Machine Learning in Retail:
a Practical Guide to Machine Learning Projects

Statistical models;Machine learning

Statistical models are about finding relationships about variables and the significance of those relationships.;Machine learning models are designed to make the most accurate prediction possible about future events or outcomes. To build a statistical model, statisticians need to understand the relationships between data to come up with parameters significant to analysis.;Machine learning requires no prior assumptions about underlying relationships between variables. A model is fed with data and the algorithm discovers patterns in data to make predictions. Statistical models are static. Once built, they are usually renewed reactively when a change in external or internal parameters has already affected their performance, or on a predefined schedule (i.e. quarterly or annually).;A model is constantly and proactively retrained based on new data, which in turn boosts its quality and precision. Thus, the performance of a machine learning model improves with time. Statistical models utilize low dimensional datasets. One doesn’t need big data to build a statistical model.;To successfully apply machine learning, the size of the dataset starts from tens of thousands of entries (i.e. sales). However, for some tasks, the approach requires fewer amounts - months rather than years - of data. Classical statistical rule-based systems analyze a maximum of 50 parameters.;Machine learning models can account for hundreds of parameters including not only historical sales data but such influencing parameters as advertising, product placement within a store, substitutes on promotion, public holidays, and weather. Statistical models need complete and accurate data to build upon. All the data has to be cleaned and verified.;Machine learning is more flexible and resilient to missing and erroneous data.

Business use case of machine learning;How does it work?;What data to use?;How to measure business value?

Demand forecasting and replenishment;A machine learning model predicts future consumer demand and recommends the optimal amount of each SKU to be ordered to reduce write-offs without affecting on-shelf availability.;Sales history, Product characteristics, Price dynamics, Sales promotions, Store characteristics, Replenishment schedule, Minimum delivery unit and minimum delivery batch size, Weather;By a reduction in overstocks and out-of-stocks. Such business metrics as write-offs and on-shelf availability can be used. Safety stock optimization;A machine learning model makes intelligent recommendations to optimize safety stock for seasonality, upcoming holidays, and sales promotions, improve turnover, and keep high levels of on-shelf availability.;Sales history, Product characteristics, Seasonality, Promotions calendar, Holiday calendar, Store characteristics, Replenishment schedule, Minimum delivery unit and minimum delivery batch size;By a reduction of safety stock. Such business metrics as stock level can be used. Intelligent pricing;A machine learning model makes intelligent recommendations on price optimization taking into account external parameters such as time of day, day of the week, weather, sales promotions, on-shelf availability, and cross elasticity.;Sales history, Sales history of substitutes, Competitors’ prices, Price dynamics, Price sensitivity, Ongoing sales promotions, Seasonality, Weather;By an increase in sales. Such business metrics as average bill amount or margin can be used. Churn rate can be monitored in the long run. Promotion forecasting;A machine learning model predicts demand for items on sales promotions and makes intelligent recommendations on promotion parameter optimization according to a defined KPIs.;History of sales promotions, Sales history of substitutes, Price dynamics, Product characteristics, Product placement within a store, Store characteristics;By a reduction in overstocks and out-of-stocks. Such business metrics as write-offs, on-shelf availability or stock level can be used. Personalized offers; A machine learning model personalizes offers for each customer and optimizes offers for the desired KPI: average bill amount, frequency of purchase, profitability, etc. The offers may include shopping suggestions, discounts, customized online-store layouts, prepacked baskets or ads.;Service usage data, History of items purchased, Logs of apps and web-sites, History of communication and responses;By an increase in sales and improved customer experience. Such business metrics as average bill amount, frequency of purchase or margin can be used. Customer experience can be monitored via customer surveys and interviews.; Staff load prediction; A machine learning model predicts future demand and staff load on a granular level (i.e. hourly demand forecast on a store level) and recommends optimal staff allocation (i.e. of cashiers or couriers).;Sales history, Store traffic, Shift schedule, Man-hour costs;By increased staff availability and decreased costs associated with personnel. Such metrics as availability, time to delivery, or time in a check-out line can be used.

Step of the pilot project;Description

Identify the business use case and goals of the project: what do you want to improve?;Optimize the amount of fresh milk (including safety stock) to be ordered from the supplier to decrease write-offs without affecting the level of on-shelf availability. Identify success criteria: how will you measure success?;The decrease in write-offs without affecting the level of on-shelf availability. Identify and prepare available data: what data do you have to solve the task?;Extract available data from existing resources including sales history, product characteristics, price dynamics, ongoing sales promotions, store characteristics, next delivery date, minimum delivery unit, and minimum batch size. Train machine learning model on available data;Build two models: 1) a predictive model that forecasts the demand for fresh milk using historical data, 2) an optimization model that: recommends the optimal level of safety stocks for each SKU and hence the amount of fresh milk to be ordered to decrease write-offs while keeping availability levels high. Test the models;First, verify the prediction model by applying it to past period data that was not used to train the model, and comparing predictions of the algorithm with the actual sales of fresh milk. Then, test the optimization model. Integrate it into the existing demand forecasting process. The algorithm recommends the optimal amount of fresh milk to be ordered, and the order is placed into the retailer’s replenishment system. A/B testing is performed by comparing the results of the experimental batch to the control group. Measure results;Measure the decrease in write-offs reached due to the recommendations. The resulting cost reduction can be used to estimate the value of applying machine learning. Deploy the solution and update the model with new data;Deploy the solution full-scale and integrate it into the existing demand forecasting process. To maintain the quality of the forecast and continuously learn from data, the models need to be updated based on new data. Identify the next steps and new use cases: what else can be improved?;Define further optimization efforts and new uses cases to test (i.e. price optimization).