While predicting the future might seem like an impossible feat, we’ve gotten pretty close (when it comes to forecasting demand, at least).
Since the ability to predict future demand has become crucial for businesses across industries, plenty of different methods of demand forecasting have been tried and tested over the years. From traditional statistical methods to those driven by AI (or more specifically ML), it’s safe to say that you have plenty of choice. Want to know which is right for you? Read on as we look at the benefits, limits, and applications of these methods.
Before we get into the nitty-gritty, let’s break down the basics of demand forecasting (for those not in the know). Demand forecasting is a type of data analysis that is used to estimate the demand for particular products or services in the future. Originally used for revenue and short-term supply chain optimisation, demand forecasting can now be used to predict:
Broadly, these methods can also be categorised by the type of data that they use. Quantitative forecasting methods - think statistical, causal, and machine learning models - rely on hard data, whereas qualitative forecasting methods use expert opinions and market research. And, different types of demand forecasting can be applied depending on:
Different stages of a product’s life cycle
The desired length of forecasting - i.e. how far into the future that you want to predict demand for - be it short term (0-3 months), medium term (3-6 months), or long term (6+ months)
The needs and capabilities of your business - whether you’re looking for something quick and cheap or are willing to foot the cost for greater accuracy, there’s a method out there for you
These methods rely on the insight and experience of experts, stakeholders, and the general public, which is turned into quantitative estimates so that a forecast can be produced. This qualitative analysis can take the form of market research, comparative analyses, or consulting with experts. Unlike other methods, qualitative analysis does not always take historical data into consideration when generating forecasts, which makes it a useful technique for:
New products or products with scarce historical data
Where R&D demands are hard to estimate
As you might have already noticed, the majority of demand forecasting methods utilise quantitative data. One of the issues with qualitative analysis is that the cost of implementing these techniques is quite high in comparison to quantitative methods. In addition to this, methods that rely on expert opinions can be affected by bias, and some of the more common methods - such as the Delphi Method and Market Research - can take several months to properly implement where quantitative methods can be implemented in as little as a day.
Statistical Time Series
Classical statistical time series methods rely primarily on historical data (ideally at least several years’ worth) in order to predict future demand. Here, a time series is considered to be a set of chronologically-ordered data points, and from this you can identify:
Seasonality - i.e. a regular variation in the data
Trends and growth rates
Often, statistical time series methods use historical data to create a ‘rolling average’ for future demand. More sophisticated methods using weighted data so that more recent data has more of an effect than older data. These methods are often cheaper and quicker to implement than other forecasting methods, but this does come with a lower accuracy than more costly and time-consuming alternatives. As a result, statistical time series methods are best suited for:
Mid to long-term forecasting
Products that are well-established and have stable demand
Predicting total demand rather than for individual products
Since time series methods require data that is stable - meaning that it follows predictable trends - these methods can fall short in predicting market saturation of a product, as well as random spikes in demand, and seemingly illogical changes in customer preferences.
Time series methods are popular because, at a glance, they make a lot of sense. What it lacks, however, is the nuance found in some of the newer forecasting methods.
Causal forecasting, on the other hand, considers the relationships between demand (or any other metric you want to forecast) and other variables. Unlike traditional statistical methods, causal models are multivariate (they can ingest a variety of data sources), with potential data sources such as internal sales data, social media activity, surveys, product features, weather, competitor pricing, etc etc.
As a result, causal models are best-suited for creating medium to long-term forecasts where demand is influenced by multiple factors and for individual products, product categories, and subclasses.The ability to ingest multivariate data also means that causal models are useful for identifying changepoints. This means that the models can identify points where an event occurs that changes the trend, and could be in response to anything from price changes or sales to product replacements and even major crises such as COVID-19. Other types of Demand Forecasting that are univariate, such as statistical time series methods, assume that these change points exist but identify where they happen from inferred trends.
Two of the more common causal models are the regression model - which defines the interaction between two variables using the least square method - and the econometric model - which considers the relationship between external economic variables and internal sales data. While these methods generally need at least two years of data to be successfully utilised, they are often more accurate than traditional statistical and qualitative methods, are relatively inexpensive and quick to implement - think anywhere from a week to a month.