Sign up

How bottom-up feature analysis kills conjoint

July 31, 2018

What makes customers decide to buy one particular product over another? This simple question has many complex answers – from price and trendiness to competitor offers, store location, mood, and a wide range of other factors.

In other words, every purchase is a tradeoff. For some customers, trendiness may be more important than cost, while for others, fast delivery may matter more than quality. To understand how to optimize their offerings for maximum revenue, retailers have traditionally turned to conjoint analysis, a technique developed in the 1970s to quantify the tradeoffs each customer is willing to make.

Conjoint analysis breaks down each product into attributes – size, fabric, color and so on – and quantify the price a customer is willing to pay for various combinations of those attributes. Problem is, conjoint analysis only works as a post-hoc technique – it’s extremely poor at making predictions.

That’s why we at Evo developed conjoint’s successor – an analysis technique with high accuracy and tremendous power to predict purchases.

Conjoint analysis is limited by its own methods

As detailed as conjoint analysis is, it’s inherently a top-down technique: it starts with a store’s sales figures, then works backward to quantify how various product features contributed to those sales patterns.

For example, if a store sold 50 units of a white long-sleeved cotton blouse, conjoint analysis would attempt to determine how much customers were willing to pay for white products, for cotton products and for long-sleeved products – then integrate those totals into a conjoint analysis of how much customers are willing to pay for a product that combines all those attributes.

This top-down conjoint analysis would then feed into predictions of expected sales for the next period, which would shape pricing and promotion strategies, as well as inventory orders for similar products.

Conjoint analysis comes with two significant problems. First, it only works as a post-hoc technique, using customer surveys to attempt to pinpoint the product features that contributed to each purchase. And second – because of this subjectivity – conjoint analysis is fairly inaccurate.

What’s more, conjoint analysis often fails to account for the factors that most influence a purchase. For example, if a celebrity is photographed wearing a long-sleeved white cotton blouse, that event may spark an uptick in purchases, completely independently of the popularity of individual product features.

For all these reasons, we at Evo set out to develop an enhanced prediction model.

Evo’s bottom-up analysis accounts for many additional factors

At Evo, we believe that “everything is a signal.” In other words, instead of breaking an item down into simplistic features like size and color, our machine learning model tags each item according to its seasonality, competitor equivalents, region, style niche, and dozens of other attributes.

Our model then analyzes how much each of those attributes contributes to sales, and generates accurate forecasts for sales at various price points over the coming period.

Our model learns from the bottom up. It starts from scratch, with no preconceived notions about what constitutes an item attribute. It simply looks for patterns in the data, discovering subtle correlations that might have escaped even the most experienced retailers. Then it uses those correlations to create a brand-new set of predictions about the optimal price for each item at each location, during each sales period.

This approach to analysis enables us to deliver precise predictions about optimal pricing and promotions – without the need for surveys. All we need is the retailer’s data on product attributes and sales – we do the rest. And our model continually learns from new data as well as retailer feedback, which means its predictions grow even more accurate over time.

Now that this powerful machine learning model has proven its accuracy, traditional conjoint analysis looks like a dinosaur by comparison – it’s slow, costly, inaccurate, and poor at adapting to shifting market conditions. By treating “everything as a signal,” our new model carries price prediction forward into a new era, in which optimization comes at the click of a mouse.

About the author

Ben Thomas is the Chief Writer and Brand Strategist at Evo, with a core focus on emerging technologies, Big Data, and the Internet of Things (IoT).

He loves to engage audiences about the frontiers of science, culture and technology — and the ways these all come together.

Hey! Was this page helpful?

We’re always looking to make our docs better, please let us know if you have any
suggestions or advice about what’s working and what’s not!

Send Feedback