Machine learning forecasting: why, what and how

With customer expectations and preferences changing faster than ever, a deep understanding of customer demand is essential to making the right decisions on marketing spend, sourcing, inventory, production, transportation, personnel , etc.

Critical business metrics such as revenue, capital expenditure, risk assessment, profit margins, cash flow, and capacity planning all depend on accurate demand forecasts, which can ultimately help businesses to estimate total sales and revenue for a defined future.

What is demand forecasting?

Supply planning, manufacturing planning and financial planning

Typically, demand forecasting includes activities such as supply planning, product manufacturing planning (eg, sourcing, R&D), and financial planning. The critical aspect of these planning activities is understanding customer product demands and how to meet those demands in the fastest and most efficient manner. By capturing future demand variability through forecasting, companies can more accurately predict customer behaviors and meet their demands with higher confidence and dramatically reduced time from order to delivery.

Supply planning has a direct impact on the bottom line. Therefore, improving planning accuracy can help reduce direct costs such as warehouse storage, shipping and transportation costs, and product disposal. Naturally, forecasting material demand, key to supply planning, becomes a priority. Even smaller and more obvious advances in demand forecasting can show huge improvements in efficiency and cost, enabling efficient use of available resources.

However, the challenge that many companies face is understanding when to apply forecasts and when not to apply them. Given the complexity of the different workflow steps and the dependencies between the large number of relevant parameters, machine learning is an ideal solution for automating demand forecasting and planning.

How can machine learning improve the accuracy of demand forecasts?

Machine learning (ML) in demand forecasting helps avoid traditional challenges associated with planning such as long delivery times, high transportation costs, high levels of inventory and waste, and the taking of incorrect decision due to inaccurate forecasts.

The goal of ML models is not only to increase the accuracy of demand forecasts, but also to free up the time of demand planners. Planners can use their time much more efficiently by focusing on the most important products or gathering more last-minute information to advise forecasts more accurately. Given sufficient volumes of historical forecast data, products and sales can generate more accurate demand forecasts using ML techniques.

Machine Learning Demand Forecasting Example: The Ericsson Model

To find out how ML can improve demand forecasting, we can look at how it has transformed Ericsson’s own approach to balancing supply and demand, including how we plan for materials, products and resources. across our global operations.

As a manufacturer of hardware components, a predictable supply chain plays an important role in our business strategy. However, as you can see in the figure below, our traditional demand forecasting workflow required a lot of resources and manual steps just to make somewhat reliable forecasts. This manual requirement meant that many data sources had to be analyzed while keeping track of master agreements and product lifecycles to make these forecasting decisions.

Figure 1: Ericsson Sales and Operations Planning Process and Machine Learning Forecasting for Decision Support

To overcome this laborious process and build a foundation for ML models, we developed an evaluation strategy to identify tasks/processes that need to be optimized. In the figure below you can see the framework with different forecast selection layers

Framework with different forecast selection layers

Figure 2: Framework with different forecast selection layers

Evaluate past demand forecasting cycles to understand future demand

In Ericsson’s model, demand forecasts are generated monthly for a defined period. Since the exact timing of demand is not a major concern for us, the assessment is made for the next three months using a percentage error for an interpretable measure as well as a bias to indicate an over- forecast and an under-forecast. The valuation horizons can remain the same over the next three to five months, i.e. for the January cycle, the valuation would be done on the months of April, May and June. This assessment is made by joining the actual sales data volumes requested to the supply demand forecast.

ML procedure for evaluating past sales cycles with MAPE

Figure 3: ML procedure for evaluating past sales cycles with MAPE (Mean Absolute Percentage Error)

Generating Demand Forecasts: The Decision Tree Approach

To generate an ML-based forecast that is better than manual alternatives, it is important to collect various sources of data that can impact product demand, such as:

  1. Previous request
  2. Future sales confirmed
  3. Other demand forecasts (at earlier demand stages, historical and current cycle)
  4. Pre-sales data (Closed-won, framework agreements)
  5. Product data (life cycle, substitution plan, product segment)
  6. Customer data (CAPEX)

With continuous demand every month, each product needs a forecast for the defined customer. Our analysis shows that there is a periodic nature of the time series (one of the statistical techniques) for each product.

With the evaluation strategy and the availability of supported data in place, the next step is to find the most suitable model among various prediction models in ML.

At first, time series modeling seemed like an obvious cause for the periodic nature of the supported data. Nevertheless, a decision tree solution has proven best suited to a regression problem such as irregular customer demands for a single product. Several decision tree models, for example Random Forests and XGBoost, provided the best results for different products and customer parameters. Once the ML model strategy is in place, the hyperparameters need to be further tuned.

Multiple decision trees are called "Random Forests" Where "Random Decision Forests"

Figure 4: Multiple decision trees are known as “random forests” or “random decision forests”

The Benefits of ML Demand Forecasting: Halve the Percentage of Errors

By implementing ML-based demand forecasting, companies can significantly reduce forecasting errors compared to manual alternatives. In the nearly 29 months since implementing ML-based demand forecasting, Ericsson’s forecast deviation performance has improved by 40-50% relatively.

Trend of the wMAPE forecast deviation (Weighted Mean Absolute Absolute Percentage Error)

Figure 5: Trend of the wMAPE forecast deviation (Weighted Mean Absolute Percentage Error)

Even though the accuracy of demand forecasting has increased significantly with the ML-based solution, there are still enough edge cases where the relevant data is not available. Therefore, demand planners are required to both validate and further increase forecast accuracy.

An ML forecasting dashboard such as this can encourage informed decisions

Figure 6: An ML forecasting dashboard such as this can encourage informed decisions (note that the data above is not representative)

As a result, ML demand forecasting has resulted in improved production lead times, as well as increased operational efficiency, cost savings, and customer delivery with better quality and satisfaction. Demand forecasting ML models are already proving to be a valuable asset for the entire product portfolio in all markets.

Valuable creation

Figure 7: Activating the value

The Future of Machine Learning Demand Forecasting

ML is quickly becoming a ubiquitous technology, but also an iterative technology that has a . This means that the sooner we deploy a solution, the more beneficial improvements can be made to the business by refining the algorithms. More importantly than the value generated by optimizing processes, reducing costs and improving margins, we have learned that ML can have a significant impact on improving overall customer satisfaction when dealing the right problem with the right approach.

The benchmark for customer satisfaction and business efficiency is rising. Is your business ready?

You want to know more ?

Ericsson’s other ML application areas include predictive network planning, anomaly detection, ticket classification and management, BOM generation, node fault prediction, transport management, forecasting freight, inventory optimization, supply planning, etc. Learn more about these areas and the other opportunities that await you on Ericsson’s Artificial Intelligence page.

Check out our other blog on applied AI for environmental sustainability.

Sherry J. Basler