Predictive analytics offers huge opportunities for every business. At the FT we have run experiments with propensity models that can help predict how likely someone is to purchase a subscription at a given moment in time. This allows us to stage appropriate interventions (e.g. sending a personalised offer to someone we believe is close to making a purchase).
With FT Strategies, we have been taking these predictive modelling capabilities to businesses in sectors ranging from insurance and retail, to arts and culture. This blog shares some of the lessons we’ve learnt at the FT, from the wider industry, and in our work with clients.
What is predictive modelling and why does it matter for your business?
One of Amazon’s many impressive feats has been shifting the expectation of consumers when it comes to delivery times. A huge range of products available for next day delivery, while Prime Now users can expect household basics brought to their door within hours. Amazon’s feat isn’t just down to an impressive network of warehouses and last-mile; a large part of the retailer’s success is down to use of data and predictive modelling.
In 2013, Amazon filed a patent for “Anticipatory Package Shipping” – an idea for using predicted purchases to get orders as close as possible to delivery addresses before the point of shipping. While many of us might feel that our shopping habits are impulsive, it turns out that we’re a lot more predictable than we might think – with Amazon able to draw on many data points (including past purchases, product searches, clicks wish lists, shopping cart and time spent on items) across a vast pool of customers.
Jenny Freshwater, a software director in Amazon's Supply Chain Optimization Technologies Group, told NPR about how her team creates machine learning models that use this data to make predictions. "It goes beyond just being able to forecast that we need a hundred blouses”, she said. “We need to be able to determine how many we expect our customers to buy across the sizes, and the colours. And then where do we actually put the product so that our customers can get it when they click 'buy.'"
Amazon’s predictive shipping models ultimately allow the company to prioritise stocking certain items in local distribution centres. This allows Amazon to reliably reduce delivery times, increase conversions, and even create entirely new products (e.g. Amazon Prime). While Amazon led the way with introducing predictive modelling to retail, businesses in every sector are starting to realise what the predictive analytics opportunity means for them.
Why are we seeing more examples of predictive modelling in business?
Predictive modelling has risen up the agenda for businesses in recent years thanks to two key factors.
Firstly, as businesses have digitised there is an increase in data available to make predictions. For example, when you watch terrestrial TV, very little data can be captured. In contrast, when you use a service Netflix or iPlayer you are typically logged in to an account, and it's possible to track similar metrics to the Amazon examples mentioned above (clicks, time spent browsing a particular category, shows abandoned, etc). This additional data allows services to predict your future viewing habits, and create services such as recommendation engines.
Secondly, the underlying technology is becoming more accessible. Most of the predictive modelling happening today is possible thanks to machine learning – a technique that makes predictive modelling faster, more accurate and can help process unstructured data (sets such as text and image). Only a few years ago, machine learning was out of reach of businesses unless they made significant investments in data science capabilities and tech infrastructure. However, machine learning has become increasingly accessible and affordable – both via services such as Google Cloud and Microsoft’s Azure AI, and a growing open source community offering increased tools and knowledge sharing.
How should your business get started with predictive modelling?
When investing in predictive modelling, it’s vital to have an understanding of what business impact you are trying to achieve and how you will validate whether you are having the desired impact. Before you build anything, you should design what will be changed for the user and decide how you will test it.
Let’s take a practical example to explore how you might do this: the FT’s first propensity model. When we created this model, the FT had a metered model that gave readers 10 free articles per month. At the time, we were attracting large numbers of registered users each week, which provided a rich source of B2C subs leads. Our goal was to increase the conversion rate of these users, and we had a hunch that predictive modelling might help.
Before technical work started, we knew it was important to break down our predictive model into the following elements:
- What are we trying to predict? E.g. We want to predict the propensity to subscribe of registered users
- What will you change for the user based on this subscription? E.g. We will send users with a high propensity to subscribe a targeted offer onsite and by email
- How will we test it works? E.g. We will measure the proportion of users subscribing through the campaign
We believe that these three questions are a crucial starting point for any predictive modelling project. In particular, having a sense of how you might measure a model’s performance is critical not just for judging the immediate success of the model – it can also help keep track of changing performance over time.
With the FT model, users subscribing through the campaign accounted for 15-20% of total subscriptions. However, as the months passed we noticed that the number of people subscribing via the promotion was decreasing. We believe this was because the model target started to become extinct over time. While we still had a pool of registered users, we were at risk of overfishing so stopped the model.
This points to the importance of seeing predictive modelling as an evolving capability – as you learn more about your users and your models, you will need to take an agile approach and be ready to move on to new opportunities.
How FT Strategies can help
Over the past years, FT Strategies has helped a number of businesses transform their data and prediction capabilities. The opportunities are clear: from driving increased conversion and creating new products, to simply making your supply chain more efficient. And we are here to help you navigate a space that can feel as complex as it is rewarding.
We’d be delighted to hear from you if you are interested in making an impact with predictive modelling. We have worked with all sizes and shapes of business in various formats – from 2-week design sprints through to longer-term engagements. And as ever, we’re always happy to start with a chat over coffee.