In our last article, “Breaking Down the Role of Data Science in Business Operations: Introduction”, we broadly discussed the various areas in which data science makes the current day decision maker smarter and more accurate. In that article, we shared the Data Science Business Map which highlighted the various departments and what sections of data science are applicable to them. In this article, we go a few steps deeper into the discussion — focusing on data-driven growth hacking. A growth hacker is defined as someone who uses creative, low-cost strategies to help businesses acquire and retain customers. I will extend the definition to include not just customer acquisition and retention but also increase in business margins.
In 2015, when we were running VoSpine, an interest-based social network, one of the greatest troubles we went through was cracking the best way to get as many users as possible with as low costs as possible. One of our techniques was virality in which we assumed that by getting media interviews, which we did, or doing viral videos on social media, the platform would get traction by users sharing it among themselves. That would turn out not to be the case and when we later ditched the platform to build MealTime — a meal ordering platform and then later Patika — a platform which connects service providers to service seekers with focus on the same virality form of growth hacking among other methods which didn’t do well, we learned a lot in terms of getting a product to the market and especially when it comes to marketing.
“Chema chajiuza, kibaya chajipendeza.” This saying makes sense today just as it did during the times of the sages who coined it. You can spend a billion dollars to market a product but still end up with little to no conversion. I see this a lot with businesses that do offers and promotions. Black Friday is here and online commerce platforms will be offering a lot of offers and promotions but are they just doing it because it is black friday or do they have an understanding of what they expect to achieve at the end? Do they understand the effects of the last black friday and are they doing anything about those insights? This introduces us to our topic of discussion: using data to understand our customers and hence make marketing decisions that cost us the least while giving us the best returns.
This article is written and backed by practical machine learning models and analytics done on a dataset for a retail company that has 4,372 customers transacting over a span of 1 year a total of 4,321,083 orders (with a total of 33,819,606 product sales) worth $ 8.3 million from their catalog of 49,688 products. All the analytics and machine learning models have been done using IDTLab — a big data analytics and machine learning platform by Insense Data Technologies.
Pareto Insights is one very simple but yet very important model that can help businesses understand their customers. In Pareto’s own words, 80% of the effects observed are as a result of 20% of the actions. In business terms, it would thus mean that 80% of your revenue will be generated from 20% of the customers. What this indicates is that your business actually exists because of these 20% of the customers; loosely put, they are so important customers that you should not lose. Identifying who exactly they are is therefore of paramount importance to the business. In our work, we noted that 80% of the revenue, an equivalent of $6.6 million was contributed by only 26.77% of the customers just as predicted by Pareto. We went ahead to get the details of these customers, a dataset which can then be used by the marketing team and the customer service team to know who their most valuable customers are and ensure they take actions that ensure they continue staying with the business. I remember as a Cashier at Equity Bank — Ngara, one of the actions that would get you fired immediately was offending any of the vital few customers. They were demi-gods and the cash officers would be running just to see them served as fast and as greatly as possible. They were not to stay in the queue, they would go straight to the manager’s office and we the cashiers would go there to get their demands, serve them on the counter and go back to them to inform them that it’s completed. They were the real bosses because without them, the bank would not be able to maintain any loan book — the greatest revenue generating service to the bank. This is not to say that the remaining 73.23% of the customers aren’t important — they too are important and in our next section we shall see the insights that can help us understand them better.
80% of the effects observed are as a result of 20% of the actions. This is The Law of The Vital Few. You should know who the vital few customers are in your business; they are the pillars as well as the foundation!
We cannot talk about “Most Valuable Customers” without looking at the Customer Lifetime Value. The Pareto Insights mentioned above is too basic to give as a conclusive understanding of the value of a customer and that’s why, in our IDTLab, we developed a model that takes into consideration the frequency, age, recency and average monetary value of every customer to not only compute the customer lifetime value but to also predict to us their probability of next purchase, its value and also their probability of being “alive” to our business in the future. Based on the recency and frequency of the customers, the following visualization shows our findings on the expected probability of future purchases for our customers.
From this visualization, the yellow to green streak at the bottom show the properties of the customers who are most likely to buy from us in the next one day (we could do for a week, month or year). From this we can see that the bottom right are the most likely to buy while the top right will most probably not buy from the business. How is this important? As a marketing team trying to send an SMS to customers with some offers, it makes more sense to send the SMS to customers who are most likely to come back and buy within a short timeframe in order for the business to notice considerable returns on investment. What this helps with is to prevent the marketers from sending marketing promotion messages or even giving offers to people who will least be converted thus helping save on the cost. A worse scenario is where I have seen businesses so much in the dark about their customers that they can’t tell if the customer has left them or not. A typical example is where a user, who used to stay in Kisumu and go to a restaurant almost daily, moved to Nairobi but almost a year later, the restaurant still sends promotions and gives offers to him. This is a burn on marketing finances and how accurately the marketing team can predict who among their customers are still “alive” is of paramount importance.
This visualization, from our retail business used in this article, tells us the probability of a customer being alive based on their frequency and recency. As can be seen, the yellow streak shows the customers with a near probability value of 1 of being alive while the top right sections indicate the customers who have left the business and hence no more money should be spent on them. Again depending on the nature of the business, it can ask itself why so many of its customers keep leaving. What are they doing wrong in customer retention practices?
From the IDTLab, we were also in a position to show the historical purchasing behavior of a single customer and from here, we can see exactly when the customer stopped being “alive” or any of the hiatus the customer has had over the time — insights that are very important for the business to offer targeted advertising, specialized attention and generally help increase retention, conversion and returns with reduced costs.
Businesses should use data to understand the CLV of their customers so as to understand the limits to how much they should spend on them or on customer acquisition.
It is one thing to apply a blanket analytics to your customers but it is another to understand how different customer segments behave. The power of machine learning models on big data is its ability to create customer segments and perform Cohort Analysis on the various segments. At Insense Data Technologies, we describe cohort analysis as the ability to understand customers based on their unique behavioral attributes. It is a powerful analytics technique to group customers and enable the business to customize their product offering and marketing strategy. The customers are grouped into mutually exclusive cohorts — which are then measured over time. Cohort analysis provides deeper insights than the so-called vanity metrics, it helps with understanding the high level trends better by providing insights on metrics across both the product and the customer lifecycle. At IDT, we look into various types of cohorts: time, behaviour and size cohorts. Time cohorts tells the business about the customers they on-boarded around a similar time. Analyzing these cohorts shows the customer’s behavior depending on the time they started using the company’s products or services. Behaviour cohorts groups the customers by the type of products or services they purchased/subscribed to while size cohorts looks at the spending amount. Obviously the behavior of low spenders are different from that of high spenders.
As you can see above based on our analysis of the retail data based on customer time cohorts, we can visualize the retention rates of the various cohorts over time and ask ourselves the relevant quetions that can help us make better decisions going forward. Assume a business kept track of the customers they acquired during the last Black Friday and are able to see their retention rates as shown above or spending habits over time, then they will be more informed in terms of how they should conduct their offers this Black Friday so as to maintain high returns, conversion and/or retention. Do not worry, that’s what we do to our clients at Insense Data Technologies. You just need to holla us at [email protected]
Assume a business kept track of the customers they acquired during the last Black Friday and are able to see their retention rates or spending habits over time, then they will be more informed in terms of how they should conduct their offers this Black Friday so as to maintain high returns, conversion and/or retention.
We cannot talk about understanding our customers and mention CLV, Pareto Insights, and Cohort Analysis and then forget to talk about RFM analysis. RFM simply refers to the recency, frequency and monetary value of the customers; more or less the same parameters we used during the prediction of customer lifetime value. A customer segmentation technique, RFM Analysis’ central idea is to segment customers based on when their last purchase was, how often they’ve purchased in the past, and how much they’ve spent overall. All three of these measures have proven to be effective predictors of a customer’s willingness to engage in marketing messages and offers. The marketing team can use these insights to offer personalized and targeted offers, achieve higher response and retention rates, impriove unit economics and increase revenue and profits. At its core, these insights can help you answer a few questions that every business struggles with:
- Who are my best customers?
- Which customers are at the verge of churning?
- Who has the potential to be converted to more profitable customers?
- Who are lost customers that you don’t need to pay much attention to?
- Which customers must we retain?
- Who are your loyal customers?
- Which group of customers is most likely to respond to your current campaign?
Companies can effectively predict a customer’s willingness to engage in marketing messages and offers by using data to conduct RFM Analysis
In our IDTLab modelling of the dataset for the retail company we mentioned earlier, we built a 10 cluster K-Means model based on customers’ RFM which can been seen in the visualizations below:
Our machine learning model for data mining, which we trained in less than 5 minutes, generated some very important insights. The following graphs show the behavior of every customer cluster
and we were able to summarize each of the clusters into named groups, each with specific actionable redemption strategies.
For instance, we advise the company that all the customers in the clusers 1 and 2 (At Risk) spent big money and purchased often but long time ago. These are customers the company needs to bring back! In order for them to do so, they need to send personalized emails or SMSes to reconnect, offer renewals, provide helpful resources. Customers in cluster 0 are champions and loyal customers — that means they bought recently, buy often, spend the most and are responsive to promotions. For this cluster, we advise the business to reward them, upsell higher value products, ask for reviews and engage them .They can be early adopters for new products and will promote your brand.
Compare the Champions to NBA’s LeBron James or Football’s Messi and Ronaldo. No team wants to lose such a player. Neither should your business!
We have focused so much on understanding our customers and how the customer service and marketing/sales teams can use data and/or the services of Insense Data Technologies to improve their output. The one section that I will briefly touch on before we conclude is on market basket analysis.
Market Basket Analysis is a data mining, frequent and sequential pattern mining to be specific, technique that helps us understand the purchasing behavour of not just the customers but also the products. It helps the business understand what people always buy together and use these insights to make numerous profit lifting decisions. Take for instance the market basket analysis we did on the retailer business we’ve been talking about; we were able to generate association rules and determine not only the frequent itemsets but also lists of consequents given specific antecedents.
Here we can see the products that appear the most frequent in most baskets/shopping carts. This is different from the total orders or purchases for a product which would be biased where one person purchases lots of products. This insight tells us the products most loved by our customers and in this case, bananas, the bag of organic bananas and organic strawberries are the pillars of this business.
But that’s just not even the half of it, the power of market basket analysis comes in when we start using the insights to understand the correlation and cuasations among the product sets.
As you can see above, buying strawberry rhubarb yoghurt increases a customer’s chances of buying blueberry yorghut 80 times fold even though the confidence is a mere 30.9 %. That was a shocker even to me!
A peek in this report shows that buying organic raspberries, organic hass avocado and organic strawberries increases a customer’s chance of buying a bag of organic bananas 5 times fold and with a confidence of 59.8%.
We even went ahead to look at their most loved product — bananas:
and we found out that a customer who buys bananas from the store has almost thrice as much chance of buying Bartlett pears, gala apples, sweet potato yam as someone who has not purchased bananas.
Such golden insights we can derive from market basket analysis are useful in making decisions pertaining to product recommendations (the famous: those who bought this also bought that), store layout planning (have bartlett pears and bananas in the same shelf not one on ground floor and the other in second floor), targetted marketing (if someone buys a banana, target them with offers on bartlett pears), cross-selling, up-selling, sales promotions, loyalty programs, discount plans among many other things.
“Data is the new oil” is becoming more hackneyed than it is being taken seriously. We talk of data everywhere but on the ground things are very different; we don’t use data to make decisions. Stop having products on discount simply because the expiry date is nigh and you want to get rid of them before they go bad; have products on discount based on machine learning insights and because you want to achieve some mission on revenue growth, customer loyalty, improving retention or anything other than to simply offload the product before it goes bad.
As a marketing lead in your company, it is upon you to try something new, try data driven growth hacking. Make decisions based on data and not the other way round: using data to try support your decisions which is what most people do. As I stated when starting, we didn’t use data in our previous startups and decisions were made based on opinions of a few. For instance, the decision to add a new feature should be based on data; before Facebook rolls out a new feature they conduct customer segmentation and roll out the feature to different samples from the different segments to understand the impact and reception. Based on the data, they can make a decision on whether to improve it, go ahead to roll it to everyone or stop it all together. At Google, high level decisions are made by machines more than humans; the humans are more or less just on the check and on the look out.
Those are big companies you say and I agree. As a small business owner, you may not be able to afford the expensive data handling operations. Thankfully, that’s the inspiration for founding Insense Data Technologies — to help retailers, founders, marketing leads, customer service teams among many others from any size of company with affordable big data, machine learning and data consultancy services.
The discussion continues on our next article…