A Data Science Roadmap for your eCommerce Business
Data Science is a big lever in eCommerce, and it’s not just the huge enterprises that can take advantage. It’s important though to choose the right data science projects for the size of your eCommerce store and the types of product you’re selling.
Below I’ve outlined a customised data science roadmap for 5 different types of store for you to use.
The Growing Store
As your store starts to grow, you’ll have more incoming traffic and active users to analyse, generating more data for your data science projects, and more opportunities to increase your profit margin.
Homepage Personalisation
Your site homepage is your shop window, and is a big opportunity to inspire your customer towards their next purchase.
Every customer is different, has specific interests and a unique history of browsing your site. With this data, it makes perfect sense to personalise your homepage to each user, adjusting which content is included (products, offers, and articles) as well as the order in which it is shown.
Every personalisation platform worth its salt will have an offering in this area.
Offer Personalisation
Nearly all stores will introduce discounting offers, whether for specific products, first purchase, or free shipping. Offers are no different from prices, and therefore each user will have a certain “elasticity” to respond to the offer. As such, just as with the homepage, it makes sense to experiment with your offers, using a machine learning algorithm to determine for each visitor the size of the discount, the applicable products or categories, and when to display the offer.
Churn Prediction
As your customer base grows, you’ll also start to see customers churn. With predictive modelling, it is often possible to predict those customers most likely to churn, and then to make an intervention, for example by sending an email with a personalised offer.
The Seasonal Store
Whether you’re selling sunglasses on Shopify or mittens on Magento, it’s likely that there is seasonality in the demand for your product. This can be a real challenge for eCommerce businesses, but don’t worry, data science is here to help!
Forecasting
Forecasting is important for all businesses in order to manage their cash flow and marketing investments, but seasonality can make this exercise tricky. Fortunately, the data science field of time series modelling is focussed on exactly this type of problem. By examining historical data versus the calendar, it can produce extraordinarily accurate predictions about future sales, enabling you to plan your budget accordingly.
Budget Optimization
With a seasonal product, you have to choose carefully both when to spend your budget and how to spend it. If you have a historical record of sufficiently diverse data about how you have invested across different channels over time, you can simulate different budget distribution strategies and predict the resultant revenue outcome.
Email Optimization
If there are particular points in the year where you need to capitalise on demand, email marketing is almost certainly part of your arsenal for triggering your users to return to your site after a lengthy hiatus. By using machine learning, you can experiment with different email content for a sample of your different audiences, and then roll out the most successful campaigns to your broader customer base when it’s the right time to strike.
The Luxury Product Store
Luxury products are usually expensive and are associated with a strong brand. The following Data Science techniques can help with the careful handling required.
Multi-touch Attribution Models
Luxury products are rarely impulse buys — customers need nurturing over time. In these cases, where a purchase decision can span over several weeks or months, it’s important to understand which marketing actions contributed to the sale. The answer? Multi-touch attribution models.
With simple attribution models, the credit for a sale will be given to either the most recent or the first marketing campaign that brought them to the site (AKA “first touch” or “last touch”). With multi-touch, every marketing action along the way is given a share of the credit — whether that’s a facebook campaign or an abandoned cart email — and machine learning models can be used to work out what weight to give to each.
Brand NLP
Your linguistic style is a big part of maintaining a strong, distinctive, and original brand voice. But tone-of-voice can be subjective and hard to measure, right? With help from NLP (“Natural Language Processing”) techniques and tools like spaCy, you can build classification models to analyse an email or a product description, and measure how similar it is to a corpus of text from your brand. It can even suggest word substitutions or rephrasing!
Fraud Prediction
With an expensive product, fraud can be harmful to the bottom line. And with fraud a growing concern amongst all stores, it needs to be tackled head-on.
Machine learnings models have had great success with fraud detection across multiple domains, dating back to early email spam classifier models, and eCommerce is no different. By training a model on behavioural data for a user, it will be able to predict the likelihood that they are a fraudster before you fulfil their order.
The Fledgling Store
Even if you’re just getting started, and data science seems a world away, there are still a few important analytical steps you can take to help you get to the next level.
Conversion Funnels
The first step with looking at your store data, is to build your conversion funnel. Your eCommerce platform likely offers this out of the box, otherwise implement an event tracking analytics tool like Amplitude. Although not advanced data science, this allows you to examine which pages and products on your store convert the best, and which traffic is driving those conversions. With that in mind, you can act to either drive more of the best traffic to your best performing pages, or you can try to optimize those pages that are not performing so well.
Lifetime Value
When you’ve got the hang of looking at conversions, the next consideration is customer lifetime value (“CLV” or “LTV”). Simply put, this is an estimation of how much you’re likely to make from an individual customer. By combining their transition history with their other attributes across your customer base over time, you can build up a model to predict this value for each user. The big advantage, is that you can start to bid more on your acquisition channels for higher LTV customers, knowing that the profits will be returned to you down the line.
Campaign Optimization
You might not know it, but if you’re running ads on facebook or another ad platform, it’s likely you’re already taking advantage of data science by using “lookalike” audiences. By firing your conversion pixel, you give facebook data about which types of user are likely to make good customers for your store. You can then imagine that they use an algorithm to cluster their users, and identify those that share similar traits to your highest value customers.
The Megastore
If you’re running a high volume store, with thousands of products, data science is a powerful sidekick to help you guide your users through all that you have to offer.
Product Recommendations
Amazon has notoriously been hugely successful with recommending, cross-selling, and upselling similar products to their customers. Underpinning this is their product recommendation engine. Comparable technology is widely available through personalisation platforms, and Shopify even offers a free version on their platform.
Recommendation engines are typically built using a technique called Collaborative Filtering, which can produce impressive results.
Merchandising Trend Predictions
If you’re in the fashion industry, selecting the right next product line can make or break your business. Analysing images using deep neural networks has been a breakthrough field in machine learning in the last decade, and one application of this is in detecting trends in clothes products. By ingesting millions of images from online stores and social networks, these algorithms can determine trends in taste across different geographies.
Return Prediction
Offering free returns is considered standard customer service in eCommerce, but it’s expensive for retailers and for the environment. In many cases, product orders likely to be returned can be identified before checkout, even with simple heuristics like flagging baskets with multiple sizes of the same item. Store tech can trigger sizing helper tools when the algorithms detect that the return likelihood is high — saving hassle for the customer and cash for the company.
Conclusion
Whatever type of store you’re running, make sure to take advantage of the possibilities that data science can bring.
Want a roadmap for your specific store or help implementing the projects above? Tell me about it in the comments below.
This article was written by Timothy Daniell, founder of Permutable, a data science consultancy in Europe working on custom projects for eCommerce businesses.
