6 Ways to Increase Retail Margin Using the Data You Already Have

Retail is being shaken hard today: taxes are changing, logistics costs keep rising, competition is intensifying. At the same time, customers are becoming more demanding, while stockouts and write-offs continue to eat into profits. But the paradox is that almost every retail chain already has everything it needs to grow.

• sales by stores and products across the network
• promotions and pricing
• CRM and loyalty program data
• website user behavior
• supply data
• staffing and scheduling data

Large retailers in the United States, Europe, and Asia have been using predictive models for years. But today these methods are becoming more accessible even for smaller players.

At dev.family, we have been building solutions for foodtech, e-commerce, and retail for many years, working with companies in the United States, the UK, and Europe. We see the same pattern over and over again: working with data consistently increases margin even for those chains that consider themselves “fully optimized.” But unfortunately, such chains are rare. Most still do not know how to work with data or even collect it properly.

That is why we decided to compile six directions that deliver results the fastest and share the experience of industry leaders and real implementation cases.

Would you like to compare your chain with the cases of market leaders? We can provide a consultation with no obligations.

Max B. CEO

Write in LiveChat

Write in LinkedIn

A Deloitte study shows that machine learning reduces forecasting errors, which directly affects the reduction of inventory and the increase in sales.

Inventory management is a balance between two extremes:

• ordering too little → stockout → lost sales
• ordering too much → frozen capital → write-offs

When you have only a couple of stores, this might be rare. You know exactly who buys what and where. But in large chains, errors accumulate quickly. Losing 3–5% of receipts due to the absence of the required item equals up to $200,000 in lost monthly revenue for an average chain.

PEPCO, Danone, Afresh, Migros — European and American retailers work with ML models that:

• forecast demand for item T in store X on day Y
• account for seasonality, weather, promotions, competitors, item availability, assortment
• adjust the forecast according to the cost of errors (under-ordering vs. over-ordering)

The results, confirmed by data from 2022–2024 and consistent with what we observe in our projects:

• +3–5% in sales at the same procurement volume
• –10–15% in excess inventory
• –20–30% in write-offs (especially in categories with short shelf life)

If you do not have weather data or visibility into shelf layouts, you can still build a simplified model. It will still deliver noticeable improvement.

According to McKinsey, personalization increases retailers’ revenue.

Customers are tired of mass offers and digital spam. Customers remain loyal when a brand “understands” them and makes very personal offers: “We noticed that you have been buying diapers regularly. Congratulations on the new addition to your family. Here is a discount on baby products.” That is it. The customer stays with you — you showed care and attention. This is why it is so important to store information about:

• what they buy
• how often
• which alternatives they prefer
• what they pay attention to on the website

A method that allows the system to “guess” a customer’s interests even if they have never purchased a specific item. For retail, it helps understand:

• which products interest a specific customer
• which categories they are likely to buy
• which items should be recommended

For example, if a customer buys baby food and personal care items, the model understands that they may also be interested in diapers or wet wipes even if they never searched for them. 

Collaborative filtering analyzes similar customers: “Customers similar to you usually buy this.” It focuses not on products but on user behavior. The method is ideal when:

• a customer has few purchases but behaves similarly to others
• recommendations must be built quickly using minimal data
• personalized newsletters, promotions, and website modules are required

For example, if a user buys shampoo and hair masks, and thousands of similar customers also buy heat protection sprays, the model will recommend them.

These models predict when a customer will need a product again based on how often and in what quantity they buy it. They are especially important for drugstores, FMCG, and categories with cyclical consumption:

• household chemicals
• home goods
• cosmetics
• pet food
• baby products

If a customer buys a 3 kg pack of laundry detergent every 40 days, the system knows: “In a week, we need to remind them to repurchase.” At the right moment, the retailer can send:

• a push notification
• a personalized coupon
• a recommendation on the main page

Search results on the website stop being the same for everyone. They begin to take into account:

• purchase history
• customer preferences
• past clicks
• categories the customer was interested in
• product margin and priority

90% of customers make a purchase within the first five search results. If search shows irrelevant items, the customer leaves. Personalized search helps:

• find relevant items faster
• recommend alternatives
• increase the average check
• promote higher-margin products

For example, if a customer often buys Korean cosmetics, the search term “cream” will show Korean brands at the top. If the customer buys premium items, the search results will adapt accordingly.

• +5–10% in sales
• –20% churn among loyal customers
• higher average check and LTV

Even without a large digital ecosystem, loyalty program data is enough.

According to IBM, automatic anomaly detection helps retail reduce operational losses.

A stockout in one store, an incorrect price on a shelf in another, a promotion upload error in the system, a broken search function in the app — retailers often notice these only days or even weeks later.

Anomaly detection identifies:

• sudden drops in product sales
• abnormally low store performance
• unexpected online user behavior patterns
• data errors and incorrect promotions
• local competitor activity

PwC reports that proper workforce planning reduces labor expenses by 5–10% without harming customer experience.

Staff is allocated in old-fashioned ways. The same number of cashiers is assigned to all stores, or schedules are planned without considering foot traffic. Often, roles overlap within a single store.

They create a model that:

1. analyzes traffic, seasonality, sales, store size, and store type
2. determines the optimal number of employees for each role
3. identifies stores with overspending and explains why
4. compares them with similar stores

Accenture highlights in its study that route and transportation load optimization using ML algorithms reduces logistics costs.

Unless we are talking about the largest market players, it is safe to say that every retailer faces the following problems:

• routes are planned manually
• trips are assigned based on dispatcher experience
• vehicles are underloaded
• delivery points are poorly grouped
• assortment growth increases supply frequency

✍️ Danone, Migros, and many FMCG brands use:

This method groups stores or delivery points based on logically connected features. It helps avoid planning each point separately and instead unites them into optimal logistics groups, making route planning simpler and the entire delivery process more efficient.

Examples of features: geography, delivery volume and frequency, type of goods.

For example, a chain has 150 stores. Intuitively, they are divided into four zones. After clustering, the model showed that:

• some stores in the southern zone are more efficiently served from the eastern warehouse
• two remote stores should be served by a separate mini-route
• two large stores form a high-load cluster requiring a dedicated trip

Classic route planning (“by map”) is often built on the dispatcher’s intuition. But once you exceed 20–30 points, the number of possible combinations becomes billions. Finding the optimal solution manually is impossible.

Metaheuristics are powerful optimization algorithms that search for the best route not through brute force but through smart approximation using mathematical strategies.

For example, they generate many possible routes, select the best ones, “cross” them, and gradually improve them. The result is a shorter and cheaper route with every iteration.

Greedy algorithms follow a simple principle: “At each step, choose the cheapest action.”

For example, select the nearest point, load the optimal volume, then select the next nearest point. It is fast, cheap, and works very well after clustering.

These algorithms:

• reduce total mileage (often by 7–12%)
• reduce the number of trips
• reduce travel time
• account for constraints (time windows, load, weight, schedules)

Savings on fuel, depreciation, and driver salaries can reach tens of thousands of dollars per month.

This method helps use vehicles as efficiently as possible: load exactly what is needed.

For retail, both overload and underload are equally dangerous:

• underload → more trips → higher costs
• overload → risks, fines, delivery delays

The algorithm analyzes the volume and weight of orders, product dimensions, vehicle capacity, temperature requirements (if any), route, and time windows. It then offers solutions when:

• vehicles run half-empty due to uneven orders
• several small stores can be combined into one trip
• a large order should be split between two vehicles
• certain product categories cannot be transported together

According to Forrester, systematic work with reviews more than doubles revenue growth rates.

Most chains learn about problems from social media or after a spike in negative reviews. And it must be admitted: not everyone knows how to respond properly. Some ignore reviews, some stay silent, some reply with a generic phrase.

We had a project where we created a model that parses reviews across all platforms, routes them through a ticketing system, and helps generate responses.

Malpa Games. Customer feedback management software for a mobile games publisher

More in our case study

View case

• automatically collect reviews from all platforms
• analyze sentiment
• identify common complaint themes
• detect negative spikes
• check whether the situation improved after fixes

A pilot produces noticeable effects within 6–8 weeks. As a result, you reduce:

• –20–30% negative reviews
• –70–80% analyst time

And as a bonus, you gain the ability to respond to problems in the network quickly.

These models do not require perfect infrastructure. They work with the data every chain already has.

That is why pilots usually take 6–8 weeks and do not require ERP changes, which is confirmed by industry standards and our projects with retail in the United States and Europe.

If we combine the effects across the six directions, even partial implementation delivers:

• Sales growth of 3–10%
• Logistics cost reduction of 5–12%
• Labor cost reduction of 5–10%
• Inventory cost reduction of 10–15%
• Write-off reduction of 20–30%
• Customer churn reduction of 20%

We highlight four stages:

1. Data collection and integration: sales, promotions, loyalty, logistics, workforce, online behavior
2. Model building and calibration, considering the cost of errors and current business processes
3. Pilot on a limited scope: 1–2 categories or several stores
4. Effect evaluation and scaling. Even during the pilot, the chain sees real numbers: forecast accuracy, savings, sales uplift, and loss reduction

Not sure where to start? We can run a quick data audit and propose the best scenario.

Max B. CEO

Write in LiveChat

Write in LinkedIn

Most retail chains are sitting on gold but use only 5–10% of it.

Machine learning is not about “complex technologies” — it is about a simple outcome: fewer errors, fewer losses, more profit. And you can start tomorrow — with the data you already have.