Our pioneering approach to Bayesian econometrics

Bayesian2018-05-15T09:21:20+01:00

Introduction

Whilst Bayesian theory is only one of several types of mathematics used by Nate Silver, it reflects a growing awareness that modern technology can be used to harness the power of techniques that were previously considered too hard to put into practice.

Bayesian is of particular interest to marketing analysts in that it enables us to deal with the very real degree of uncertainty that we face in designing models for practical application in the real world. In also enables us to draw conclusions for smaller amounts of data or when the data is hard to come by as well as incorporating prior knowledge into models. The main advantage of Bayesian over traditional (or ‘Frequentist’ approaches, as they are known) is more accurate predictions.

A lot of the content written about Bayesian appears aimed at the academic or mathematician. We’ve therefore compiled our own introduction which can be downloaded from the link below.

DOWNLOAD OUR FREE WHITE PAPER

Why Bayesian matters to marketers

Whilst the 20th Century was dominated by classical (or ‘Frequentist’) statistics the last decade has seen a resurgence of interest in Bayesian approaches to determining probability. This has particularly been the case in marketing, where Bayesian approaches are increasingly being sought across the many areas where modelling can provide a competitive advantage. So what has an approach inspired by a 250 year old theorem got to do with modern, omnichannel marketing?

“Inside every non Bayesian there’s a Bayesian struggling to get out”

Dennis V. Lindley
Decision Theorist

The interest reflects something of a paradigm shift in the way we view probability which is particularly appealing to marketers. Marketers in many sectors today are facing competition from some previously unlikely sources; they are facing greater complexity in the choice of tools in their marketing armoury due to the explosion of digital; and they are facing greater need for accountability. To accommodate these demands, they are looking for a way of making more informed decisions, but in the real-world are often faced with a paucity of data.

In reality, marketers are rarely making decisions in a vacuum and have a wealth of previous knowledge and experience to draw on. A Bayesian approach enables these views on the underlying (and often incomplete) data they are presented with to be taken into account rather than simply relying on counting what is there.

For instance, market researchers relying on traditional statistical approaches often report that predictions are too aggressive and unrealistic. Their model might suggest, for example, that a price of reduction of, say, 10% or 20% would result in an increase in market share that experience tells the marketer is too large.  When uncertainty is taken into consideration, market share predictions become more realistic and tend to agree with current and past experience. Bayesian approaches therefore offer a more elegant solution to an important problem.

If marketing is really about understanding human behaviour, then Bayesian approaches enable us to study the complexities of behaviour in a more realistic fashion than was previously possible.

Our approach to Bayesian

One of the reasons why Bayesian was previously largely overlooked was the sheer complexity in its application in anything but the simplest problem.

Whilst the actual mathematics that goes into a Bayesian model is still pretty challenging, technology now helps some of the pain out of the number-crunching. We have developed our own proprietary Bayesian modelling platform called BayesIQ™ about which you can find out more below.

However, there is more to developing a Bayesian model than sophisticated technology and clever mathematics. The model needs to be grounded in the context of the prior knowledge that exists around the problem in question. This is where our team’s experience of working across many industry sectors plays a key role in working with our client teams to uncover and incorporate the existing insight, and real-world limitations that can help develop a more accurate model.

Far from being a ‘black-box’ solution, we can then carefully adjust our model to see how much reliance is placed on the data versus the prior knowledge. We then work closely with the client team to review the model’s predictions and translate them into clear and actionable marketing insight.

BAYESIQ™ – BAYESIAN MODELLING PLATFORM

BayesIQ™ is the software on which we run our Bayesian models and we believe it to be the only fully Bayesian enabled multiple regression application specifically designed to run marketing mix models.

BayesIQ™ is underpinned by world leading academic input and our extensive experience of market mix modelling. It allows MetaMetrics to run models specifying differing types and strengths of prior knowledge in order to deliver more robust, stable and actionable outputs.

The advantages of the Bayesian environment are widely known but the challenges of running Bayesian marketing mix models have up to now been significant. This is a result of the statistical knowledge required as well as the need to rely on academic open source software that was never designed to handle the demands of market mix modelling and, as a result, is time consuming and cumbersome to use in practice.

BayesIQ™ allows us to seamlessly incorporate external information, logic and common sense into our econometric models which in turn means that we reduce the amount of judgement and intervention required by the analyst. We are also able for the first time to quickly specify, update and iterate Bayesian marketing mix models.

We are able to carefully adjust how much reliance we want to put on the data versus our prior knowledge. Poor quality or sparse data can be heavily constrained by tight priors gained from other markets or our database of normative information. Very rich and robust data may only need weak priors to account for known effects and allow the rest of the dataset to work around the known to estimate what we do not know.

Retail chains or regions can be modelled as belonging to the same distribution or “family”, rather than simply as entirely unconnected separate entities, with their ability to differ from each other controlled by the Bayesian framework.

BayesIQ™ puts this ability into the hands of the marketers for the first time.

Find out what Bayesian can do for you…

GET IN TOUCH