Skip to main content
Placeholder

Learnings from AI applied to consumer and shopper strategy

Welcome to our (nearly) spring newsletter!
Written by: Foresight Associates
Date: February 18, 2021

We are excited to announce that we will be hosting a session at the Quirks Global Virtual Event for Marketing Research and Insights Professionals. Our topic will be a case study on piloting the use of AI in quantitative consumer and shopper strategy work. We are co-presenting with one of our clients from The Coca-Cola Company and with our partner firm Amass.

I hope you enjoy reading a few of our learnings in this issue, and will join our webinar where we share the full presentation: “Using AI to develop an early warning system for brand performance.”

Click HERE to register (free for end clients):

Room 4 | 11:30 am-12:00 pm | Wednesday, February 24, 2021 – main session presenters live for Q&A
Room 4 | 4:40 am-5:10 am | Thursday, February 25, 2021 – replay for APAC community

Please feel free to reach out to us with any questions, or if you feel this work could be relevant to your business!

Vittorio Raimondi

Managing Director, Foresight Associates LLC

Using AI to develop an early warning system for brand performance

Background

Unless you have been living under a rock, you have likely heard about the coming wave of artificial intelligence. AI has graced the covers of the Harvard Business Review, The Economist, Forbes, and more in recent years and it is positioned as a revolutionary change in the way we will use data for decision-making.

Given our expertise in analytics and decision-making at Foresight, we’ve been keeping an eye on the AI trend for some time. We know that AI has a clear use case in marketing – but so far it is extremely tactical. Targeted content, customer service bots, in depth forecasting and allocation, recommendation engines, dynamic pricing engines, voice recognition and even promotional content creation are examples.

Yet there never seemed to be a clear applicability to our strategy work – despite the fact that we are nearly 100% data-driven. And so we (and we assume our clients) were left wondering: is this AI business all hype or the real deal?

Top Learnings

We were surprised to find that AI can be easily applied to the types of datasets we work with in strategic modeling projects. In fact, one of our learnings was simply that AI is not an entirely separate discipline from traditional statistics – rather it builds on those methods. With incredible efficiency, the model was able to develop accurate predictions for the beverage universe based on testing a few common machine learning algorithms and methods.

We see a clear use case for our client in the form of an early warning system dashboard that can improve the way we think about Key Performance Indicators. This use case will only strengthen as we move from a pilot with a constrained time series to an expanding database in real time.

Test Case

We set out to test a use case for AI in the realm of strategy making, re-visiting a project we had worked on with The Coca-Cola global team to identify Key Performance Indicators (KPIs).

The challenge, familiar to many in CPG, is that amongst a dizzying array of metrics that brand managers track – from controllable activities to point of sale and consumer metrics to equity studies and macro factors – it is often not clear what really matters for growth. What should be prioritized and what should be ignored? How do you know when a brand is at an inflection point: will what matters today be what matters tomorrow?

In this pilot, we set out to see if AI could do a more efficient and proactive job of identifying KPIs than us humans.

Here are a few of our top learnings about bringing AI into the realm of consumer and shopper strategy:

1. AI is not as exotic as it sounds

When you imagine AI, what comes to mind? The Terminator? Amazon Alexa? Your Netflix account? In this case, it’s helpful to clarify what we mean: AI is simply “the science of getting computers to act without being explicitly programmed.” The most common use case today is machine learning – which is “the ability to learn without being explicitly programmed.” This includes:

  1. Reinforcement Learning – in which the machine figures out the correct actions to take over time using copious amounts of data and gradually making fewer mistakes over time (for instance, self-driving cars).
  2. Supervised Learning – in which the machine is trained by humans to correctly classify data with labels (for instance, identifying features in images).
  3. Unsupervised Learning – in which the machine does its best to fulfill an objective given certain parameters by identifying patterns without human training (for instance, recommendation algorithms like Netflix).

 

For our purposes, we are unlikely to ever have enough data points to allow for true reinforcement learning. But we can take advantage of Supervised Learning in the form of brand classification, and Unsupervised Learning in the form of clustering and neural networks.

Artificial neural networks are designed to work like our brains: they “learn” and remember to build mental models based on many inputs. The particular technique we used (called Long Short-Term Memory) iterates over time-based events, “learning” how various events in the model are connected, and performing millions of calculations to provide prediction outputs.

2. AI can move our KPIs from backward to forward-looking

When we ran our model over about 3 years of data across the various sources, the AI was able to provide us a unique set of predictors for category share gain for each brand in the beverage universe. This was fundamentally treated as a classification problem: how likely is the brand to gain (or lose) share over a certain forward-looking time horizon, and what drives that prediction?
For instance:

  • A Diet/Light Sparkling Brand’s model was most focused around its base price compared to the category, and its current Yearly+ consumer base.
  • An Orange Juice Brand’s model was most sensitive to Distribution and Future Consumption Intent at the Weekly and Daily level
What comes out is a unique model for each brand in the

We think this type of modeling could help organizations move from choosing KPIs based on habit, history and management preference to being fact-based and future-focused. Further, they can be completely customized and continuously evolve as certain factors will pop up over time that weren’t factors before, and others may become less so. This may signal an important inflection point for the brand’s strategy.

3. There is a clear use case for a global organization as an 'early warning system'

How does this come to life for a global team like Coca-Cola? Sitting in Atlanta, they need to track performance across over 40 markets worldwide, identify opportunities and gaps to the company’s growth model, and isolate which teams to talk to when issues arise.

We built a prototype of a dashboard that could use this AI engine to provide a “risk” meter for a brand. Think of it like an early warning system for a hurricane – except the disaster we are predicting is a loss of share. How likely are you to gain/lose share based on the KPIs the model has identified, when is it going to happen, what is the expected share change, and which KPIs are driving that prediction? Further, how are those KPIs becoming more or less important?

Wireframe prototype for using this model

This can provide value not only for one brand but for a collection of brands or markets, with warning alerts sent to the relevant team members when risk of future share loss crosses a certain threshold.

It can also provide predictions and warnings not only for your brand, but for the portfolio in aggregate, as well as for the competitors likely to capture that lost share.

Through the clustering algorithm, we can also identify brands with similar characteristics that may have different KPIs which could become relevant for the brand of interest in the future (or if a new brand appears in the market, what KPIs are most likely to matter).

Once an opportunity is identified, marketing mix model coefficients can be integrated into the interface such that the brand team can explore how to solve the issue based on their controllable marketing activities like media spend, price, and distribution. Our client felt this would allow the team to support more pro-active decision making when facing a gap to plan.

4. The benefits are primarily speed, scalability, and automation

The use of AI does not necessarily mean brand new methods. Many of the algorithms we ended up using in the model – clustering, neural networks, etc. – are tried and true analytic workhorses. But the biggest difference from the status quo is in the promise of a completely automated system, with a data feed that continues to update and learn in real time, integrating sources from across the organization, with hyper efficiency.

As the engine is expanded across markets, categories, and new sources – including factors like regional data and unstructured social data – it will continue to learn and improve, allowing for the insights to incorporate new predictors. Further, the model will be able to continuously test different analytic approaches and evolve its own methodology as time goes on without human intervention.

It is important to note that this model answers a different set of questions from marketing mix models. This is about efficient predictors of share change, across multiple data sources – which may or may not be controllable. It is more analogous to reading the weather forecast. If the clearest signal for your brand’s success is the number of cases of product you have on display in stores, its important to recognize and track that – even as you model all of your controllable initiatives for growth.