Data – decision critical, not just a record

Data will be considered a decision-maker, an influencer, and an input to our actions – not just a record of what has happened

Data as a decision-maker is a long emerging trend. Today’s challenge of “too much data, too little insight” will be a thing of the past. Data will envelop our everyday experience.

The COVID-19 pandemic has accelerated our understanding of data. Even the most data-averse guzzled graphs, became familiar with simulations, and learned to distinguish predictions from projections. We saw how critical data is to decision-making, but we also came to understand its limitations.

What does this mean for marketing?

The realisation that data has limitations is a wake-up call for marketers. We have become captivated by the collection of data and attempts to build an ever more complete picture of customers. This is a Sisyphean task – or one that can never be completed – given such data dates from the moment it’s created and therefore must constantly be refreshed.

Marketers have been particularly enamoured with machine learning (ML) networks for predictive modelling and collecting sufficient data to feed these ravenous beasts – even if few of us would want to explain how they work. The rather muted role ML in pandemic analytics, however, was a reminder that it’s not the only predictive technique.

By 2032, we predict a new maturity in marketers’ approach to data-based decisioning; one that uses ML selectively but also makes use of computational and statistical alternatives.

Simulations: an expertise-based approach

Predictive ML models are based on correlation. During the training process they find patterns in large volumes of historical data that explain how a desired outcome is achieved. Once enough data is crunched, and a suitable pattern found, the training process is complete, and the pattern or predictive model can be applied to real world data.

The model does little to explain the cause that links input to output, it simply finds correlations. By contrast, simulations use domain expertise to define the predictive model, designing what input will cause what outputs.

In agent-based simulations, for example, domain experts predict how agents will behave and run that behaviour in a variety of contexts – introducing new external factors or changing small elements of the agents’ behaviour – to project probable outcomes. The ‘Washington Post’s’ Covid-19 simulations are a classic example of projecting different outcomes for the spread of the Covid-19 virus. They looked at the impact of social distancing measures and how these could flatten the infection curve.

These simulations used expert knowledge to codify how the agents (humans) would behave in different circumstances. In a socially distanced society, there would be fewer interactions between humans and so fewer opportunities for the virus to spread: the infection curve was therefore flatter. Simulations use less data and processing power, but some also consider them more appropriate for strategic decisioning because they are based on causation and expertise. There is even some suggestion they are a more accurate predictive tool.

Swarm decisioning to predict outcomes

Like simulations, swarm decisioning uses human experience and expertise to predict outcomes. Unanimous AI has pioneered a technique, inspired by the behaviour of honeybees. Small groups of domain experts are asked to make predictions about future events, such as sporting fixtures, election results and the likely impact of product launches and marketing campaigns. Participants are given just 60 seconds to predict the outcome of these events but, even within that short timeframe, they can and are influenced by each other.

The technique relies on domain expertise amplified by algorithms which weight strength of conviction and has remarkably accurate results. The prediction for President Trump’s approval rating at 100 days was accurate to the integer (or whole number). The prediction was made before his inauguration and, at 41%, it was by far the lowest rating ever given to a US President.

Synthetic decisioning for poor data

The availability of high-quality data has always been the greatest constraint on analytics. Synthetic data is an increasingly popular solution to this problem.

Generated from mathematical rules and statistical modelling, synthetic data reflects real-world data but avoids many of its pitfalls, such as availability, quality, cost, and privacy constraints. When thoughtfully created, it can remove the sort of bias that can perpetuate discrimination and include outlier data that might otherwise be missed.

We agree that in the future most data used in ML will be synthetic. In the future, we expect analysts to create their own business metaverse (miniverse?) where they can play with data, simulate the outcomes of different scenarios, try out bold ideas, and test their preferred strategies before implementing them in the real world.

A question of balance

Although marketers will continue to use ML for predictive modelling, we predict a more balanced approach in which other techniques, especially those which take a more strategic and expert-driven view, are also used.

We think expert-driven techniques have a significant advantage over ML predictive modelling in that they use less data and less computer processing power. Although synthetic data will go a long way towards resolving the first problem, by 2032 excitement about ML will be tempered by awareness of the financial and environmental costs.

Going forward, businesses will have a more considered approach to data-driven decisioning – knowing exactly what business value they are seeking and using a portfolio of decisioning techniques to help achieve it. Synthetic data will lift the constraints on analysts and business leaders, allowing them to create virtual versions of their industry and markets, and test strategies before they implement them.

This is the second in the series of three articles derived from WPP’s Thoughtful Data 2032 which follows Data 2030.