APPLICATIONS OF DATA SCIENCE FOR RETAIL 

(1-A) Store Segmentation

What is store-segmentation?

One blanket strategy for Sales, Marketing, Pricing and Ranging for a company having stores at multiple locations will not work in today's time of intense competition. Instead, companies use Store Segmentation which is the process of grouping stores with common characteristics. It enables optimization of all store level activities that are suitable for that segment of store. It goes beyond just finding the right location: “it’s about personalization for stores".

It is especially important when the company does not have enough data about customer-demographics or past transactions (E.g., while opening a new branch). 

Based On what characteristics are stores segmented?

Store-segmentation process necessitates a complete analysis of the entire market, not only customer’s needs and shopping habits but also knowledge of changing market conditions and competition.

Usually stores are segmented using clustering based on characteristics like socioeconomic profile of the trade-area, competition, sales-area and special services that stores offer. Some such characteristics are shown in figure above.

What next after store-segmentation?

After segmenting stores using clustering, frequent item-sets and association rules analysis will be applied for each cluster for the purpose of target marketing. This way, a retailer will be able to find the differences between the clusters in terms of the items sold at the stores. 

(1-B) Product Segmentation

What is product-segmentation?

One-product-fits-all approach no longer works in retail. Along with differing personality traits come different customer needs and preferences. To cater to them, companies use ‘Product-Segmentation’ i.e., offering different versions of its product to different groups of people. E.g.  a company might have a base product that provides the core functionality, and several versions built around that core that cater to individual industries. 

Companies spend extra efforts, time and money on product-segmentation as they view it not as expansion of one product but as expansion of its product range.

What are benefits of Product-Segmentation?

(1)Monitor individual products

When products more closely align with a given customer-segment, company is able to take each product in turn and monitor its performance. This allows it to get a better picture of how each customer-segment responds to its product by letting it monitor the product most closely aligned with them, rather than one nebulous product. 

(2) Cater to different customer groups

With one nebulous product  some customers often feel they are paying for features they do not need and others may feel they do not have enough features. Product segmentation, when built around core buyer personas, allows a company adopt balanced approach between these two extremes, acquire more customers, and improve customer retention. 

(3)Understand target market to determine the right pricing

Once there is no one-size-fits-all approach in product offering, the company can also abandon one-size-fits-all pricing. 

As a segmented product tailored to different customer bases, company is able to adopt a pricing strategy that is ideal for each customer base.

(4) Explore new growth opportunities

The great thing about segmenting product is that it makes it easy to uncover growth opportunities. 

As company examines the types of customers and feedback received from them, a picture will begin to form of how it can further segment its base product to appeal to a wider range of people and unlock new growth levers.

(5) Segment willingness to pay data

If company has "willingness to pay data" it can segment it by customer groups, but that data will be less than useful so long as there is only one identical product for everyone to purchase. 

Creating multiple products that serve different segments allows company to segment that data  as per product-segments and price them accordingly.

What are product segmentation strategies?

Once a company has decided that it wants to use product segmentation, it must consider following four key product segmentation strategies:

(2) RFM Analytics

What is RFM Analytics?

RFM (Recency, Frequency, Monetary) analysis is a proven marketing model for behavior based customer segmentation. RFM analysis is based on the marketing adage that "80% of  business comes from 20% of customers." 

It groups customers based on their transaction history – how recently, how often and how much did they buy. RFM analysis is used to identify existing customers who are most likely to respond to a new offer. 

Judging customer value on just one aspect (Monetary value of transaction) gives an inaccurate report of customer base and their lifetime value. As a customer may buy big but only once, or his big purchase might have been long time back. That’s why, RFM model assigns each customer three numerical scores: Recency-Score, Frequency-Score and Market-Value Score. Then the three scores are combined to get a combined RFM score and customers are ranked based on RFM Score.

How are customers classified based on their RFM score?

Customers are classified into segments based on this combined score as shown below:

(3) Purchase Likelihood Analytics

What is the likelihood to purchase?

The likelihood to purchase (LTP) indicates the high probability of some visitors who are more likely to make a purchase within an expected time frame (normally 6-12 months). 

The exact time-point when customer's purchase intent will be converted to actual purchase depends upon where the customer is in the four stages of customer journey as shown below:

On which factors does LTP depend?

Likelihood to purchase  (LTP) depends on factors like

(1) Seasonality
(2) Satisfaction of existing customer
(3) Customer Demographics
(4) Advertising

How is LTP measured?

Likelihood to purchase  (LTP) can be measured by surveys like

(1) Brand tracking Surveys
(2) Product testing Surveys
(3) Customer Satisfaction Surveys
(4) Ad Testing Surveys

In addition, for e-commerce, continuous analysis of user's every interaction(E.g.Click-stream-analysis, wish-list analysis, abandoned-cart-analysis etc.) can provide useful information about where customer is in his journey and help in estimation of his LTP. Customers with high LTP are found by customer-segmentation using ML algorithms which learn from buyer behavior and predict the probability of a purchase.  

How is customer LTP related to CLTV?

A customer’s lifetime value (CLTV) and likelihood to purchase (LTP) are related and they fall into one of three categories: low, moderate, or high. 

Due to intense competition and increasing marketing costs, acquiring new customers is becoming expensive. As lifetime values decrease, companies and agencies can use LTP segments to bring more conversions, improve ROAS, and optimize marketing budgets. Growth marketers can make use of high LTP segments on costly digital ad-platforms to turn more clicks into purchases. 

What are benefits of analysis of LTP?

(4) Pricing Analytics

What is Pricing Analytics?

Pricing analytics is a technique that provides companies with the tools and methods to better perceive, interpret and predict consumer behavior w.r.t. price changes.

Why is Pricing Analytics Required?

Pricing merits a detailed analytics because pricing power comes from understanding ‘what consumers want’, ‘where they shop’, ‘which offers they respond to’ and ‘how much they are willing to pay for a product or service’. 

Further the retail space is becoming highly tech-driven with 80% customers using price-comparison tools. Hence, industry leaders are abandoning manual processes and adopting data driven pricing tools which utilize price-analysis data from various channels of a company’s operations, enable competitive pricing decisions and a better ROI. The latest trends in retail Pricing Analytics are shown below.

However this does not mean that modernizing technology alone is enough for deciding Pricing-Strategy of a company. It must choose a Pricing Analytics Solution that is able to answer questions like:

(1)How is Pricing Influencing Brand Associations?
(2)Is current Price-Brand Association Optimal?
(3)Does the solution provide comparison with competitors' Price-Brand associations?
(4)Does the solution takes both price and value into account?

What are other benefits of Pricing Analytics in addition to setting  optimal price?

(5) Cross-Selling And Up-selling

5-A Cross-Selling

What is cross-selling?

Cross-selling is the process of satisfying the additional needs of customers that the primary product is unable to fulfill. It means encouraging customers to buy complementary products, increasing the overall cart value. 

For example, shopping from the in-flight product catalog, such as a pair of headphones, would be a cross-sell. Some great companies which very successfully use cross-selling are Amazon, ASICS, IKEA, Apple, Sephora etc.

What are advantages Of Cross-Selling?

(1)Helps increase overall sales value.
(2)Better customer-comprehension and engagement.
(3)Greater convenience to customers via useful suggestions.
(4)Helps discover similar/complementary products/services.
(5)Creates an environment of mutual benefit for businesses.
(6)Helps to promote and sell a larger variety of products.
(7)Promotes innovation of new products.
(8)Increases customer lifetime value.
(9)Enhances customer-satisfaction.

What is up-selling?

Up-selling is the technique whose end goal is to increase the cart value not by introducing new items but by increasing the size or quantity of the initial purchase decision. For instance, upgrading an economy airplane ticket to business class would be an Up-sell. Some great companies which very successfully use up-selling are Spotify, GoDaddy, Booking.com and DELL.

What are advantages Of Up-selling?

Why cross-selling and up-selling are so powerful?

As shown above, it is far easier to get more business from existing customers than acquiring new customer. Both cross-selling and up-selling are so powerful as they are aimed at getting more revenue from already existing customer and improve customer-retention. Increasing customer retention by as just 5% can increase profits from 25% to 95%³.  

According to a 2022 HubSpot Blog survey⁴ of more than 500 sales professionals, 72% of salespeople who up-sell and 74% who cross-sell say that it drives up to 30% of their revenue.

In HubSpot survey of 2022 the response to question: “What percentage of your company's revenue comes from cross-sell or up-sell?” the answers obtained are depicted above which clearly show the power of cross-sell and up-sell. This is particularly true for e-commerce site. Up-selling can raise sales by 10-30% on eCommerce sites, while cross-selling raises profits by over 30% on an average¹. 

Moreover, reports say that 77% of revenue comes from up-sell, renewals and cross-sell² in the case of B2B businesses.

What are strategies for cross-sell & up-sell?

(6) Recommendation Engine

Introduction

Ever wondered how Facebook suggest "people you might know," and you know the majority of them! Or, how Amazon knows what kind of product you need and show you the most attractive options? Or how can Netflix figure out what kind of movies or series you might like? They use a recommendation-system. A recommender system is a delicate way of bringing user close to relevant content from fast sea of information.

Some other top companies which use recommender system are:Spotify, LinkedIn, YouTube, Instagram, Quora and Google.

What is market potential of recommender system?

The global market for the usage of Recommendation Engine was valued at USD 2.69 billion in 2021. It is anticipated to surpass USD 15.10 billion by 2026, reporting a CAGR of 37.79% during 2022-2026.

What is a recommendation system?

Technically a recommendation engine is a data-filtering system based on ML algorithms to recommend products, services, and information to users. They are combinations of information filtering and matching algorithms that bring together content and user.

From where do they get data?

The data can be explicitly or implicitly collected.  Explicit data is generated by a user taking a direct action that indicates user's preferences (for example, information about past activity, ratings, reviews, gender, age, income-profile). Implicit data is generated during interaction of user with website E.g. device used for access, click-stream, location, and dates.

What do they do with data?

Recommendation engines discovers data patterns in the data set by learning consumers choices and produces the outcomes that co-relates to their needs and interests.

How do the recommendation engines work?

Types Of Recommendation Systems

(1) Demographic Systems:

Such systems take available user data (age, gender, location, etc.), classify it into specific audience segments, and then put it in a bigger picture to fill the gaps in the data.

Demographic-based suggestions are widely used on content-aggregation websites and in the general eCommerce marketplace. Usually, this type of recommendation provides a background operation in case there is no other information available.

Advantage:
Simple and can be implemented with even a limited set of data to deliver broad suggestions. 

Disadvantage: 

To make it work, this system requires full-on market-research as a foundation. 

(2)Knowledge-Based(Cognitive) Systems

Knowledge-based systems are systems where suggestions are based on a user’s preferences and certain constraints that are based on a degree of domain-knowledge. Rules are defined that set context for each recommendation. These do not have to use interaction history of a user  but can include these as well. 

Advantages:

(A) Given the way the system is built up, the recommendations can be easily explained.

(B) It is better suited to complex domains where prior data is lacking. As prior data is not needed it doesn’t suffer cold-start up problems.  

Disadvantages:

Building up this type of framework can be expensive. 

(3) Content Based Filtering

A recommendation like ‘products similar to this’, is a typical instance of this type of approach. Here algorithm recommends products which are similar to the ones that a user has liked in the past. For example, if a person has liked an article on Lionel Messi then this detail is stored by the blogging site including rating given by him in his ‘Profile-Vector’. All information of other articles is stored in ‘Item-Vector’. The Algorithm finds ‘Cosine Similarity’ between these two vectors (either by top-n method or rating method). This Similarity can also be calculated by the ‘Euclidean Distance’ or the ‘Pearson Correlation’.

With an approach like this, the more information that the user provides, the higher the accuracy. Hence often the metadata collected from a user’s history and interactions is also used by algorithm.

(A) Advantages:

This recommender engine technique does not need any additional data about other users since the recommendations are specific to this user. Also, this model can capture the particular interests of a user and suggest niche objects that very few other users are interested in. Hence, niche eCommerce stores (Discogs and Artsy etc.) or content aggregation websites (Mashable and The NextWeb etc.) use this recommender system.

(B) Disadvantages: 

Due to privacy and regulatory issues often personal metadata and individual transactional data can be missing at the outset. So, a recommender system cannot draw inferences for a query due to lack of sufficient information (cold start problem)

(4) Collaborative Filtering:

The collaborative filtering method collects and analyzes data on user behavior, online activities, and preferences to predict what they will like based on the similarity with other users. So far as user data collection is concerned collaborative filtering casts a much wider net, collecting information from the interactions from many other users to derive suggestions for one user. This type of recommender system is common in the eCommerce marketplaces.

(A)Advantages: Collaborative filtering has higher accuracy than content-based filtering. 

(B)Disadvantages: 

Popularity bias: Refers to system recommends items with the most interactions without any personalization.

Item cold-start problem: Refers to when items added to the catalogue have either none or very little interactions while recommender rely on the item’s interactions to make recommendations.

Scalability issue: refers to lack of the ability to scale to much larger sets of data when more and more users and items added into our database.

(5) Hybrid Recommenders:

A content-based filtering model will not select items if the user’s previous behavior does not provide evidence for this. 

A Collaborative filtering cannot provide recommendations for new items if there are no user ratings upon which to base a prediction.

Hence hybrid recommenders that combine above two approaches are becoming popular. They take advantage from both the ‘representation of the content’ as well as the ‘similarities among users’. There are two main ways of combining 

(A) Combining Item Scores

(B) Combining Item Ranks

Advantages:

This  is latest recommendation system and can outperform both ‘content based’ and ‘collaborative’ recommenders in terms of accuracy. E.g. Netflix uses a hybrid recommendation engine. 

Disadvantages:

These are sophisticated and costly to implement. Suitable only for companies with huge product base and user-base.

(5) SVD, SGD and ALS Recommenders:

The three disadvantages of collaborative filtering listed in (4) above can be solved by matrix-factorization algorithms like  SVD (Singular Value Decomposition), SGD (Stochastic Gradient Descent) ALS (Alternative Least Squares). Among this ALS is scalable across a distributed data-set of size petabytes.  

How To Choose Recommendation Systems?

So, Finally your organization has decided to adopt recommendation system. Congrats! The road-map below will help you choose the right one:

(7) Demand Forecasting And Planning

What Is Demand Forecasting?

Demand forecasting is a prediction of the willingness of consumers to purchase specific products/service, at a specific price, in a specific time frame. It enables a business to estimate sales and revenue for that time frame. It also estimates the possible demand hike or fall that may occur in the market.

What Data Sources Are Used For Demand Forecasting?

Demand forecasting uses data like historical sales, the time of day, season, past marketing campaigns, customer preferences, expert-insights, economic trends, opinions of sales force, market conditions of the period.

Demand-Forecasting  Vs. Demand-Planning

Demand Forecasting is process that generates forecasts of customer demand. Demand Planning, on the other hand, is process that converts those forecasts into actions by breaking down financial budgets into categories, and stores. 

Demand-Forecasting Vs. Sales-Forecasting:

(A)Sales-forecasting tools cannot calculate the sales you lost because a business ran out of stock. For example it sold 100 units of an items as there were only 100 units in inventory (while actual demand for that item was 250 units). So the business lost sale of 150 units to its competitors. Thus Sales forecasting creates a self-defeating strategy because the retailer will continue to understock high-performing products. 

(B)Sales-forecasting tools rely on past sales to extrapolate trends into the future. Demand forecasting goes beyond simple trend extrapolation, accounting for hundreds of factors that influence demand for each SKU in every channel (price, events, product families, assortment, product cannibalization, etc.)

Why should every retailer use Demand Forecasting and Planning?

Demand Forecasting has innumerable benefits in retail industry as listed below:

According to Gartner⁵ demand forecasting is the most used ML technique for retail. It can help big or small retailers both to minimize cost and maximize profits by:

(1) Optimal inventory management

Inventory is the largest investment that the retailers make (even above real estate, merchandising or advertising). Inventory cost includes cost paid to supplier, costs associated with running distribution centers, costs associated with transporting stock to the stores, carrying cost (cost from shrink, clearance-price). Inventory also has high ‘opportunity-cost’ (ties up cash-flow and shelf space which could have been used for more promising products). Inventory affects the most important retail KPI: GMROI. 

Forecasting how much of every SKU needs to be stocked in every store and distribution center is mission-critical for modern retailers and both underestimation and overestimation of demand have harmful consequences as the below diagram shows:

(2) Choosing optimal prices for every SKU

Armed with accurate forecast, the retailer can choose optimal price for items. For example, raising prices of items whose demand is expected to rise or offering them only in combos and also ensuring their optimal stock. Conversely declaring sale on an item that is likely to see less demand and clearing shelves for more promising items.

(3)Improved Customer Experience, Optimal Product Availability 

Accurate demand forecasting ensures that at the time of peak demand of an item/brand there is enough stock of it and no customer returns dejected due to ‘Out-Of-Stock’ notice. Similarly in brick-mortar shop it ensures there are always optimal number of sales-executives present.

(4) Better Assortment, Merchandising and Promotion

In retail business (be it brick-mortar shop or eCommerce website) the assortment of items and great merchandising is what differentiates successful players from failures. Demand forecasting helps in this. It also helps in designing promotional campaigns, sales and discounts.

(5) Smart staffing and smart shift scheduling

Payroll expenses are major cost-components for retailers. Correct estimation of demand ensures that only optimal staff is present in a given store , in a given shift throughout the year. This saves huge money.

(6) Assess risk of launching new product/business

Launch of new product/business is always more risky due to lack of historical data. Forecast for such products no longer need to be ‘guesswork’ as some demand forecasting methods like ‘Qualitative Forecasting Model’ or ‘Macro-level External Demand Forecasting’ can reveal demand for products you don’t yet offer or new business being launched.

(7) Run Better Marketing Campaigns

Understanding customer demand makes the marketing efforts more intentional and cost-effective. For example:

(A) When a drop in sales is coming, a business can run limited-time discounts or double down on your loyalty program and encourage repeat purchases. That’s also a great time to explore low-cost digital marketing ideas or social media contests. 

(B) Leading up to peak sales periods, you can partner with local influencers or experiment with paid ads on social media and Google to capitalize on the high demand.

(8) Have better budget accuracy and maintain positive cash flow

All the benefits of demand forecasting boil down to better management of budget and cash flow. Armed with accurate forecast a business is always confident of meeting its fixed and variable expenses. Good demand forecasting helps to set and stick to a budget and helps predict revenue, profit and cash-flow. Any other business plan like capacity planning, expansion planning etc are also impossible without demand planning.

(9) Reduction in lost Sales. Less error in supply chain

As per Mckinsey Digital⁶, AI-powered demand forecasting can reduce mistakes by 30 to 50% in supply chain networks. It leads up to a 65% reduction in lost sales due to inventory out-of-stock issues.

(10) Better margins and timely discounts
Forecasts can show when demand is likely to be both low and high, helping retailers know when to offer a discount in order to drum up business, or when demand is high,  tweak pricing to get the best margin possible.

(11) Fresher Items To Customers And Reduced Wastage

Food waste costs retailers about $18.2 billion annually⁷.Supermarkets and grocery stores have a particularly tough time ensuring freshness of products. Fresh produce has a limited shelf-life meaning demand forecasts are crucial to avoid wastage & ensure customer satisfaction.
(12) Carbon foot-print reduction and reduced storage-cost

Being sustainable and ‘Green’ business is must for retailers and it also enhances their brand-value. Gone are the time of just throwing away unsold inventory, at the end of season, to free-up shelves. Right solution is accurate demand-prediction which enables optimal use of storage and also saves storage cost by 10 to 40%⁶.

How big opportunity is demand-forecasting?

Even in developed countries like US 34% of top 50 retailers have poor forecasting accuracy. Due to their large scale of operation, demand forecast deviation of just 1% in either direction (over or under projecting) could potentially cost a company up to millions of dollars. This shows how big opportunity demand-forecasting is for both retailers and data-scientists.

What are techniques of demand forecasting?

Many intuition based retailers start demand forecasting with wrong question: “How much do I need to sell?”. This leads them to endless spiral of wrong actions. Savvy retailers start with right question. “What is the true customer demand for all of our products?” and then use suitable demand forecasting techniques from below to find its answer:

(1) Qualitative Techniques:

Qualitative forecasts are built up from market research from four main sources:

(A)Public opinion surveys, which provide a sense of consumer confidence.

(B)More focused surveys can reveal consumers’ purchase intentions segment-wise.

(C)Insights from Consultants, analysts, and other experts based on their work within the industry.

(D)Leadership insights of top management.

This type of model is best suited for retail businesses that do not possess any previous or historical data. For example, companies that launch a new product, the drastic time difference between prior and planned period, etc. Delphi method, customer surveys, and sales force composite method are qualitative forecasting methods.

(A) Delphi Method: 

The Delphi method involves tapping a panel of demand forecasting experts outside the company to get their anonymous opinions on the company's market forecast. The idea of the Delphi method is to encourage the group to arrive at a consensus as to what the proper demand forecast is after several rounds of deliberations.

(B) Market Research:

Market research involves conducting customer-surveys, one on-one interviews, and focus-groups with customers. It provides a wealth of insights about an organization's customers that management simply would not find in their internal sales data.

(C) Sales Force Composite: 

The sales force composite method rests on the idea that, due to their proximity to customers, sales teams have intimate knowledge of their desires, challenges, pain points, and feedback and also aware of what competitors are doing. This method puts company's sales team in control of creating the demand forecast. 

Producing Forecasts with qualitative techniques can take a few weeks.

(2) Quantitative Methods

These methods leverage current/historical statistical data to predict future outcomes. Quantitative methods are considered to be more objective than qualitative methods as they use verifiable facts and figures. However, quantitative data still has to be contextualized to be of real use to a business. Some examples of quantitative methods are:

(A) Trend Projection: 

Trend projections is a simple and effective technique that uses verifiable historical sales data to predict future sales volumes. It works for businesses with at least 18-24 months of trading data using which a time-series is plotted and future demand is forecasted  assuming trading conditions remain the same. One can make this more sophisticated analytics  by incorporating data from web analytics, environment data, historical shipments and known future orders, and competitor pricing. In such case forecast can even flag emerging short-term trends. 

Producing basic time-series forecasts may take a few hours.

(B) Econometric Method:

These method uses complex mathematical models to establish relationships between sales data and external forces that influence customer demand. It aims to derive and fine-tune an equation to establish a reliable historical correlation. The projected values of the variables that influence demand are then plugged into the equation to produce a demand forecast.

(C) Barometric method: 

The barometric technique is more sophisticated, using statistical analysis and economic indicators to generate a demand forecast.

(D) Causal Model:

This model accounts for factors that may change the initially forecasted demand. Here, data is split into two different factors i.e. controllable and uncontrollable factors. Controllable factors include marketing efforts, location, pricing, sales, and visual merchandise. Uncontrollable factors include weather, politics, competitors, natural calamity, and more. 

This model is a little complex since it acquires a many variables to look upon. In addition to that, factors like weather and natural calamity are hard to predict accurately. However, causal model  is best suited for retailers in the volatile market, multi-channel businesses, data-driven retailers with a lot of metrics over time. Developing sophisticated causal models can be a year-long process.

Types Of Demand Forecasting Based On Economy Or Time:

1.Industry-level Demand Forecasting: 

This calculates the total quantity of the products of an industry required in the market.
2.External(Macro)Level Forecasting

This forecasting is done by calculating the relation between IIP and economic environment (including general employment level and national income level). 
3. Internal Micro Level Forecasting 

The more control a retailer has over daily business operations, the better will be the returns. Internal micro forecasting reviews the day-to-day operations and identifies areas for improvement. 
4. Short-term Demand Forecasting: 

Useful in tactical decisions for the short period-based industries. Forecasts are useful for <1 year period.
5. Long-term Demand Forecast: 

Used For businesses or industries that make long term operation and trading commitments. Such terms may be >2 years.
6. Passive Demand Forecasting: 

It is used in  that are stable and have a conservative development plan (e.g. small or local businesses).
7. Active Demand Forecasting

Used for businesses looking for aggressive expansion and diversification. Suitable for startups and businesses in the growing phase.

How To Create Demand Management System? 

1. Create a Demand Management Database:

This involves identifying available data-sources (based on forecasting method to be used) and creating a system for gathering and storing information from them in a database. This is an iterative process for creating demand forecasts. A single database as data-source helps to protect the integrity of projections.

2. Gather Data from the Sales Department
This is technically part of the data gathering process, but the nature of sales data makes it a separate component, as it is distinct from most supply chain data. This step allows sales teams to input and review their demand estimates independently from other departments in the business.

3. Manipulate and Analyze the Data:
With the database in place, organizations can begin manipulating and analyzing the information to discover actionable intelligence. Examples of analysis techniques include filtering, comparisons, drill-downs, and monitoring changes over specific time-frames. Results are reported in real time by using a dash boarding tool.

4. Develop a Method of Creating Forecasts:
A simple demand forecast can be geared towards a point estimate for a product or category over a given period. Accurate forecasts are the product of a collaborative effort between managers and teams from different departments.

5. Evaluation:
Demand management also requires a system for ongoing evaluation and improvement. Demand management software can make simple work of this task with automatic reporting and alerts. After evaluation if need is felt new data can be acquired.

Why Legacy Demand Prediction tools must be replaced by data-driven methods?

In the past, retailers did SKU-level forecasting only for the most important products and covered the rest of the assortment in category-level forecasts. Much more nuanced understanding of each SKU is needed in today’s dynamic retailing environment. 

On top of that, even smaller retailers must address the omni-channel reality  and price comparison tools of today’s shopping environment. But those factors are just the tip of the iceberg. Beneath the surface, hundreds of other factors influence demand. Not accounting for each one will undermine your forecasts, leaving you with excess inventory and lost sales. Here are just a few of the subtle variables that shape demand:

Seasonality: 

Seasonality has always been important part of a Retailer's life. But for retailers with regional or national footprints each store and each product may have its own unique seasonality.
Price elasticity: 

The simple adage: “sales go up when prices go down” is not enough in modern retail sales forecast. Sales depends on promotions,market conditions assortments and many other contextual factors.
Promotional uplifts:

Forecasting uplifts is always difficult as  it depends not only on offer but also on consumer perceptions and sales-channels. 
Supply chain/vendor lead times: 

When vendor lead times are measured in months, stores have to place orders before knowing the demand. In many cases, variations in lead-times affects the availability of your merchandise.
Different store types: 

Medium to large retailers often have multiple store types, for example, express/convenience, pop-up stores, malls, standalone showrooms, and outlet centers, eCommerce Store, eCommerce App etc. So data for demand forecasting must be gathered from all these sources.                                                                      Geo demographics:

Even for stores within the same city, the communities surrounding each of different stores have different tastes, shopping patterns, and interests. These can impact per-store-per-SKU sales quite dramatically  Retailers can’t simply offer the same assortment and inventory levels across all stores. The personalized assortment has quickly become the preferred method of allocation of inventory.                              

Competition: 

Each product has unique competitive environment.Retailers must keep an eye on merchandise they compete on in terms of pricing and promotions in-store and online.

Product demand cannibalization: 

Promotions, new product launches etc. draw customers’ attention to one SKU will also draw their attention (and purchases) away from other SKUs. Such product cannibalization is a challenging variable to account for in demand forecasting.
Assortment depth vs. diversity: 

Assortment strategy influences demand. Variety reassures customers that the retailer can meet their individual tastes while depth gives customers confidence that they can buy what they want when they want. At the same time, retailers do not stock same inventory level for all products as  usually 20% of products yield 80% of profit. Traditional demand forecasting tools can not manage this complexity. They cannot match the big data  and AI based analytics tools that modern retailers use. When using legacy demand forecasts, retail analysts must build hundreds of demand forecasting constraints into their models and adjust them for each SKU in each store. Even an army of analysts could not do this effectively for thousands of SKUs across hundreds or thousands of stores.

(8) Customer Churn Prediction

Customer churn is one of the most vital data points for businesses to track. The best way to prevent customer loss is to predict if some of them are inclined to leave you. Customer-churn analytics help in this. The recent bankruptcy of an American retailer Toys R is largely attributed to lack of insights provided by churn analytics.  

Definition Of Customer Churn And What is Churn-rate?:

Customer churn is the rate at which customers leave your company. It is a useful metrics to understand the amount of business-loss. Churn rate is the number of customers who leave a product over a given period of time, divided by the total remaining customers over that time. 

Online businesses typically treat a customer as churned once a particular amount of time has elapsed since the customer’s last interaction with the site or service.

Why Customer Churn is undesirable?

A churned customer means lost revenue and the marketing costs involved with replacing that customer with  a new one. Further it is always more difficult and expensive to acquire a new customer than it is to retain a current paying customer. 

How big is the problem?

The annual cost of client attrition is $4 billion in the U.S. and Europe and around $10 billion globally. Though severity of problem is industry specific; both online and offline retail suffer from very high churn rate as shown in diagram above.

What methods are used for churn prediction?

(1) By calculating each customer's churn-factor:

A customer’s churn-factor is measured by dividing the time since the customer’s last activity by the customer’s activity-frequency.  If a customer’s churn factor is high, it is more likely that the customer has already churned. Using churn factor to analyze customer behavior considers each customer’s behavior in context, creating a simple yet very powerful churn prediction. Further analyzing customers’ risk of churn in the context of their behavioral patterns gives even deeper understanding of each customers’ behavior, increasing potential to retain them longer.

(2) By Machine Learning Models based on static data and metrics:

Many retailers have started using ML based churn-prediction models long back, which use techniques like logistic-regression or binary modelling. These approaches can identify a certain percentage of at-risk customers, but they are relatively inaccurate. This inaccuracy hurts retailer two ways: if the marketer is unaware of a customer about to churn, no action will be taken for that customer. Conversely, unnecessary incentives would be offered to a happy customer wrongly predicted as at risk customer, resulting in loss. 

(3) ML based techniques which take into account real-time data and metrics:

Hence in today's dynamic retail environment even modern models are used. They calculate Customer Life-Time-Value(CLV) based for each and every customer using a new research-based techniques. They are based on combination of  continual dynamic micro-segmentation of customers and a predictive behavior modeling system.

The micro-segmentation intelligently and automatically segments the entire customer base into a hierarchical structure of ever-smaller behavioral-demographic segments. This segmentation is dynamic and updated continually based on changes in the data.

The predictive behavior modelling system is based on the fact that the behavior patterns of individual customers frequently change over time. In other words, the “segment route history” of each customer is an extremely important factor determining when and why the customer may churn

By merging the most exacting micro-segmentation available anywhere with a deep understanding of how customers move from one micro-segment to another over time – including the ability to predict those moves before they occur – an unprecedented degree of churn analysis accuracy is attained.

What are benefits of churn-analytics? 

(1) Identify customers about to churn
(2) Identify Pain Points
(3) Remove pain points and implement churn-prevention methods
(4) Provide better customer experience

(9) Supply Chain Analytics

Supply chain analytics refers to the processes organizations use to gain insight and extract value from the large amounts of data associated with the procurement, processing and distribution of goods. There are various discrete components of a supply chain, such as warehousing, inventory management, procurement, order management, carrier partnerships, and more.

Its importance can be judged from the fact that cosmetic giant Revlon had to file for bankruptcy¹² as it failed to repay $3.7 Billion debt mainly due to failure in addressing its supply-chain woes.

A century old discipline of  supply-chain analytics is becoming highly sophisticated due to evolution of ML and better data infrastructure. Achieving end-to-end supply chain analytics requires bringing information together across the procurement of raw materials and extends through production, distribution and aftermarket services. The goal of such integration is supply chain visibility: the ability to view data on goods at every step in the supply chain.

The days of reactive supply chain management are long getting over. With the help of Data Science and automated supply chain management tools, the inconsistencies in any stage can be predicted in advance or detected in early stage and remedial measures can be taken. Thus, by managing the supply chain risks, Data Science takes care of the entire business.

The management of supply chain risk requires working with  a large number of input attributes. From the manufacturing of an end product to its delivery to customers, the process involves risks at all its stages. The complexity and unpredictability of this task makes the job more suitable for data scientists as they first convert all inputs to data points, model various scenarios, anticipate market changes, minimize risks, optimize costs and efforts. They not only manage risks but optimize the supply chain so that waste is reduced and productivity is improved.

According to another classification, there are four categories of supply-chain analytics as below:

Following are some important benefits of data-driven supply-chain strategy.

(10) Marketing Mix Model:

Retail marketing mix refers to the variables that a retailer can use in marketing methods to design an effective marketing strategy. 9 such important variables are shown in diagram below. This is called 9-P Marketing Mix Model.

(1) Product Mix Analytics:

Product is the basic element of every organization. An organization is nothing but a collection of Products, People and Processes. Collection of all the products manufactured by company and sold by retailers is known as product-mix. Product mix refers to the length, breadth and depth of the products. The ‘length’  here is the total number of products that are present in the product-line while the ‘breadth’ refers to the number of product lines that are offered by the company and finally the 'depth' means the various varieties of a particular product in one particular product line. As an example, the diagram below shows Product-Mix for a kids clothing manufacturer.

(2) Price Mix Analytics:

As price is the single most important variable affecting retailing buying decisions, the entire retail organization is focused on determining optimal pricing for its products. The calculation of retail price should always be based on the markup and not the cost that is involved. Before any pricing decisions are made, a company must establish what is objective of its pricing strategy. Depending upon product, location, market condition, economy etc. there may be different objectives of pricing strategy as shown below:

The pricing strategy of a retailer must have following characteristics:

(1) It should be consistent 
(2) It must consider positioning of retailers, sales, profits, expenses,costs and ROI. 
(3) It must enable retailer to achieve desired cash flow,overall growth, and profitability.
(4) It must take into consideration market-condition, position of the product in market, customer-perception, various stages of a product life cycle through which the product is passing, the competitive strategy, the overall retail marketing mix and government regulations. Following are some popular pricing strategies.

Following are the components of price mix:

(3) Promotion Mix Analytics:

Ideally Price-Mix Analytics and Promotion-Mix Analytics must be undertaken together as they are inseparably related as the price-waterfall below shows:

What is promotion mix analytics?

Promotional pricing (e.g. Black Friday deals, boxing day deals, and New year deals) is a pricing strategy intended to attract interest and increase sales in the short term with primary intent of getting maximum sales in minimum time. 

Promotions are one of the most significant initiatives that can impact the retailer’s profitability and margins. However, as per HBR 55% Of Sales Promotions are ineffective. This underlines the need of using ML based promotional analytics. 

How Price-Sensitivity And Promotion-Affinity Are Closely Related?

To improve ROI from promotions the key metric is Promotion-Affinity. Promotion-Affinity score is closely related to Price-Sensitivity score. It primarily measures the impact of earlier promotions in terms of transaction- and customer-based success factors like (fig. below):

(1) Increased Revenue And Margin.
(2) Customers’ willingness to buy product without a promotion.
(3) Increase in the number of transactions due to discounts.
(4) Volume purchased of the product promoted.
(5) Basket product variance.

Due to such close relation between Pricing Analytics and Promotion Analytics companies should combine both scores together using tool like price-promotion matrix so that an optimal balance can be identified for each product being sold(Fig. below). Products are then placed into one of four quadrants of the matrix.

(1) High price sensitivity and high promotion affinity :

Price these items by setting the lowest prices and maximizing discounts.
(2) High price sensitivity and low promotion affinity: 

Used for necessity-type products. Keeping regular prices low, at a level below the recommended retail price but above what a promotional price would be.
(3) Low price sensitivity and high promotion affinity: 

Price these items close to the recommended retail or highest competitor prices but with discounts that lower promotional prices as much as possible.
(4) Low price sensitivity and low promotion affinity: 

Price these items close to their highest competitors’ prices, and they reduce or stop promotions on them.

A big online retailer, based on successful pilot project using this strategy, expects its sales revenue to increase by 3-5% and profits to grow by 2-4% over next 3 years.

What are benefits of Promotional Analytics

(1) Get approximate ROI of future promotions by building “what-if” scenarios
(2) Analyze data and reports in a single interface. No siloed data systems.
(3) Get the optimal promotion mix.
(4) Measure the effectiveness of past promotions with the right metrics.
(4) Forecast demand and never go out-of-stock during promotions.
(5) Understand the impact of promotions on the basket and trips.
(6) Get actionable recommendations for optimizing or running new promotions.
(7) Predict future sales.
(8) Build a data-backed promotion calendar with granular insights.

(4) Place Mix Analytics:

In simple terms, Place-Mix ensures that the availability of the product should be close to the place of consumption so that the customers can buy it easily. However a good Place-Mix is much more than just about geographical location and includes considering following aspects. 
 

If all the above are taken care of then the customer would happily buy the product. So, “Place-mix is an arrangement of distribution channels, both physical and non-physical, through which the product is made available to customers for purchase. It is the set of decisions a company undertakes to make the product accessible to its target customers at right spot, at right time, in right form,with right approach in the most cost-efficient manner”

The place mix is made up of two major components that can help a company meet its distribution goals and business targets. Physical Distribution  and  Channels Of Distribution

(A) Physical Distribution

Physical distribution is a set of activities concerned with the efficient physical movement of goods from producer to the consumer. It includes the four major activities as below:

(B) Channels Of Distribution

Channels refer to various essential linkages between manufacturers,intermediaries and end-users involved in the firm’s distribution network. It’s the network of individuals and institutions that helps to establish frequent contact with the customers and offer required information about the items to the customers. They also assist in financing by providing credit and helps in after-sales services. Thus, it’s that one entity that undertakes all risks related to the performance of the distribution function. 

Types of channels of distribution are shown in following diagram:

Choosing the right distribution channel is crucial as both the price and promotion strategy depend on the selected chosen channel. Thus, a manufacturer should consider the following factors while selecting a channel of distribution.
(1) Nature Of Product: 
Perishable items or heavy industrial items use short channels of distribution while for consumer-durable a long channel including agents, wholesalers and retailers is utilized.
(2) Nature Of Market:
The choice of the channel-structure is also decided by the number of consumers and their geographical concentration. Customer dispersion also plays an important role in deciding channel-structure. 
(3) Nature Of Company:
When selecting a distribution channel, a corporation must look into its market size, size of its major accounts and its ability to establish cooperation between intermediaries. Thus, a prominent business  or a corporation that wants more control can use short channel. Also, a financially strong corporation can negotiate and set up an exclusive distribution channel. 
(4) Intermediaries Consideration:
The manufacturer must select the intermediaries that offer the most outstanding marketing services such as storage, shipping, credit and packing who can guarantee the success of new products. 
(5) Environment Consideration:
While choosing a distribution channel, every corporation should pay attention to government rules and also with social standards and values. 
(6) Economic Considerations:
When the economy is robust, a corporation can use the long distribution channel with various intermediaries. However, the corporation should choose the quickest and cheapest means of distribution if there are many fluctuations in the economy.

After appropriate channel of distribution is selected the last missing piece of place-mix is choosing distribution strategies. It can be one of the three below:

 (11) Sentiment  And Emotion Analytics:

What is Sentiment Analytics?
Sentiment analysis (also known as opinion mining) is a ML and NLP based technique through which one can analyze a piece of text to determine the sentiment behind it. Using basic Sentiment analysis, a program can understand whether the sentiment behind a piece of text is positive, negative, or neutral. It focuses not only on polarity (positive, negative & neutral) but also on emotions (happy, sad, angry, etc.). This score can be more fine-grained into five (or any number of ) categories as needed as shown in fig. below which shows conversion of response on 5 point scale into five polarities.

Similarly polarities ‘Positive’ and ‘Negative’ can be converted into further fine grained categories and even emotion analysis can be performed as shown in fig. below: 

It is very important to analyze emotions as they are directly related to six pillars of customer experience as the table below shows:

Colin Shaw's Emotional Signature framework also suggests that by clustering customers (based on their emotion) into 'Attention Cluster', 'Recommendation Cluster' and 'Advocacy Cluster', businesses can derive long and short term value.

Why Perform Sentiment Analytics?

According to the survey,80% of the world’s data is unstructured. This data needs to be analyzed and converted to structured data. Then only algorithms can detect sentiments. Sentiment Analysis is also required to store this huge amount of data in an efficient and cost-friendly manner. Sentiment analysis can solve real-time business-issues (see fig below)

Techniques To Perform Sentiment Analytics?

The techniques are classified into three main categories: Lexicon Based( Or rule Based) Models, ML Based Models, Hybrid Models. Their further classification is shown in diagram below. 

Three main categories are described below:

(1) Rule/Lexicon Based Sentiment Analytics:

This algorithm is based on manually created lexicons that define positive and negative strings of words. The algorithm then analyzes the amounts of positive and negative words to see which ones dominate. Rules can also be set around other aspects of the text, for example, part of speech, syntax, and more. The architecture for a lexicon based sentiment analysis tool is shown below:

This approach is easy to implement and transparent when it comes to rules standing behind analyses. Disadvantage with this approach is it can not handle big data. Also rule-based approaches are limited because they don’t consider the sentence as whole and miss complexity of human languages. They also require regular update of rules. The steps in rule based sentiment analysis are shown below:

(2) ML Based Or Automatic Sentiment Analytics:

This method relies exclusively on ML techniques and learns on received data. It starts with creating a dataset that contains a set of  texts classified either as positive, negative, or neutral. Then learning begins on this data and continues as a semi-automatic process using chosen algorithm (Naive Bayes or Logistic Regression or Linear Regression or SVM or Deep Learning or LSTM or state of the art Transformers) learns on data until the  system achieves some level independence, sufficient  enough to correctly assess the    sentiment of new, unknown texts. The architecture for a ML based sentiment analysis tool is shown below:

The basic steps of ML based sentiment analysis are as below:

Advantage of this method is that it can analyze peta/exa byte scale data unlike the rule-based algorithm which can handle only small data. Disadvantage with this method is that accuracy depends on quality of data on which model learns is critical. Further, the method lacks explainability, meaning, it’s impossible to tell why a particular text was  classified as positive or negative.

(3) Hybrid Method For Sentiment Analysis:
This one combines both of the above mentioned algorithms and seems to be the most effective solution. It’s because it combines high accuracy provided by machine learning and stability and explainability from the rule-based approach.

Real-life Business Applications For Sentiment Analysis

(1)Voice-of-Customer(VoC)  or Voice-Of-Employee(VoE) Programs
E.g. Amazon is great at customer feedback analysis 
(2) Net Promoter Score (NPS) surveys 
E.g. This is done by almost all retailers by asking customer “How likely are you to recommend us to a friend on a scale 1-10?”
(3) Improved Customer Service Experience
E.g. TripAdvisor greatly polished it service and improved customer engagement using opinion mining.
(4) Feedback on Product Experience
E.g. UBER used sentiment analytics to discover if users liked new version of their app.
(5) Brand Sentiment Analysis
E.g.  Boeing did urgent crisis-management when  it figured in the top 5% of worst brands. because of a plane crash in China and publication of weak quarterly results.
(6) Social media sentiment analysis
E.g.,When sentiment of #UnitedAirlines hashtag showed 69% negative responses , company immediately took corrective actions
(7) Market research
E.g., Companies like Frank-Knight analyze industry data on the real estate market can reveal which city/area has potential of price-appreciation.
(8) Campaign Monitoring
In 2012, using sentiment-analysis, the Obama administration investigated the reception of policy announcements during the 2012 presidential election.
(9) Stock Market Analysis
When Moderna announced the completion of phase-I of its COVID-19 vaccine clinical trials. a strong rise in the stock price of Moderna was seen. But the stock stumbled after a it lost a patent. Using sentiment analysis, traders analyze such news in real-time and profit from market.
(10) Compliance Monitoring
Tools like ScrapingHub can help fetch compliance documents from large websites like 'Financial Conduct Authority'. and use intelligent classification to find the right content among millions of web pages.

Why is Sentiment Analysis challenging?

Human languages have many nuances and not all have been conquered by algorithms so far. Some examples are:

(1) Subjectivity:
E.g. ‘The laptop-screen is small for its prize’
(2) Contradictory emotions:
E.g. “The performance is great but TV is awfully expensive”
(3) Context:
Many algorithms see individual words or phrases but not context of entire sentence.
(4) Irony & Sarcasm:
E.g. The movie was so gripping that sleep gripped me in whole show."
(5) Comparison:
E.g."Much better than alternatives" does not give information about which alternatives were considered.
(6) Competitors: 
In survey done by company X, if a  comment about competitor company Y is "Company Y also offers great service", then actually this is a negative/neutral comment for company X. But for detecting that one needs to design a custom model.
(7) Emojis:
Sentiments expressed using emojis require lot of per-processing and contextual information as same emoji can have different meanings in different customer clusters.
(8) Idioms:
E.g. "not my cup of tea” will probably confuse algorithms
(9) Neutrality:
E.g. statement “This mobile is black” has no emotions.
(10)Ambiguity:
E.g. the word ‘wish’ in two statements below have positive and negative emotions respectively.
“I wish I had seen this product earlier.” and “I wish the delivery was done faster”
(11) Negation:
User may write “I can’t not buy another Apple Mac” then only LSTM or Transformer models can handle negation.
(12) Audio-visual Content:
Detecting sentiment in sound or image requires subtle analysis of tone or light/angle. This is much more difficult than text.

(12) Market Basket Analytics:

What is Market Basket Analysis?

Market basket analysis (MBA) is frequent item-set mining based technique. It is used for identifying customer behavior, buying patterns, and the relationship between products and content delivery by the retailer inside the store (or on their online shop). It works by analyzing past transaction data and looking for combinations of items that occur together frequently in transactions. 

The most well-known example of MBA is seen when one shops on portal of eCommerce giant like Amazon. At the early stage of checkout ‘Frequently bought together’ suggestion pops up as below. This is derived by analysis of data of past millions of transactions.

So, from customer's point of view MBA gives an experience of adding items to shopping cart as shown below:

However from data-analyst perspective, MBA is  study of customer buying habits which clusters associated items in clusters (or baskets) so that better recommendation can be generated for next transactions. See fig below:

How Association Rules Are Obtained From Transaction Data?

We take example of a dataset of an online store from UCI ML repository⁹. The dataset includes 406,829 records and 10 fields. Analysis (code not shown) revealed following facts:

(1) Most orders occurred between 10:00–15:00 Hours. 

(2) Maximum number of people purchased <10 items per invoice.

(3) We Could find top-10 best seller items and their count.

(4) A bar plot of 20 most frequent items bought

Before further discussion we need to define three important metrics used in Market Basket Analysis. These are Support, Confidence and Lift. 

(A) Support: 

If we have made an association rule from data in a restaurant  that "customers who buy sandwich and cookies are also likely to buy drink"  i.e. IF {sandwich, cookies} THEN {drink} or IF {Antecedent} THEN {Consequent} ; then the probability that the antecedent event will occur is the support of the rule. The support of an item helps to identify keystone products. Hence, if a sandwich and cookies have high support, then they can be priced to attract people to the store.

(B) Confidence: 

The probability that such customer will also buy a drink is called the confidence of the rule. Confidence can be used for product placement strategy and increasing profitability. Placing high margin items near associated high confidence (driver) items can increase the overall margin on purchases. 

(C) Lift: 

The lift of the rule is the ratio of the support of the left-hand side of the rule (sandwich, cookies) co-occurring with the right-hand side (drink), divided by the probability that the left-hand side and right-hand side co-occur if the two are independent.

Now coming back to online store example, we used apriori algorithm with support=0.001 and confidence=0.8 and generate association rules using R. 

rules <- apriori(tr, parameter = list(supp=0.001, conf=0.8))
rules <- sort(rules, by='confidence', decreasing = TRUE)

One can also use other algorithms like AIS,STEM or FPGrowth for mining rules instead of apriori algorithm used above. We got 89,697 rules, and statistics for  Support, Lift and Confidence. Just showing top 10 rules below:

Scatter plot of these 10 rules with Support, Lift and Confidence on 3 axes is shown below:

Finally the graph of top 10 rules showing associations they represent was obtained:

What are applications of MBA in retail? 

The disclosure of “Correlation Relationships” among huge amounts of transaction records can help in many decision-making processes such as:

(1)Selective marketing.  Better Advertising and Promotions.
(2)Planning ledge space.
(3)Designing optimized store Layouts.
(4)Designing of catalogs, UI/UX and improving CX.
(5)Using correlations between items for cross-selling
(6)Increase customer-engagement and customer-comprehension.
(7)Boosting sales, Precise Targeting and Improved ROI
(8)Improving customer experience
(9)Optimize marketing strategies and campaigns
(10)Identifies customer behavior and pattern
(11)Optimizing inventory for each store or warehouse.
(12)Recommendations on Netflix,Amazon Prime, Spotify  
(13)More In-Store Traffic.

What are types of market basket analysis?

(13) Customer LTV Analytics

What is CLV or CLTV?

Customer lifetime value (CLTV, or CLV) is the total revenue a retailer is likely to receive from an individual customer during their relationship with company. It represents a customer’s value to retailer's business over a certain time period.

Why is CLTV important metric for retailers?

(1)With CLTV analysis,  a retailer can quantify the value of its customers.  Particularly in  for eCommerce  the net margin averages just 4.6%, LTV is a make-or-break metric that can not be ignored.

(2) CLTV gives clear idea about direction that a company is taking over long time. A growing CLV means a company is doing well—customers are happy and will be giving more money over the lifetime. But, a declining CLV means that company is getting less money from each customer and needs to fix something fast.

(3) CLV is an easy metric to look at to see the overall health of a product in terms of both revenue and customer retention. 

ARPU (average revenue per user) doesn’t show is whether or not the customers are going to keep paying that/higher amount in future. 

Similarly, customer retention won’t show if  the customers who stay with company would paying more or less in future. 

Revenue retention will show if the customers are upgrading or downgrading—it won’t show whether the number of customers is fluctuating.

CLV gives a more robust picture of present state of company as well the revenue it can expect over the course of a customer lifespan. Its importance is described in depth at 9 real-life case studies¹¹ of companies like AMAZON, BONOBOS,CROCS,HEAR AND PLAY MUSIC,NETFLIX, KIMBERLY-CLARK, ZAPPOS, U.S.AUTO-PARTS and STARBUCKS.

How is CLV Calculated?

There are five main methods. First three are shown in fig. below:

(1)Historical Approach

Suppose 20 customers brought $1,240 in profit over a 3-month period. So,ARPU (3 months) = $1240 / 20 = $62 and ARPU (12 months) =  $248 per year per customer. By historical approach this $248 is itself CLV. The historical approach is valid only if all customers have similar preferences and stay with the company for the same period of time.
(2)Cohort Based Method)Predictive (ML Based) Method

A cohort is a group of customers who have similar characteristics and made their first purchase during the same month. Using cohort analysis, analysts calculate the average revenue per cohort (ARPC) instead of per user. E.g. In fig. above we have calculated the ARPC per month for two cohorts named 'Customers from January 2018' and 'Customers from March 2018'. In this case, purchases differ from month to month for two cohorts and the March cohort will likely go silent from Month-4.
In addition to calculating CLV, cohort analysis can help you find the number of loyal clients, improve CLV by finding the points where purchasing drops off, and accurately assess ad-campaign performance.

(3)Predictive (ML Based) Method
This is the most complicated but accurate method for calculating CLV. It requires calculation of T, AOV, GM,ALT and CHURN RATE. If 120 transactions occur in 6 months then avg transactions/month T=120/6 = 20. Similarly if total revenue of November month is $12,000 and there are 20 orders then Avg Order value (AOV) = 12000/20 = $600. The formulae for calculating GM, ALT and CHURN-RATE are shown in fig.above.

(4)Retention Rate And Discount Rate Based Method

The formula for this method is as below. Where GML is calculated using formula:

GML = Gross Margin (%) × Average Total Revenue per Customer

And R is calculated as below:

(5)Google Analytics Based Method

The LTV report in Google Analytics can tell you data on users’ behavior at their first stages of interaction with the company based on page views, goals, events, and trends. Basically, you just click the buttons to generate a report and Google Analytics breaks it down for you.

What Are Benefits of CLTV Analytics?

Understanding CLTV enables decisions around:

(1)Managing customer-relationships, time-intensive customer support and costly returns.
(2)Targeting customers for loyalty or “VIP shopper” programs?
(3)Maximizing marketing ROI.
(4)Measuring business viability: High CLV is a sign of product market fit, brand loyalty and viable business.
(5)Providing clarity on CAC. E.g., CLV should be at least 3 *CAC.

(6)Helping to achieve steady growth: CLV is a great metric to monitor and optimize for growth.
(7)Practicing Value-Tier-Segmentation. Move customers to higher value segment. 
(8)Preventing customer churn
(9)Measuring Your Ad Performance

Why must CLV Analytics be re-emphasized?

Although documented evidences of CLV analytics are well-known,a UK study¹³ found that only 34% of the marketers they surveyed were “completely aware of the term and its connotations.” And only 24% of respondents felt their company was monitoring CLV effectively. At the same times 9 real-life case-studies¹¹ of some great companies ow world shows how they achieved tangible benefits by CLV analytics.

(14) Fraud Detection And Prevention

Why retail fraud analytics?

Retail transactions are fraught with danger of various kinds of fraud. E.g. in US alone the fraud and losses from other retail “shrink” totaled $61.7 billion in 2019¹⁴, up from $50.6 billion the year before. In the year 2022¹⁶, the number neared $100 billion. The problem is also getting worse  for eCommerce space¹⁵ as the chart below shows.

Retailers are prioritizing new resources to safeguard their brand with 44.5% indicating loss prevention as an area of investment, 60.3% increasing their technology budget and 52.4% are increasing their capital and equipment budget. Many are investing in technologies like RFID, computer vision at POS and license plate recognition. All these solutions need data analytics.

What are types of retail fraud?

Although the fraudsters are becoming increasingly innovative there are some well-known categories of retail fraud as shown below:

One more classification¹⁷ divides all retail frauds into 12 categories as follows:

Clearly just POS analytics or cameras are not enough to tackle all these varieties of frauds. Retailers must use advanced predictive and preventive analytics using rich data that is characteristics of retail space to stop fraud before they are attempted.

How Machine Learning and AI can help in tackling retail frauds?

(1)Predictive Analytics 

AI fraud-detection services – aided by predictive analytics –  can integrate with retail payment processing systems at the POS. The ML models empowering fraud-detection services learn to spot patterns associated with fraud. 
Examples of predictable fraud activities:
Employee Theft – Discounts and write-offs at the point of sale can hide fraud. For example, irregular transaction patterns associated with certain item types or individuals could signal that a cashier is easing inventory out the door for friends, family, or a pseudonym. Trained algorithms can flag transactions matching high-risk patterns for review by management. 
Identity Theft – Transactions far away from normal shopping areas may signal fraud. Detailed customer profiles, enhanced by transaction geolocation data, empower identity theft protection services.
False Returns – AI algorithms can learn to identify patterns using transaction data in real time for high-variance activities like false returns.

(2)Anomaly Detection

Applications for detecting deviations from normal payment activity compared to historical data over a time period which can be scaled differently for separate investigations. 

These include prescriptive analytics systems which recommend a “next best action” immediately following detection. They function especially well with a steady stream of fresh training data from recent transactions.

(15) Augmented Reality. Better CX

What is augmented reality (AR)?

Augmented reality is the blending of computer technology and science to create an entirely new medium through which users can interact with the digital information around them. Augmented reality describes the use of information technology to augment reality. 

Why is it necessary to adopt augmented reality(AR) in retail?

Many feel AR is a useful but far-fetched solution, particularly in retail.However, after the launches of Apple's ARKit and Google's ARCore and after several businesses have already successfully deployed and benefited from it, AR is no longer a future possibility for retail -- it's here. In fact, 75% of consumers¹⁸ now expect retailers to offer an augmented reality experience. 

Research, by mobile app developer Apadmi, based on the surveyed up to 2,000 consumers who use retail apps shows us trends. The survey explored the main expectations from Augmented Reality apps and their further popularity.

(A)29% of surveyed expecting retailers would invest more into both AR and VR platforms.
(B)33% of them are ready to use improved AR benefits to preview the size of an item in real before buying it.
(C)29% of them want to check how they can use stuff before purchasing it.
(D)25% of them would use AR features to see all differences of variations of product colors or designs, before choosing one.

Thus, if retailers do not adopt AI, customers coming with google glasses will force them to adopt it as is clear from the report¹⁹ that the global AR market is projected to grow from $6.12 billion in 2021 to reach $97.76 billion by 2028. 
How AR benefits retail-stores?

Which ML Techniques Are Used In AR?

AR would not be possible/viable without some techniques in ML:
(A) Object Labelling
(B) Object Detection And Recognition
(C) Text Recognition And Translation
(D) Automatic Speech Recognition

Any Real-World Use-Cases Of Successful Use Of AR In Retail?

Several reputed brands have already deployed AR solutions in their businesses and seen big jumps in revenue, profit, customer-loyalty and brand-value. Some such examples are^18,21,22:

How Should I decide if my retail-segment is ripe for AR? 

In the chart below segment-wise penetration of AR in different retail-segments is shown. Depending upon your segment get ready for AR revolution!

REFERENCES

1. https://www.mckinsey.com/business-functions/growth-marketing-and-sales/how-we-help-clients/clm-online-retailer

2. https://www.unboundb2b.com/blog/b2b-marketing/customer-success-for-sales-reps/#

3. https://hbswk.hbs.edu/archive/the-economics-of-e-loyalty

4. https://blog.hubspot.com/sales/cross-selling

5. https://www.gartner.com/smarterwithgartner/improve-the-supply-chain-with-advanced-analytics-and-ai

6. https://www.mckinsey.com/~/media/McKinsey/Industries/Semiconductors/Our%20Insights/Smartening%20up%20with%20artificial%20intelligence/Smartening-up-with-artificial-intelligence.ashx%20str%209

7. https://refed.org/downloads/Retail_Guide_Web.pdf

8. https://doi.org/10.1155/2021/7160527

9. http://archive.ics.uci.edu/ml/datasets/online+retail

10. https://www.criteo.com/wp-content/uploads/2019/04/Criteo-CLV-Report-2019.pdf

11. https://barnraisersllc.com/2017/01/15/case-studies-customer-lifetime-value-clv/

12 https://www.glossy.co/beauty/mounting-debt-and-supply-chain-problems-lead-revlon-inc-to-bankruptcy/

13 https://www.bloomreach.com/en/blog/2021/customer-lifetime-value-guide

14. https://nrf.com/media-center/press-releases/retail-shrink-totaled-617-billion-2019-amid-rising-employee-theft-and
15. https://www.statista.com/statistics/1273177/ecommerce-payment-fraud-losses-globally/

16.https://nrf.com/media-center/press-releases/nrf-reports-retail-shrink-nearly-100b-problem

17.https://www.vendhq.com/blog/retail-fraud/

18. https://blog.hubspot.com/marketing/augmented-reality-retail

19. https://www.fortunebusinessinsights.com/augmented-reality-ar-market-102553

20. https://thinkmobiles.com/blog/augmented-reality-retail/

21. https://econsultancy.com/14-examples-augmented-reality-brand-marketing-experiences/

22. https://www.forbes.com/sites/bernardmarr/2021/09/13/10-best-examples-of-augmented-and-virtual-reality-in-retail/?sh=6c4659366626