Demand Estimation Methods: From Traditional Statistics to Predictive Systems

The primary unknown for any company is how much it will sell and what quantity of product the market will request. The reason is clear: demand dictates the needs for production, storage, distribution, purchasing, maintenance, human resources, and a long list of other factors that affect the entire organization, as the departments influenced by sales extend far beyond the supply chain itself.

In this sense, there are stable products with low-fluctuation demand that have traditionally allowed for accurate forecasts based on historical activity. However, day by day, demand is becoming more volatile due to fast-consumption trends and fashions on one hand, and increasingly frequent weather fluctuations on the other. These can result in heaters not being sold during a winter campaign due to unusually warm temperatures, or demand skyrocketing during a cold snap occurring outside its typical months.

We have a very recent example of unexpected and massive demand. The need to frequent open spaces, avoid social contact, and spend more time at home caused by the COVID-19 crisis triggered requirements for pools and bicycles, which remained sold out in both physical and online sales channels and accumulated significant waiting lists.

Events such as the pandemic we are experiencing are impossible to foresee, but it is possible to estimate the population’s needs in the short and long term, using major unexpected events like this as a reference. The lockdown in China, followed by the one decreed in Italy, gave the world many clues that this would likely be the path followed by other states, which was inexorably going to affect consumption trends.

With demand forecasting or prediction systems, 100% accuracy in estimates would not have been achieved, but there would have been enough effectiveness to largely adapt industrial supply to consumer needs. As the name suggests, these tools perform future forecasts or predictions of sales behavior based on objective data of different types depending on the chosen estimation system. Let’s look at the options available and their optimal application environments.

Traditional Statistical Forecasting

Statistical forecasting has been used for decades to perform sales estimates based on historical activity. These systems start from the premise that what happened in previous campaigns will, with high probability, be what happens in future ones. It is assumed that campaigns are stable, leaving unexpected events or situations exposed, which can cause sales to plummet or demand to experience significant increases that cannot be addressed.

Nevertheless, traditional forecasts, with the automation experienced in recent years, continue to be the most widely used due to their low cost and because they do not require knowledge beyond the use of Excel or the Enterprise Resource Planning (ERP) systems already implemented in organizations. They are effective in very stable markets with low SKU turnover, such as plumbing spare parts, and offer good results for obtaining medium- or long-term forecasts.

Demand Planning with Machine Learning

The use of machine learning for demand planning is based on data exploitation at the highest level; that is, it requires both internal and external data sources to make accurate forecasts. Unlike traditional statistical analytics, the use of artificial intelligence allows for the extraction of information from proprietary and contextual sources, working with both structured data (sales history, purchases, financial data…) and unstructured data (weather forecasts, trend analysis, marketing actions…).

Machine learning systems compile all available data—and we are talking about thousands of data points (big data)—process them with mathematical algorithms, and determine patterns from which they transform recorded items into information. In each of these actions, machine learning models learn behaviors and adapt their forecasts, offering greater efficiency. However, this leap in automation and precision requires more technical resources and implementation by technical data science teams.

Each system is suitable for specific business and market conditions, with machine learning showing more effectiveness in situations where working with historical activity is not possible, such as:

• planning for unstable and changing demand,
• launching new products,
• and short- and medium-term forecasting.

Demand Prediction

Demand prediction represents a significant leap from models that forecast sales to those that predict them. The nuance may seem small, but it is very important. Models based on demand forecasting estimate the demand that might occur, which does not imply total certainty, while new models predict what will happen—what the behavior of buyers will be based on market circumstances, the economy, society, etc.—providing the capacity to react to sudden events that change market conditions. Prediction makes it possible to gain the certainty that forecasting does not offer, as it does not just estimate; it shows future demand with high levels of efficiency.

To perform predictions, we again need data; the more that is obtained and the higher its quality, the more accurate the information teams will have to guide business processes. Work is therefore carried out with internal and external sources: ERP, CRM, sensors, work calendars, economic indicators, weather, historical activity, market trends, promotional campaigns… which will be processed with predictive algorithms to transform them into quality information about future market demand.

Demand prediction systems show their maximum efficiency and utility in medium- and long-term strategies. They also demonstrate great results for making predictions in volatile and changing environments and in the launch of new products for which previous information is lacking, as they take into account a multitude of endogenous and exogenous variables to predict what will happen, with high levels of correlation with future reality.

Conclusion

There is no doubt that data exploitation is indispensable in business management. Today, in the era of big data and artificial intelligence, organizations that do not join this revolution will cease to be competitive. However, each tool is indicated for a specific type of product and activity, so before deciding on forecasting, estimation, or prediction systems, it is necessary to study the intrinsic conditions of the company itself, its products, and its markets to choose the most appropriate and effective option for each production environment.