How predictive algorithms are transforming retail industry
Very often, rather than a “boiling the ocean” approach, it is prudent to focus analytics on high impact measures that deliver actionable insights into what really matter to the business.
By Ashwin Malik Meshram
Technological innovation is transforming the retail industry in unprecedented ways. But are companies able to capitalize on it? In our experience across multiple organizations from diverse geographies, analytics is still under-leveraged in most companies, and has inadequate focus on what really matters to the business.
In spite of investing heavily in analytics resources, we have seen companies struggling with all sorts of challenges, sometimes pertaining to fundamental business metrics, like who their most valuable customers are, what are they buying, purchase patterns, which products are being recommended to customers, seasonality, demographical and geographical preferences, most profitable suite of products, etc.
And very often, rather than a “boiling the ocean” approach, it is prudent to focus analytics on high impact measures that deliver actionable insights into what really matter to the business.
Measure #1 – Predict the online return propensity using pattern recognition algorithms
We have often seen companies having a better handle on store returns, but struggle with respect to online returns. For one of our clients, while store returns remained flat at 4-5%, online returns hovered around 30-35%, going even as high as 40%.
As one of the ways to address this, we developed an algorithm to predict the online return probability, generating a real-time score during online transactions.
This score reflected the propensity of return incidence based on historical trends in transactional attributes and customer demographic parameters.
Called as Returns Propensity Score or RPS, this score was used to implement differential returns policy, sales strategies, return control measures as well as loyalty program refinements.
This multi-variate scoring algorithm was integrated with machine learning capabilities to ensure any shift in customer behaviour is incorporated in future scores.
RPS was also utilised in drafting return policy with proactive measures to curb return frequency, besides refining loyalty programs and streamlining inventory and supply chain management.
Measure #2 – Intelligent product recommendation algorithms enhance customer experience, and help you sell more.
We have often seen customer conversion rates, and cart sizes, steadily decline for newer transactions.
Often, online product recommendations were either too similar or completely unrelated to the products already in the shopping cart, rather than products most often bought with the product in the cart.
A recurrent cause of such customer behaviour comes from online product recommendations on the website. Products recommended to customers are either too similar or completely unrelated to the products already in their shopping cart, rather than products most often bought with the product in the cart.
Companies fail to realize that products recommended to a 22 year old female from a Metro, should be different from those to a 45 year old male from a Tier 2 City. It should also be different in May versus December. Companies would often overlook profitability and tend to promote products with low margins, which dive into unprofitability when adjusted for shipping, handling and product returns.
The most effective way to address this is to deploy intelligent product recommendation algorithms which suggest products in the order of propensity to buy, based on customer behaviour and historical purchase patterns. These algorithms factor in multiple attributes like customer demographics, past orders, similar customer preferences, seasonality, and especially product profitability and profitable product combinations.
Measure #3 – Store-site identification models cut through and zero in on the right location
Historically, new store locations are identified based on judgement or intuition, often uninformed or generic market data. The company’s own customer preferences, patterns and nuances are often overlooked.
Critical business decision like whether to open a first store in a new city versus second store in an existing city is driven by individual perceptions, rather than customer behavior, profitability modelling or scenario planning. Moreover, all stores are often stocked similarly or without any inventory optimization, thus overlooking profitability as well as regional customer behavior.
A store site identification model utilizes historical data to answer these questions and enable data-driven decision-making. The algorithm identifies new store locations (going down at the zip-code level) based on current customer distribution, online purchase patterns, return behavior, remoteness from existing locations, product preferences, customer demographics, etc.
Zip-code based online purchase patterns are also combined with store site identification algorithms to develop a “store product portfolio” for each store to arrive at the most optimum range of products to offer at stores.
Measure #4– Product identification algorithms, combined with insightful dashboards, help discover effective ways for customer engagement
In most organizations, a wide net is cast to find new customers or wake up dormant customers, rather than a ROI-based, analytical, targeted approach. Not every lost customer is equally likely to come back to the website or store, and of those who do, not every one of them is equally easy, or profitable, to engage.
A set of algorithms, focussed on identifying the right set of products and a right set of customers, can be deployed to ensure marketing campaigns end up doing their job effectively and efficiently.
To attract new customers, product identification algorithms carve out segments of products bought most often by new customers. Promotional campaigns can then be run using the identified products as hooks to target new set of customers.
An intelligent BI dashboard highlights critical business metrics and emerging patterns across multiple dimensions. It’s startling to see often that over 50% of customer base never transacts on the website more than once. Valuable, active customers tend to become dormant after few years, and eventually lost. Algorithms can combat this, by identification of high value customer groups along with likely product recommendations so that targeted marketing campaigns can be deployed to win back identified lost, valuable customers.
Measure #5 – Use algorithmic customer segmentations to identify and incentivize the right customer group and behavior
Often customer behavior is incentivized incorrectly, without factoring in profitability. We have seen return rates for the same set of customer base increasing once they were upgraded from Silver loyalty group to Gold and then to Platinum. Many times, customers are upgraded based on their sales before adjusting for the returns. Similarly, wrong set of customers are also penalized unfairly. We have seen a blanket ban on preferred payment method (e.g. Invoicing after delivery in Europe, Cash on Delivery in India, etc.) for a certain set of customers (e.g. new customers or customers from certain region) without factoring in specific, relevant data points like historical customer behavior, product profitability, seasonality, etc. E.g. There is no reason to deny any payment method to a new customer who is purchasing a backpack, if historically it has had a very low return rate. Return behavior is also seasonally variable, with gifting times like Christmas, New Years driving up return rates.
A right set of predictive algorithms help target the right customer, promote the right product, and incentivize the right behaviour.
Data, as they say, is the new oil. And it’s important to wield and tap into the power of this big data using the engine of analytics and the intelligence of algorithms.
The author of the article is the Director at BIA.
