Misconceptions about machine learning demand forecasting

Machine learning demand forecasting has the potential to transform the way metals companies operate, but leaders might be hesitant to use the technology based on uncertainty about the return on investment and outdated information on the topic.

Following are three misconceptions metals leaders might have about machine learning demand forecasting along with what it can actually offer.

Misconception 1: Using a spreadsheet for demand forecasting works just as well as ML models

Demand forecasting is a complex process that must take many variables into account. After years of improving spreadsheet-based methods and manual internal processes to achieve greater accuracy, metals leaders are understandably hesitant to augment their current processes with machine learning demand forecasting without better understanding the value it offers.

Reality: Metals leaders should consider the implications of being slightly more accurate over time rather than waiting for large accuracy gains to make a change. While machine learning demand forecasting models might not make perfectly accurate predictions every time, they can beat the results of traditional forecasting methods in small but provable ways. Predicting the future will never be a precise science, but metals leaders shouldn’t underestimate the value of being even a little bit more accurate than their competition.

In addition, any forecasting method will be more accurate closer to the prediction date – a reality that equally affects the weather forecasts we use every day and demand forecasts for metals products. Automation and integration that are possible with machine learning demand forecasting tools allow metals leaders to request updated predictions more often, improving the quality of data used in sales and operations planning processes.

Misconception 2: Machine learning demand forecasting requires a full-time data scientist

Metals leaders might believe they need to have a focused team of data scientists to use an ML model for demand forecasting. Historically, companies have used data scientists to create custom models that took months to develop and resulted in software that no one else on the team understood or knew how to use. This limitation barely differentiated machine learning demand forecasting from other manual methods that also relied on deeply specialized resources.

Reality: Today, companies like Microsoft have developed self-training ML tools, such as Microsoft Azure™ AutoML, that can take a history of data, automatically apply multiple models to make predictions, and identify the best ML model for the data before returning a forecast with an indication of confidence level. Tools like AutoML allow metals leaders to quickly evaluate and receive the benefits of machine learning demand forecasting without committing to the expense of a full-time data scientist.

However, when metals companies choose to work with a team of specialists to help guide them in their digital transformation journeys, they can enjoy an added benefit of initially consulting with a data scientist who can evaluate business data and make recommendations. Tools like AutoML are effective for speeding up implementations and enabling business users once a tool is in place, but initial guidance from a specialist is indispensable if companies don’t want to leave potential accuracy gains on the table.

Misconception 3: Machine learning demand forecasting can only be used to predict upcoming purchases or sales orders

Some metals leaders forecast for raw material needs – or upcoming purchases – and use this information to plan operations. Others forecast for finished good needs – or upcoming sales – and then work backward to determine material and resourcing implications. Some companies might create both types of forecasts, particularly if they manage a mixed manufacturing environment. The mindset that forecasting can only be done for the starting or finishing point has historically stemmed from the inflexibility, politics, and long lead times associated with traditional forecasting approaches.

Reality: With machine learning demand forecasting, metals leaders can forecast with the frequency or granularity they prefer for any item number with a relevant history in their enterprise resource planning (ERP) system. Planners can then use these predictions for demand midstream in their fulfillment processes to provide a robust dataset to their master scheduling solutions, which helps improve the value of their solution’s scheduling recommendations. Workers also benefit from these predictions by receiving improved daily recommendations about decisions such as which truck they should send from one site to another or which production order they should prioritize.

Even if it’s only as accurate as manual approaches that have been previously used, machine learning demand forecasting frees up time for metals leaders to manage the configurations of their ERP systems and master scheduling solutions in line with business strategy and market realities. The use of ML for forecasting vastly expands the possible scope and frequency of forecasting, improving the quality of data fed to other technology investments.