Produktionsprognosemodell: ML-Lösung für das Unternehmen Neue Produkte

Tags:

Business Intelligence (BI)Data Engineering

About the project

Forecast that balances demand and spending

The Novi Produkty Group of Companies is one of the largest producers of beverages and snacks in Ukraine, operating on the market for over 21 years and included in the Register of the country's largest taxpayers. The company exports its products to 23 countries worldwide, has over 1,500 employees, and serves over 106,000 retail outlets. The brand portfolio includes more than 220 items, including SHAKE, NON STOP, PIT BULL, REVO, KING’S BRIDGE, APPS, EAT ME, and PRYRODNIE DZERELO. The company has implemented a certified food quality and safety management system. The group is also actively involved in charitable activities.

The client's problem

With the growth in production volumes and expansion of sales markets, the company faced the risk of an imbalance between demand and production volumes. Intuitive planning led to either overproduction or product shortages: both scenarios resulted in financial losses and lost market opportunities.

Key challenges:

excess production created illiquid inventories and write-off costs;
product shortages led to a loss of sales volume;
The lack of a forecasting tool complicated planning in conditions of fluctuating demand.
historical sales data was of poor quality, in particular due to the absence of zero balances;
It was necessary to reduce losses without compromising production flexibility.

To solve the problem, the company sought a tool that would allow it to forecast production volumes taking into account demand, seasonality, and data limitations.

Case information

Built a predictive model that synchronizes demand with production

The tool has become part of the company’s operational and strategic planning.
Our approach:
It all started with participation in a tender: we built the first version of the model based on a limited set of features, which, despite its simplicity, showed the highest accuracy among all participants.

After launch, we focused on adapting to real business: we held a series of workshops with the client, identified features that influence demand, and tested several ML architectures. As a result, we implemented an ensemble model with an accuracy of over 80%.

The difficulty arose due to unrecorded zero balances at some retail outlets: the system confused the absence of sales with the absence of demand. Approximation methods corrected these distortions and significantly improved the forecast.

Work results:

We have implemented a forecasting model that allows us to plan production more accurately:
thanks to the ensemble ML model, a prediction accuracy of 80% has been achieved;
reduced risks of overproduction and shortages;
data inaccuracies corrected: absence of zero balances taken into account;
the model helps determine how much to produce and when;
The foundation has been laid for scaling to other product categories.
Instead of reacting to shortages or surpluses after the fact, the company began planning volumes in advance, thus moving from a reactive approach to proactive production management.