A Next-Product-To-Buy (or NPTB) model is designed to help retailers and marketers improve the effectiveness of cross-selling product recommendations by predicting the product each customer would be most likely to buy next.
Although they’re common in the retail sector, they’re also widely used in banking, where they are used to target offers at customers with greater accuracy. They’ve even been used in content marketing for recommending TV shows and videos that users may want to watch next.
Cross-selling is a sales technique that originated in the traditional retail sector and quickly moved into e-commerce. The concept is simple: when someone has put a given product in their basket, you recommend other products that would go with it (i.e. the batteries that aren’t included, the case, or the screen protector).
The customer gets the benefit of getting all the accessories they may require for their product, and the retailer turns a single-product purchase into a multi-product purchase, helping boost their basket size and Average Order Value (AOV).
By contrast, up-selling offers a larger or more expensive version of a product type to a customer to encourage them to spend more than they otherwise would have done. Sometimes this might be a more premium version of a product, at others it might be a size upgrade. The “super-sizing” sales model used by fast food outlets, offering you the opportunity to “go large” is a great example of this, but both techniques are seen everywhere if you look for them.
In the pre-Internet days, these cross-selling recommendations would have been made by salespeople in shops (remember those?) When you bought a given item, the salesperson would think about the other items you might want, or didn’t know you wanted, and recommend them. These days, where a rapidly growing percentage of products are sold online, those recommendations are generated by algorithms used in product recommendation engines.
On the product page, before the system knows much about you, the algorithms will look at the product you’re viewing and promote products to cross-sell, perhaps as “You might also like” or “Customers who bought this also bought”.
When you add the item to your basket, you may see a more personalised recommendation in either the add-to-cart overlay or in a set of product recommendations shown on the basket itself. Similarly, after you’ve purchased, it’s common to continue to be targeted via marketers using both email and display ads. You might even receive a sales call.
There are a number of reasons why increasing cross-selling should be viewed as one of your ecommerce team’s key objectives. Firstly, customer acquisition is a costly business, and many companies don’t recover the costs of acquisition until after the second order is made. Increasing the number of items purchased, and the AOV, helps recover these costs more quickly.
Secondly, it’s much easier to sell something to a customer who’s already in the purchase process, or even who has just purchased, than it is to sell to a “cold” sales lead. Even in the world of telephone sales, research has shown that you’ll usually make more by cross-selling than you will in making outbound low conversion rate cold calls.
From a customer psychology perspective, selling customers a wider range of products is thought to improve not just the “share of wallet” but also the “share of mind”, and is thought to add “costs” of switching, helps improve the relationship with your brand, and improves customer retention and reduces your reliance upon acquisition for so much of your sales revenue.
If the customers are on your site buying, or if they’ve just bought something, you should be aiming to sell them a bit more. The other key benefit is that the extra data helps you build a better picture of their buying behaviour and tastes so you can improve your targeting.
If your customers typically place few orders with you per year, or if you sell products that have a long lifespan, such as electronics, the importance of cross-selling is even more critical. You may not see them very often, so you need to capitalise when you can.
I suspect there’s a degree of “banner blindness” around some product recommendations these days, much in the way that website users used to subliminally ignore banner ads, before we all just blocked them with Chrome plugins. However, done right, serving recommendations from next-product-to-buy models should improve the user experience by recommending things the customer might need or might want.
They’re quite benefits-driven as recommendations go, so they do add value, I think, especially if they’re so accurate that they take compatibility into account. If you’re using the NPTB approach to target a customer directly, perhaps by email or even telephone, then you can make sure you’re promoting the right product to the right customer at the right time.
At the most basic level, an NPTB model looks at the products a customer has purchased and then predicts which products they will buy next so the products can be targeted at the customer. The models fall into two main types: Collaborative Filtering models and Acquisition Pattern Analysis models.
Acquisition Pattern Analysis models aim to predict the next step for a customer. By examining the behaviour of other customers, what do other customers who purchase a given product, or product type, buy next. Basically, it looks at their sequence of product or service acquisitions. The products may not necessarily be purchased at the same time, or particularly close together.
One great example of this, which I recently experienced myself, followed the purchase of a new phone. I didn’t buy a case at the time and then subsequently got served recommendations for phone cases in the days that followed. The really clever thing I also observed was that some retailers who targeted me weren’t connected to the site I purchased my phone from.
They’d taken advantage of a feature of Adwords that allows advertisers to target ads at customers who have visited pages on other advertisers’ websites. The timing was perfect and I purchased a phone case as a direct result. The ads were useful, rather than invasive, so this worked for me.
While the Acquisition Pattern Analysis models look at the sequence of product acquisitions, Collaborative Filtering looks at the associations of different products (or product types) across purchases from all customers to identify other items that are purchased with the one being purchased or purchased recently.
This commonly used algorithm has been made popular by sites such as Amazon and Netflix and is fairly ubiquitous on e-commerce sites around the world, thanks in part to its relatively straightforward ease of implementation.
One of the key studies on NPTB was written by Knott, Hayes and Neslin. It showed that the model generated stronger response rates, more revenue per responder, and more revenue per customer than other common methods, such as heuristics. They reckon the NPTB model doesn’t just identify customers who will buy the product anyway - it identifies the ones who want the product and are likely to respond to marketing on it. That’s of particular use in direct mail, where you increase your profits not just by mailing to those who will respond, but by not mailing those who won’t…
I suspect that quite a lot of major retailers use NPTB models, but not many disclose that they do. There’s been some published research on electronic appliances, which has shown that customers have a tendency to acquire them in a common order. There have also been some studies on how to target the customers using direct mail and other marketing techniques, but not that much has been published from the ecommerce sector. The related model family of Next-Best-Offer models have been studied in the banking and finance sector. Logically, you could see the same approach working with content or online courses.
Next-Product-To-Buy models aren’t without their drawbacks. One important thing to remember with customer behaviour is that the data you’re observing only shows the customer’s relationship with your business. Most of us aren’t loyal to individual online retailers and commonly shop around, so if customers don’t buy a cross-sell product from you, it doesn’t mean your algorithm was wrong, it may just mean they bought it from someone else.
This also means there’s a finite window for cross-selling to be effective. If you’ve just bought a new phone or laptop, you’ll probably buy a new case for it within the first week or two, after that you’re probably no longer in the market for one. You can therefore tone down the email marketing and cut back the bids on ads targeted at these customers once they go past your typical purchase window.
There’s also a risk that NPTB models could over-recommend popular products which reduces the opportunity to sell other lines, perhaps with a higher margin. However, this can be avoided quite easily with the right approach.
Another useful way to get value from NPTB outputs is in educating customers. If your data suggests there’s a high probability that a given customer would be a likely user of a particular product type, and they haven’t purchased one from you yet, it’s worth trying to educate them about the benefits that using the product might bring to them. Assuming they’ve not purchased the same product type from a rival, you might both provide them with useful and carefully targeted content and generate an extra sale which helps strengthen your relationship.
Traditional NPTB models are designed to predict only what a customer will purchase. They often look at the product category level, rather than the product or product variant level, which improves their accuracy a bit. For example, if you’ve just bought an iPhone XS, you’ll be a good candidate for a matching case, but the exact style and colour of case you’ll choose is going to be much harder to predict.
As you might imagine, the practical application of the Next-Product-To-Buy methodology also benefits from knowing when a customer is going to purchase a particular product type. This allows you to ensure you market to them within the right time frame, which helps improve return on marketing investment by making ads more timely and relevant, and therefore, more likely to convert.
The Prinzie and Van del Poel study is one of the most interesting ones when it comes to predict the when aspect. They studied purchases of electrical appliances at a major Belgian retailer and were able to predict both the product type customers were most likely to buy and when they’d be most likely to buy it.
The most effective features in NPTB models may depend on the business, but they’re likely to fall into three main groups: the product type or need, the order in which the product was purchased, and the time that has elapsed since the purchase of the previous item.
I think it goes without saying that the crucial factor above everything else is that the customer needs to “own” at least one product before you can apply it (unless you’re doing this at the basket level).
|Type||Rather than using the product or product variant, consider using the overall product type (i.e. washing machine, iPhone XS case) as this reduces granularity. Prinzie and Van den Poel (2007) broadened this to look at "Need", so a washing machine was covered under "Cleaning clothes and dishes" along with tumble dryers and dish washers.|
|Order||The order in which products were purchased is critical. If you're decorating a house, it's probable that you'll buy a number of decorating related products before you move onto tidying up your garden.|
|Duration||The amount of time that has elapsed since a purchase is going to have a big impact on the likelihood to buy a related product. If you've purchased a new laptop computer, you'd be likely to buy a case or sleeve for it fairly soon to keep it safe, and the probability of you buying this will drop with time.|
The Prinzie and Van den Poel study includes some extremely useful information on the features they engineered for use in their models. Here’s a summary of some of them.
|Monetary||Total amount spent on all categories divided by length of relationship|
|Total amount spent on a specific category divided by length of relationship|
|Maximum amount spent on a single purchase|
|Minimum amount spent on a single purchase|
|Average amount spent on each purchase|
|Variety||Total number of products bought|
|Total number of products bought by category|
|Total number of product types bought|
|Average number of product types bought|
|Average number of product types bought|
|Demographic||The degree of urbanisation|
|The gender of the customer|
|The customer's dwelling type (i.e. house or flat)|
|Brand loyalty||Number of purchases from the given brand|
|Number of purchases from the top brands|
|Total number of brands purchased|
|Number of brands divided by number of products|
|Price||Total number of products purchased in each price quartile|
|Number of products purchased during promotions|
|Percentage of products purchased on promotion|
|Number of products on "lowest price guarantee"|
|Returns||Number of products returned|
|Percentage of products returned|
Quite a few studies have looked at NPTB and a number of different statistical techniques have been used. Back in 2002, logistic regression, multinomial regression, discriminant analysis and neural networks had all been used and researchers found them to be similar in overall accuracy.
This year, Utku and Akcayol compared deep learning techniques for NPTB against Random Forest, Autoregressive Integrated Moving Average (ARIMA), Convolutional Neural Networks (CNN), and Multilayer Perceptron (MLP) models to see how more modern approaches worked on this as a time series problem.
They found that a deep learning approach beat the MSE of a Random Forest model by 15.3%, beat ARIMA by 13.4%, beat a CNN by 11.2% and beat an MLP model by 11.2%.
Yes. I’ve used a category-based RFM variables to achieve similar results, and the underlying features are similar to those used in the Prinzie and Poel model.
NPTB can easily be applied to direct mail marketing and there have been several studies in this field. While it’s a field that’s dying out these days, if you want to use this, you’d create a group of target product categories (i.e. phone accessories, laptop accessories, desktop accessories) and then create a piece of direct mail focusing on each product category.
You would then use the model to target those customers who had a high probability of being a good fit for the product category and were within the typical period in which customers would usually purchase from the category. This can help improve the relevance of your direct mail marketing and reduces the marketing costs, since you’re not using the unscientific “spray and pay” approach.
Anand, S. Hughes, J. Bell, D. and Patrick, A. (1997) - “Tackling the cross-sales problem using data mining,” in Lu, H., Motoda, H. and Liu, H. (eds.) Proceedings of the First Pacific-Asia Conference on Knowledge Discovery and Data Mining, 331-343.
Utku, A., and M. A. Akcayol (2020) - “Deep Learning Based Prediction Model for the Next Purchase,” Advances in Electrical and Computer Engineering, vol.20, no.2, pp.35-44, 2020, doi:10.4316/AECE.2020.02005
Kamakura, W. A. (2008) - Cross-Selling: Offering the right product to the right customer at the right time. Journal of Relationship Marketing, 6(3-4), 41–58. d
Knott, A., Hayes, A., and SA Neslin (2002) - “Next-product-to-buy models for cross-selling applications,” Journal of Interactive Marketing, Volume 16, Issue 3, 2002, Pages 59-75.
Li, S., Sun, B., Wilcox, R. T. (2005) - “Cross-selling sequentially ordered products: An application to consumer banking,” Journal of Marketing Research, 42, 233-39.
Prinzie, A., & Van den Poel, D. (2007) - Predicting home-appliance acquisition sequences: Markov/Markov for Discrimination and survival analysis for modeling sequential information in NPTB models. Decision Support Systems, 44(1), 28–45. doi:10.1016/j.dss.2007.02.008
Matt Clarke, Wednesday, March 03, 2021