Driving analytics by insight

Analytics has long played a pivotal role in the success of landscapechanging companies. Google, Facebook and Amazon have all made full use of their vast data sources from the outset. But, impressively, analytics also drives creativity. Industries that traditionally relied on gut decisions now look to data. Take Netflix, famous for not only recommending favourite movies to users, but also for creating entertainment by data guidance. Analytics helps pick the right mix of talents and resources – programme types, casts, producers, and investment - and dynamically determines the course of a new series in response to viewership figures. This affects decisions such as increasing or downplaying the role of an actor, changing a storyline, or even pulling a series.

Companies derive valuable insights into their customers, operations and marketing strategy through analytics. The people who play critical roles in this process are data scientists, and they typically receive their training in computer science – especially machine learning and artificial intelligence, as well as statistics and applied mathematics. Such technical training equips students with the ability to work with data and search for answers. In some cases, the technical ability plays a differentiating role, a matter of whether a problem can be solved or not, especially when the problem is large-scale or complex. Imagine an e-tailer who needs to deliver parcels as required by its customers in specified time windows. Failing that would mean unhappy and possibly lost customers. Solving the problem at the scale of hundreds of thousands of customers with hundreds of delivery staff in a city is a very challenging task.

Unlock the full potential of data

However, to release full power, I would like to argue that analytics has to be guided by and incorporate business insights. In this process, the new generation of tech-savvy business school graduates has a pivotal role to play. In business schools, we teach how to frame a business problem correctly and ask the important and critical question. We have accumulated a very solid and comprehensive knowledge database about the behaviour of consumers and decision makers, financial systems, and market dynamics. Such understanding should play a guiding role in exercising analytics, and needs to be incorporated into the analytics process. In this way, the full potential of data can be unlocked. Neglecting this can lead to undesirable consequences. I will elaborate my argument based on a couple of research projects we have completed over the past few years.

Optimising Tmall supermarket

Two years ago, The Decision Analytics Group worked with Alibaba on a recommendation problem for their Tmall Supermarket. Different from its traditional business, Tmall Supermarket runs a direct selling business model: Alibaba buys and owns the inventory and sells to its customers. To increase customer stickiness, Tmall Supermarket runs a channel called " 今日瘋搶 (Today's Best Deals)," which is positioned to sell a selection of the most popular products from the tens of thousands of products on offer. A simpleminded approach for selecting the products would be to take those with the highest sales. A more sophisticated approach is to use the many established recommendation algorithms in machine learning such as collaborative filtering. Previously, Tmall Supermarket used a customised approach that combined some business rules such as "A and B cannot be sold together" as well as machine learning algorithms to select products. Typical recommendation algorithms rank items based on their relevance and recommend to customers the items with highest relevance.

We examined the problem and the current practice and identified some drawbacks. First, the objective of the problem was not clearly defined: Was it to maximise the clicks, the revenue, or the profit? Maximising total relevance may not optimise any of these, though possibly positive related. Second, the previous approach ignored the mutual influence of products contained in the channel. For example, when two competing products were offered, their sales would be diminished by each other. Thirdly, customers' substitution behaviour, switching to another product when one's most preferred product is not available, has long been established as a trait in business research but was also ignored in the previous approach.

Boosting performance

In view of these drawbacks, we proposed a new approach that clearly reflected Tmall's objectives and captured the customer behaviour and cross-product influence. In particular, we made use of the preferenceranking customer choice model, which is well researched in marketing literature. Marketing research shows that customers often follow a twostep process in their shopping process. In the face of overwhelming choices, they will first quickly scan the products and boil down to a smaller group, called the consideration set, and then deliberate more carefully on this narrowed selection to decide what to buy. This customer choice model allows us to capture customer substitutions for any given selection of products. The analytics model we built and the solution process are both simplified due to the incorporation of consumer behaviour knowledge.

The outcome of our business insightdriven effort was impressive. The implementation of our approach on Tmall Supermarket brought significant improvements: a 7.4% improvement on the conversion rate (the proportion of customers who purchased among those who clicked the products), and a 16.9% improvement on the amount of goods bought.

There are better revenue models than auction

Since the inception of the internet economy, advertising has been the most important revenue model. In particular, despite all the controversies about user privacy, the internet allows advertisers to target their customers effectively. For each search keyword or each user impression, there are potentially many interested advertisers, and a classical way of deciding the ad resource allocation is auction; simply speaking, whoever brings the highest payoff to the website wins. But event-based auction has a few wellknown drawbacks. For example, the first time you are exposed to an ad, suppose you are aware of it, there is a higher chance you may click it. If you have clicked it, the chance you will click it again is extremely low. This applies to a targeted group as well. Therefore, the number of clicks you will receive from a targeted group for a particular ad first increases sharply, then increases slowly with the number of times you display the ad. In other words, the number of clicks is an S-curved function of the number of impressions.

In marketing theory, this diminishing marginal effect is termed "advertising wearout effect." Such dynamic effects are hard to capture by perevent based biddings, although the industry has proposed various measures to correct this; for example, dynamic updating of the clickthrough rate (CTR), which measures proportions of individuals who see an online ad and subsequently click on it. Another major drawback associated with the auction mechanism is that it may induce fraudulent behaviour. For example, advertisers may intentionally set very low prices at the beginning of the day and wait to win the bids until the later part of day when the budgets of other advertisers have been used up. To prevent such behaviour, again, the advertising firm has to introduce some smoothing-out scheme into the supposedly event-based auctions.

Advertising wears out over time

In view of the drawbacks of the current selling mechanism, we proposed to plan the advertising delivery by a globally optimal allocation of ad resources. We no longer relied simply on eventbased auctions, which by nature are short-sighted and cannot produce a coordinated strategy. Instead, we made use of existing advertising theory, namely, the characteristic S-curved function of ad clicks. To get the number of clicks, one may feel tempted to simply take the product of CTR and the number of impressions. This, however, would be wrong, since the CTR is changing dynamically. In other words, it is the insight that advertising will wear out over time that has inspired us to think about a planning approach, which then further creates the need to make a prediction of the number of clicks for a given number of impressions. In end effect we have produced an analytics task that is new to the community.

We developed a prediction model, and then based on that, an optimisation model to allocate the ad resources. We tested our method at a leading world-wide advertising platform and found an improvement of over 10% to revenue, a huge boost to profitability.

Ask business-savvy questions

Our two research stories above confirm a rather ubiquitous observation shared by many business executives. Current industry practice into customer behaviour may not be answering the right question, even though it may be technically efficient. As we often say, it is more important to do the right thing than to do things right. Business school education teaches a holistic and fundamental view of business problems. It equips our graduates with the ability to ask business-savvy questions, more relevant to business strategies, long-term goals or shortterm objectives. When one questions, "Why are we doing this?" it may lead to different perspectives and fresh formulation of analytics problems.

Incorporating business insights like the S-curved customer response function will simplify the data analysis job and reduce the need for big data. In fact, unlocking the full value of analytics requires talents that know both analytics and have a sound grounding in business knowledge. As of today, engineering-trained talents learn business insights on the job. But the chances are that many take the status-quo for granted and do not ask the deeper question as to why things work in the current way. Many business school graduates, on the other hand, often lack the technical skills that allow them to embark on the analytics journey in the first place.

The new generation will win by doing analytics with insight

Analytics is becoming increasingly important to business success. To make the right strategy, design the right product, and deliver the right service all analytics need to be empowered by insight. After all, analytics is a tool that helps make more informed business decisions. Right now, business schools and business graduates are punching below their weight. They are not playing the important role that should be theirs. With more analytics courses introduced into our curriculum, our new generation of graduates will be better equipped to make a profound change to the current analytics landscape. Analytics should go beyond operational and technical levels. Analytics should be guided by business strategy and support the execution of strategy. The new generation of analytical competitors will win by doing the right analytics. Business insights will remain central to this process.

Let us be ready to make our unique contribution.