Data analysis for sales forecasting and optimization of the product matrix for the Purina brand

Client

Purina is a brand of pet food and treats, a subsidiary of the food products manufacturer Nestle.

Context

As brands cannot display all their products on store shelves, they often choose to showcase only their top products with good profitability or high off-take rates – the rate at which items are sold off the shelf.

Given the numerous products and the constant updates to sales data, the product matrix needs to be regularly updated. Despite the complexity of this process, in many companies, it’s done manually, using tools like Excel, relying on experience and intuition.

Task

Simplify the planning and maximize the automation of this process by utilizing intelligent data analysis through the application of machine learning.

Solution

1 We have developed a tool called Sales Forecaster. This solution is based on machine learning algorithms that, when fed with historical sales data, will create forecasts and generate a product matrix for trade marketers and analysts.

2 By default, the service takes into account historical data from past periods. However, there is the option to upload more recent sales documents for new periods (in our project, this included data from the past three years). Users can also define necessary variables: the number of positions requiring forecasts, desired category proportions – for instance, 60% cat food and 40% dog food – and the breakdown between wet and dry food, and so forth.

3 The Sales Forecaster has the ability to create forecasts based on the intended goal: either increasing sales or enhancing profitability. Users can adjust filters to compare graphs and observe how the requested parameters dynamically recalibrate.

The machine learning model utilized by the service takes into consideration various factors, including geography, sales at specific locations or chains, seasonality, data from other products in the category, the introduction of new products, planned brand promotional campaigns, or recent promotions by competitors, and even weather conditions.

4 Sales planning in Sales Forecaster can be done for one month, three months, six months, or a year, allowing you to create short-term strategies or plan sales in a long-term perspective. Furthermore, the tool calculates both the error percentage and profit percentage based on historical data.

Dmitry Pukhov

CTO Cleverbots

The core of the service relies on Prophet, an open-source software developed by Facebook. It’s specifically designed for forecasting the behavior of time series data in the future. This “oracle” looks for patterns in time series and can consider seasonality and trends. Prophet’s main feature is its flexible approach and the ability to account for shifts in behavior during weekends and holidays, and if available, even changes in customer activity by days and hours.

Alex Kapustin

Head of project office

Sales Forecaster not only generates a new assortment matrix but also displays on a graph which SKUs need to be replaced and the profit that would be gained in such a scenario. Comparing the model’s forecast with actual data after implementing the tool is interesting – the difference in assortment and the growth it generates is very clearly visible.