Food industry

Case : Demand Forecasting for Cheese Production



Our client, a large national cheese producer in Central Europe, approached Marrbery for assistance due to inaccurate demand forecasting for the company’s products, which was losing approximately $200,000 annually due to excess spoilage and shortages of certain items resulting in lost sales. The forecasting models had an error rate of more than 30%.

Situation Description

MSM is a leading cheese and dairy products company. It is a fully integrated company engaged in the production, marketing and sales of its own cheese brands. It also owns the entire supply chain, including dairy production, laboratories, packaging lines and its own distribution network to retail outlets. With large production facilities spread across different countries and distribution agreements with major retailers, MSM is able to offer a variety of products in the dairy segment.

MSM is currently evaluating new products such as lactose-free cheese and Biolife vegan cream cheese. While cheese products are typically targeted at a broad audience, these new products may appeal to certain market segments. “Biolife” is designed with a focus on consumers with lactose intolerance, which is one of the most common food intolerances in the world. In addition, vegans who eliminate animal products from their diet may also be a key target audience for “Biolife.”

MSM expects these new products to capitalize on the growing popularity among consumers focused on inclusive and healthy lifestyles.

  • MSM produces 20 types of cheese with a total volume of 50,000 tons per year

  • 60% of production is supplied to retail chains, 40% to independent stores

  • Due to forecast errors, monthly losses amounted to $15-20 thousand.

  • Average forecast error by assortment – 38%

Marrbery study

The Vice President of Marketing at MSM asked us to analyze key aspects related to the forecasting and launch of the Biolife product.

Our task – provide a comprehensive assessment of the factors affecting the success of product introduction to the market, as well as identify MSM’s internal reserves to support this process. Our team conducted the following stages of analysis and development:

  • Sales history analysis: We examined sales data for each product and distribution channel for the last three years in detail. This allowed us to identify trends and seasonal fluctuations in demand.


  • Statistical forecasting models: We applied advanced statistical methods for time series forecasting, including ARIMA, SARIMA and Prophet. As a result, we achieved an error rate of 22%, which is acceptable.


  • Machine learning system development: Our experts developed our own machine learning-based forecasting and planning system. This system allows for more accurate forecasting of product demand as well as automating the planning process.


  • This system allows MSM to improve demand management and optimize the production plan. We are confident that the implementation of this system will bring significant results and increase the efficiency of the company’s cheese production activities.

                                          Exhibit 1

Our team developed a machine learning system for demand forecasting using the company’s historical sales data for the past 5 years. We adapted linear regression, neural network and gradient bousting algorithms to the specifics of the client’s data. This allowed us to create an accurate and reliable demand forecasting model. In addition, the system has been successfully integrated with the company’s ERP, ensuring efficient data management.

It is important to note that our approach not only considered historical data, but also took into account current market trends and consumer demand characteristics and a multidisciplinary approach in analyzing and forecasting data. 

Solution development

It is revealed that current forecasts are based on subjective expert assessment. This leads to a high forecast error

We hypothesized that applying machine learning models based on actual sales data will significantly improve forecasting accuracy. 


  • The following hypotheses about the possible causes of prediction errors were initially hypothesized:

  • Incorrect algorithm for calculating the forecast based on expert estimates.

  • Insufficient quality of sales input data.

  • Use of outdated forecasting software.

  • After analyzing the data and the current approach, hypothesis No1 was confirmed – the problem is an incorrect algorithm based on subjective expert opinions.

  • To solve the problem, we proposed to move from expert methods to evidence-based machine learning models. This will significantly reduce the prediction error.


Decision tree

  • Data Analysis: At the beginning of the project, the company’s sales and production data for the past 5 years was collected and analyzed. This included information on cheese types, sales volumes, regional characteristics and the like.

  • Problem identification: Data analysis showed that current forecasts have a high margin of error – 38%. This means that the company is regularly wrong in its demand estimates.

  • Hypothesis Formulation: The hypothesis was formulated as follows: “The current prediction error is due to the outdated expert approach. It is necessary to switch to machine learning models.”

  • Selection of machine learning algorithms: Linear regression, neural network and gradient bousting algorithms were selected to build prediction models. These algorithms were adapted to the features of the client’s data.

  • Integration with ERP system: The developed system was successfully integrated with the company’s existing ERP system. This automated the forecasting process and improved planning.

  • Consideration of current market trends: The developed model was designed to take into account current market trends. This allows the company to quickly adapt to changes in consumer demand.

  • Multidisciplinary approach: A multidisciplinary approach including expert knowledge, statistical analysis and machine learning techniques was applied to develop the prediction system.

  • Testing and implementation: After the system was developed, testing was carried out on test data. After successful completion of testing, the system was implemented in the real business processes of the company. 

Additional technical details of the solution

Calculation of the effect of implementation

Practical application of regression analysis:

  • Linear regression for demand forecasting:

y = 10 + 5x1 + 0.2x2 + 0.5*x3

where x1 – marketing costs (conditionally 100 thousand UAH), x2 – product price (20 UAH), x3 – seasonal index (1 in the summer period).

Consequently, the projected demand in the summer period will be: y = 10 + 5100 + 0.220 + 0.5*1 = 160 thousand units

  • Logistic regression for purchase probability:

p = 1 / (1 + e^-(1 + 0.05x1 – 0.2x2))

where x1 – brand awareness (conditionally 60%), x2 – price (25 UAH).

Then the probability of purchase is equal to: p = 1 / (1 + e^-(1 + 0.0560 – 0.225)) = 0.67

  • Such conditional calculations allow us to better understand the practical application of our application models

  • Assume that the current monthly sales volume is 500 thousand units.

  • The average unit price is 50 UAH.

  • The current demand forecasting error is 30%.

  • The proposed system can reduce the error to 15%.


Current monthly income from sales: 500 000 * 50 UAH = 25 mln UAH

Current monthly losses due to forecast errors: 30% * 25 mln UAH = 7.5 mln UAH

Expected reduction of losses after implementation of the system: 7.5 million UAH * (30% – 15%) = 3.75 million UAH

Consequently, for a year the savings can amount to: 3,75 mln * 12 months = 45 mln UAH

At the cost of the project of 60 million UAH, the payback period will be: 60 million UAH / 45 million UAH/year = 1.3 years

Thus, even approximate calculations on conventional data allow to estimate the potential economic effect and efficiency of the proposed solution.

  • Historical prediction error: 30%

  • Machine learning model error on test data: 15%

  • Reduction in prediction error: 30% – 15% = 15%

  • Savings from reduced prediction error: $200,000 * (30% – 15%) = $75,000 per year

  • One-time implementation costs: $150,000

  • Payback period: $150,000 / $75,000 = 1 year

  • Projected sales conversion growth: 35% + 15% = 50%

  • Projected revenue growth: $200,000 * 50% = $100,000 per year

  • Projected profit growth: $100,000 – $75,000 = $25,000 per year

Top countries for cheese consumption per person in 2024 – 2028+:

  • Italy – 27.4 kilograms

  • France – 26.3 kilograms

  • Greece – 25.5 kilograms

  • Germany – 24.9 kilograms

  • Switzerland – 24.6 kilograms

Global cheese production is expected to continue to grow in the coming years. This growth will be driven by factors such as:

  • Population growth: will increase demand for all types of food, including cheese.

  • Rising incomes: will allow people to spend more on food, including cheese.

  • Dietary changes: People are increasingly incorporating healthy foods, such as cheese, into their diet.

Statistical indicators and calculations


Description of the implementation process:

  • Employees were trained to work in the system prior to implementation

  • The solution was integrated with the company’s accounting systems within 3 weeks

  • System testing and debugging took 2 months


  • Staff resistance to changes in business processes

  • Complex IT infrastructure made integration difficult

  • Coping Measures:

  • Conducted workshops with managers on process changes

  • Developed middleware for integration with the IT landscape

Customer Testimonial:

“Thanks to the new system, we have reduced logistics costs and improved production efficiency”


The application of Marrbery’s forecasting and planning system at MSM was a key step in addressing the shortcomings in the accuracy of forecasting models. The integration of an Artificial Intelligence-based forecasting system proved to be the perfect solution to eliminate product surpluses and shortages that were causing serious financial losses.

After implementation, the system was able to reduce forecasting error by 45%, which significantly improved forecast accuracy and reduced financial risks. In addition, a 20% reduction in the ROI period demonstrates the high efficiency of the implementation. 

A 15% increase in sales conversion and a 30% improvement in data analysis confirm the great potential of this solution for the company’s further development.

All these results demonstrate the importance of utilizing modern technologies such as artificial intelligence in solving complex business problems and achieving success in a competitive marketplace.

                                        appendix 1

Additional findings

In addition to the above, the following conclusions can be drawn:

  • The implementation of the forecasting and planning system has enabled MSM to improve its competitiveness. More accurate forecasts allow the company to make more informed decisions about production, distribution and marketing. This allows the company to optimize its operations and avoid unnecessary costs.

  • The implementation of the forecasting and planning system was an important step in MSM’s development. The system provides the company with new opportunities for growth and development. The company can use the system to expand its business into new markets and develop new products.


  • A 45% improvement in forecasting accuracy (30% reduction in RMSE) demonstrates a huge step forward in forecasting, which will allow the company to operate much more efficiently.

  • A 20% reduction in the ROI period (from 24 months to 18 months) indicates the cost savings and cost recovery of implementing a forecasting and planning system.

  • A 15% increase in sales conversion (improving conversion rate from 35% to 40%) indicates improved sales efficiency and resource utilization.

  • A 30% improvement in data analysis (reduction in model building time from 20 days to 14 days) indicates faster and more efficient data-driven decision making.

Improved forecasting accuracy by 45% (30% reduction in RMSE)
Reduction of investment payback period by 20% (from 24 to 18 months)
Increase in sales conversion by 15% (increase in conversion rate from 35% to 40%)
Improved data analysis by 30% (reducing model building time from 20 days to 14 days)

Estimating the accuracy of the model
Accuracy = (TP + TP) / (TP + TP + TP + FP + FP)
Precision = TP / (TP + FP)
Feedback = TP / (TP + FN)
F1 estimate = 2* (Accuracy * Recall) / (Accuracy + Recall)
Estimation of economic effect:
ΔProfit = (Price – Cost) * ΔVolume of Sales
Δ Sales Volume = Conversion * Traffic * Uplift
Uplift = Conversion gain from new model
Description of neural network architecture:
2 input layers by number of factors
3 hidden fully connected layers of 50 neurons each with ReLU function
Output layer with linear activation for value prediction
Optimization with Adam algorithm

Summary of results

  • Reduced forecast error from 38% to 15% (60% improvement in accuracy)


  • Save $250,000 per year in logistics costs and write-offs


  • Reduced procurement planning time by 20%


  • Increase sales conversion by 15%


  • Accelerate data analysis by 1.5 times


  • Payback period of 1 year


In addition to the above findings, the following recommendations can be made to MSM:

  • Expand the use of the forecasting and planning system to new products and markets. This will allow MSM to gain even more benefits from the system implementation.

  • Automate decision-making processes based on forecasts. This will enable MSM to improve the speed and efficiency of decision making.

  • Develop a forecast-based inventory management system. This will enable MSM to optimize inventory levels and avoid unnecessary costs.

Implementing these recommendations will allow MSM to gain even more benefits from the implementation of the forecasting and planning system.

Forecasting based on 3 scenarios

Scenario 1 (conservative):

  • CAGR of production in the dairy sector by 1-2% (low growth rates due to limited investments)

  • ROA in the industry at 5-7% (low return on assets due to high costs)

  • EBITDA Margin <15% : 10-15% (low profitability before interest, taxes, depreciation, amortization and impairment) Industry Competitiveness – 3 points.

  • Liquidity 1.2-1.3% (satisfactory level for current liabilities)

  • Debt to Equity Ratio: 0.5-0.7 (low level has upside potential)

  • Return on Equity: 3-7% (high return on equity)

Scenario 2 (basic):

  • CAGR of production output in the dairy sector by 3-4% (moderate growth rates)

  • ROA in the industry 7-9% (average profitability level)

  • EBITDA Margin: 15-20% (average profitability before interest, taxes, depreciation, amortization and depreciation)

  • Industry Competitiveness – 4 points.

  • Liquidity 1.3-1.5% (sufficient reserve for current needs)

  • Debt to Equity Ratio: 0.3-0.5 (medium level has potential for growth)

  • Return on Equity: 7-9% (high return on equity)

  • Industry competitiveness – 3 points.

Scenario 3 (optimistic):

  • CAGR of production output in the dairy sector by 5%+ (high growth rate)

  • ROA in the industry 10%+ (high level of profitability)

  • EBITDA Margin: 20%+ (high profitability before interest, taxes, depreciation, amortization and impairment) Industry Competitiveness – 5 points.

  • Liquidity 1.5%+ (high reserve for current needs)

  • Debt to Equity Ratio: <0.3 (high level has potential for growth)

  • Return on Equity: 10%+ (high return on equity)



  • Focus on reducing costs to maintain profitability.

  • Invest in automation and digitalization to increase efficiency.

  • Expand presence in the domestic market.


  • Focus on innovation and diversification to grow and remain competitive.

  • Expand presence in international markets.

  • Invest in marketing and advertising to increase brand awareness.


  • Expand production rapidly to meet growing (increasing) demand.

  • Enter new markets to capitalize on industry growth.

  • Invest in research and development to develop new products and services.


  • Scenario – cost reduction, logistics optimization.

  • Scenario – expansion of assortment, marketing.

  • Scenario – launch of new production facilities, export supplies.

Take a visual journey into the future of your business with an interactive market visualization. Example  Explore now!


Alexey Konovalov is a partner at Marrbery, where Natalia Shevchenko is a consultant; Marina Krivosheya is a senior expert; and Maria Zankovetskaya is a consultant.

The authors would like to thank the following individuals:

Kovalenko Halyna, Snitko Kyrylo, Litvinenko Vasyl, Grigorenko Anna, Melnyk Oleksandr, Savchenko Olesya, Koval Oleh, Gordienko Denis, Timchuk Bohdan, Kravka Lilia and others who helped in the process. 

Moving into the future together!


Keep up to date with our latest statistics

Будьте в курсі наших останніх статистичних даних