces their readability.”

The Securities and Exchange Commission adopted the new ‘language’ to adapt to the demands of an increasingly digital world.

Three-phase introduction of XBRL financial reports

In the years before the regulation was enacted, all firms would experience parallel trends in the readability of their reports, no matter their share value. In phase one of the regulations’ introduction that began in 2009, firms with share value in excess of $5 billion were required to file their financial reports in both formats. In 2010, phase 2 was initiated and firms with share values between $5 billion and $700 million adopted XBRL. Phase 3 followed in 2011 whereby all remaining firms had to adopt both report formats. “We compared the readability of reports from the high value phase 1 group of firms before and after the regulation was implemented, between 2008 and 2009, with those of firms from phase 2, who were only required to submit financial reports in HTML format in those years,” Li explains.

New regulation brings readability reduction 

The team noticed a dramatic reduction in readability of HTML-formatted reports between 2008 to 2009 for the high share value firms that were submitting in both HTML and XBRL formats. This was not seen with the firms who were not yet required to write reports in XBRL. “We used a tried-and-tested DID design method of analysis to conclude that this increased complexity and decreased readability of the financial reports is due to the new regulation,” Li concludes. “The regulation has unintended and undesirable side-effects.” Li is also quick to mention that the timing of the introduction of this regulation coincided with a period of financial crisis: “The study was carried out to control for the impact of the crisis. Even when this is factored in we still see that the regulation increases the complexity of the financial reports.”

Added complexity could bring investment losses 

HTML-formatted financial reports are relatively easy for humans to read and so are still in use by most analysts and investors even after the introduction of XBRL. These findings are not good news for the U.S. Securities and Exchange Commission, who implemented the regulation with the purpose of improving access to information. At present, reports are not produced, processed and exported by computers alone, and human manual intervention makes the process unintentionally messy. With less readable reports, firms will receive less attention from the public and investors, who cannot easily understand the figures. Li warns of the financial impact this could have on firms: “When an investor cannot read a firm’s financial report, they won’t understand its financial situation and will be less likely to invest in their stock.”

Whether used to predict weather patterns, determine how exchange rates might fluctuate, or in which direction consumer preference might evolve, accurate forecasts are key to success in business. In many industries, sufficiently large companies hire data scientists to provide such predictions while others buy forecasts as a service. 

But while decision-makers would often like their forecasting experts to come up with one single accurate prediction, in reality even just one expert can think of several forecasting methods, each of which usually requires tuning and some technical choices. This means companies are spoiled for choice when deciding which methods to choose. 

Such methods include traditional statistical models, whereby past data is modeled to come up with predictions. Then there are machine learning approaches, where a computer by way of an algorithm is able to automatically find hidden data in statistics without being told directly where to look. Within machine learning techniques there are also different methodologies (for example ‘random forests’, which uses decision trees, or ‘deep learning’ using multi-layered transformations of the data). Managers, when faced with a whole cloud of forecasts and what they sometimes see as too much choice, perceive this as a problem. 

But there is a solution. As in life outside of business, diversity can be an opportunity and instead of selecting one expert or one forecasting technique at all cost, there are ways to aggregate the cloud of prognosis into a single meta-forecast. This work (itself a form of machine learning) can be conducted by computers and is both automated and safe.  

At HEC Paris we have developed forecast aggregation tools to help businesses make decisions. So far we have used these in various industries: to forecast exchange rates at a monthly rate and with a macro-economic perspective [1]; to forecast electricity consumption [2] with France’s largest electricity provider, EDF; to forecast oil production (on-going); and even to forecast air quality [3]. 

So how does it work? First, we must design all the automatic processing of expert forecasts. We do this by creating formulae and algorithms, associating each expert with a weight and have the weights vary over time depending on the past performance of each individual. Once programmed, a machine then uses the algorithms to perform the desired aggregation automatically as a black and does not require human supervision. 

This way of aggregating expert forecasts performs well both in practice (i.e. when used for practical purposes) and in theory, and comes with what we call ‘theoretical guarantees of performance’, which is what makes it so safe. At HEC, as well as other research centers around the world, we have produced aggregation techniques that perform almost as well as, say, the best expert, and even the best fixed (or linear) combinations of experts. The important thing it to gather forecasts from a multitude of experts and we believe this is beneficial, as it increases the chance that at least one of them will be good. Of course aggregating forecasts does have its limitations. Firstly, constructing or getting expert forecasts that actually exhibit diversity is not always an easy task, as experts are often clones of each other and predict similar tendencies.

Secondly, this black-box aggregation of expert forecasts that comes from various sources (human experts, statistical models, machine learning approaches) does not attempt at all to model the underlying phenomenon, which is in strong contrast to more classical statistical or econometrics approaches. With aggregation, all efforts are put into forecasting performance and nothing is invested on the modeling. Of course this works fine if the final decision-maker only wants to make good decisions and not  try to understand her/his environment!  

As we continue to develop this aggregation solution at HEC, which is both on the theoretical and practical side, we have begun approaching various companies for R&D contracts and/or proofs of concepts. Our aim is to look at business problems even more closely related to the ones studied at HEC Paris, such as forecasting volumes of sales to manage the supply chain. Perhaps our work will help put decision-makers’ minds at ease when they face what can be an overwhelming choice of forecasting options.

Whether grabbing a coffee-to-go at the shop down the street or flying halfway across the world, consumers tend to enroll in loyalty programs for just about every possible purchase. On paper, it's a win-win relationship: the customer gets a tenth coffee for free or a free flight against accumulated miles, and the company gets repeat business. These schemes should be all the more successful as “rewards reinforce positive purchase behaviors,” says Valeria Stourm, who studies such loyalty programs.

But, in fact, consumers use their points a lot less than might be expected. One third of the $48 billion in rewards issued in the United States every year are never redeemed. And we’re not just talking about complicated programs whereby a threshold of so many points (typically, airline miles) needs to be reached to obtain a specific reward. Even simple, boring, a-point-for-a-dollar at the supermarket programs see customers stockpile points instead of using them to instantly reduce their basket price. Hoarding is neither particularly profitable for the customer (unredeemed points may expire, depending on retailer rules) nor for the business (outstanding points are "stuck" on the balance sheet). So why don't customers systematically collect the shopping rewards to which they are entitled? Faced with that puzzle, three researchers, Valeria Stourm, Eric T. Bradlow and Peter S. Fader, developed a model to understand why customers persistently stockpile points in so-called “linear programs” (the kind designed explicitly not to encourage stockpiling, with free-to-use points). 

Why consumers accumulate loyalty points

The novelty of the model presented by the team of researchers in their paper is that it brings together three different approaches to explaining why customers of linear programs stockpile: economic, cognitive and psychological motives. A potential economic motivation is that customers forgo the opportunity to earn points on purchase occasions in which they redeem. The second is a cognitive incentive: the transaction comes at a “cost”, albeit not a financial one. “Perhaps customers find it annoying to take out their loyalty card and ask the cashier to redeem,” suggests Valeria Stourm. The third, more sophisticated explanation is psychological. Based on mental accounting, this explanation supposes that we have different mental slots for different currencies, even when they are of equivalent monetary value; for example, poker chips in a casino as opposed to cash, or money budgeted for leisure as opposed to money budgeted for groceries. In this context, it means that customers do not perceive cash and points equally. With this model, a $10-purchase made by paying the equivalent of $3 from the customer's points account and $7 from his or her cash account would be analyzed as follows: The $3 are a loss to “points wealth” and the $7 to “cash wealth”. By redeeming, the customer also loses the opportunity to earn $0.10 in points (or whatever percentage of the price is set by loyalty program rules). Such debits and credits are evaluated separately as gains and losses in each mental account. It may sound silly to assess the use of points in terms of loss or pain, but the researchers did hear one customer from a linear program express sadness about point redemption in an interview: “It makes me feel sad because I don't have any points left on my card.” In addition, the researchers relied on prospect theory, which assumes that losses and gains are perceived differently when making choices. “Put simply, losses loom larger than gains, so the pain of losing $5 feels greater than the pleasure of winning $5,” says the researcher. 

It's more strategic to encourage redemption, and loyalty, if only from a purely marketing perspective.

Empirical confirmation from retail shoppers

The researchers estimated their statistical model on data from a large supercenter chain in Latin America. By tracking the purchases of a cohort of more than 300 customers over three years, they indeed found that many were sitting on a goldmine of unclaimed points: even though no minimum is required to redeem points, only 3% of all purchases had redemptions associated with them, even though redeeming could reduce basket price by 30% on average. “It doesn't mean customers don't care about the loyalty program, since they still show their card to earn points, and 40% of the customers in the panel eventually redeemed during the observation period,” comments Valeria Stourm. She and her co-authors hypothesized that customers waited until they had enough points for the loss in terms of points to be compensated in terms of cash (redeeming attenuates the pain of cash loss). And indeed, when testing the statistical model, the researchers found that their analysis based on dual mental accounts was relevant to forecast customer behavior. Cognitive motivations (fixed costs) also helped to explain observed customer purchase behaviors, but there was much less evidence to support explanations based on economic motivations.

Encouraging customers to redeem points?

Having thus identified the cognitive and psychological drivers of redemption behavior, the researchers were able to offer suggestions for companies to rethink their loyalty programs. Valeria Stourm remained cautious, however, explaining that the results of the policies evaluated in their study would be applicable to those particular cases only. The policies compared included offering rewards points for every basket price, regardless of redemption choices (an economic policy), automating redemption (cognitive policy), for instance by automatically reducing the customer's basket price if his or her stockpile is greater than 15 points, and a psychological policy that allows customers to redeem up to 100% of the basket price instead of the current 50%. But is it actually strategic to prod customers in that direction, since redemption comes at a financial cost for companies? Valeria Stourm admits that it's debatable: “From an accounting perspective, an unredeemed point is a cost you don't incur, but it's also a liability or deferred revenue on the balance sheet.” In her opinion, it's more strategic to encourage redemption, and loyalty, if only from a purely marketing perspective. “Customers who experience rewards may become even better customers,” she says.