CDO Interview: Product-level predictive analytics at retailer Very
The Very Group Chief Data Officer Steve Pimblett explains how data, insight and most importantly action are changing the face of one of the UK's established retail concerns.
Just after Steve Pimblett joined The Very Group in 2020 as the retailer's first CDO, he had a conversation that gave a clue as to the task at hand.
"I always remember speaking to the managing director who ran Home and Garden. He said ‘Great, you're someone who understands data, now you can tell me what garden item to sell next'," Pimblett recalled. "And I said, ‘Well, that's quite interesting because I thought what you would have asked me is which of our customers have got a garden?'"
Helping to transform the Very Group from essentially four separate product-oriented businesses, each with its own P&L (and bonus structure), to a customer-centric concern that transcends artificial barriers has been Pimblett's mission ever since.
But change doesn't always come easy. Formed from the merger of Littlewoods (founded 1923) and Shop Direct, The Very Group's roots run deep. Older companies, like older people, tend to like what they know, and mergers sometimes create cultural clashes.
Consequently, Pimblett spends a lot of his time addressing cultural barriers, explaining why it makes sense to move from a category-run business to one that sees one customer served by four sectors. He talks a lot about agile, the value of predictive and descriptive analytics and the actions that managers need to take to drive value from data.
This last step is critical. All the data and fancy analytics in the world are so much pie in the sky if they don't lead to action. Insight must be directed to solving real issues, and proposed interventions must be pinned to something tangible.
"What really resonates is when you can tie the insight to the action, a better decision, growing the top line, optimising margin, improving Net Promoter Score. Tie it to a business outcome."
Very big
Headquartered in Liverpool, Very is a substantial business serving 4.4 million customers with 2,000 brands and 250,000 product lines. Divested of bricks and mortar, it is purely app and online, and delivers about 50 million parcels a year.
The company has digitised and applied intelligence to its supply chain and has automated its warehousing facilities.
And it has also moved into the financial side, with flexible payment options including buy-now-pay-later, pay-in-three and click-and-collect. It is now one of the biggest de facto lenders in the UK.
Underpinning all of this, of course, is data.
Data centres of excellence
Pimblett's data team numbers 150 individuals grouped into several centres of excellence. They include platform teams with skills in storing, scaling and deploying data in the cloud; teams focused on business intelligence and dashboards; teams experienced in handling financial data; data science teams; teams with skills in specific platforms; specialists in security, risk and governance; and so on.
These centres of excellence serve the analytics needs of the business verticals. The retail vertical needs to know what stock to buy, availability, what price or promotion; in the financial services vertical expertise in fraud, credit risk analysis and compliance are invaluable; in marketing analysing customer communications and customer experience come into play. And of course there are many areas that cross over. The idea is to avoid duplication.
"We build once and deploy across all the verticals," Pimblett explained.
Product-level predictions
In the early days of big data, retail was very much a poster child. There was even a TV ad featuring a coffee shop that changed product lines in response to the weather forecast. That was largely big-data washing, but, says Pimblett, those predictive capabilities are very much here today.
Take purchasing. The optimum goal in predictive purchasing is to predict what customers are going to want, then buy one too many of everything, so that no customer is ever disappointed and no stock is wastefully stored.
The Very Group has developed specialised machine learning models that can crunch multiple variables, from the macroeconomy, to regional differences in demographics, seasonal trends and weather patterns in order to plan promotions, optimise stocking levels, and pick product lines.
"It goes right through to the dynamics of the product itself. For example, in fast fashion, you've never sold that particular dress before with that particular colour and pattern. How many should you order? That's a different predictive model than estimating how many kettles you're going to sell over the year."
You can't wear a kettle or make tea with a dress. Likewise, there is no one-size-fits all predictive model covering all products. To iterate towards the best answer for each use case, domain or product category, Pimblett's team uses a technique called champion/challenger. This is a way of comparing multiple competing strategies in the live production environment to determine which delivers the best results. The champion is the existing model, while challengers are new approaches that may end up stealing the crown for a particular use case.
"We can try different predictive models even down to the individual product-line level, and see which is the best at predicting that particular product," he said.
The tech stack
The Very Group creates its own machine learning and analytics models in house, frequently using those available from AWS and Teradata as a starting point, and partnering with those companies in their further development. Often, Pimblett's team will start with an "off the shelf" model then refine it iteratively.
"We might be able to get a champion live really quickly because we leverage some out-of-the-box service. But within a few days or weeks, we'll have developed our own IP that can start to beat it."
Some retailers steer clear of AWS, seeing Amazon as a competitor, but The Very Group, which moved its analytics stack from an IBM data centre onto AWS two years ago, is not among them, Pimblett said.
"We don't see it as a negative that Amazon is in retail. We see it as a positive because they've gone through the learning. They understand the industry."
The Very Group is also a major Microsoft customer, particularly Power BI for business intelligence.
The partnership with Teradata goes back a long way, too. Having been a big user of its tools and database on-premises, Very is using the company's VantageCloud and ClearScape AI/ML for much of its analytics, data management and modelling.
"We're starting to find more and more use cases that we want to drive value from with the massive amounts of data that we've got in Teradata," said Pimblett. "We can unlock that data with descriptive analytics and predictive analytics that drives the right actions."
Because analytics is an iterative process, it lends itself perfectly to the MVP approach, he added.
"Start with the action that is tied to the business. That gets sponsorship, and it gets conversation and collaboration going. Then start small, build agile and iterative, realise the value and get that feedback."
On 5th October Computing is hosting the IT Leaders Summit 2023. The ITLS is a two-day event bringing together the most senior and influential voices from IT leaders in the UK at Down Hall in Essex. Register today.