There's been a lot of hype regarding using machine learning (ML) for demand forecasting, and rightfully so, given the advancements in data collection, storage, and processing along with improvements in technology. There's no reason why machine learning can't be utilized as another forecasting method among the collection of forecasting methods already being utilized — or maybe not.
Are expert systems using machine learning?
To be honest there is no single tool — mathematical equation or algorithm — designated just for demand forecasting. Given the capabilities of some of the more advanced software solutions, they could be mistaken as using machine learning to automatically build the models, when in fact they're using business rules.
Expert systems are an integral part of artificial intelligence” (AI), also known as “symbolic artificial intelligence” in the conventional sense. Expert systems are a set of production rules that connect symbols in a relationship, for example, an if-then statement, and they're the most successful form of symbolic AI.
The expert system procedures use business rules to make deductions and to determine what additional information is needed, (i.e. what questions to ask, using human-readable symbols). This is differentiated from “ sub-symbolic artificial Intelligence,” which includes the newly revived research area of deep learning neural networks.
Are expert systems really learning?
Technically, these early symbolic artificial intelligence systems, also known as expert systems, are not actually learning. Expert systems are not a part of ML, unless you have a decision tree, like random forest or gradient boosting, to learn the rules of the expert system from the data.
ML not only creates but changes the rules as it learns versus always using the same set of production rules. If systems do include a learning algorithm, then they would indeed be a so-called "sub-symbolic” approach. When it comes to demand forecasting and planning, the most successful form of symbolic AI are expert systems, which use a set of production rules, not real machine learning.
So, beware of software vendors who claim to use machine learning to automatically forecast your products. If they don’t have a learning algorithm like neural networks, gradient boosting, and/or multi-stage models (a combination of neural networks and times series models), then they're not truly using ML.
Can we use ML algorithms to forecast time series data?
Over 30 years ago, Spyros Makridakis started a series of forecasting competitions known as the M-series. The purpose was to learn how to improve forecasting accuracy and how such learning can be applied to advance the theory and practice of forecasting. A series of M1, 2 and 3 competitions were conducted in the 1980’s and 1990’s.
In 2018, Spyros Makridakis and his team conducted the M4 competition. The purpose of M4 was to replicate the results of the previous competitions and extend them to include an increased number of series and ML forecasting methods to evaluate both point forecasts and prediction intervals. The five major findings of M4 were:
- Out of the 17 most accurate methods, 12 were ‘‘combinations’’ of traditional statistical approaches.
- The most significant finding was a hybrid approach that utilized both statistical and ML features which had an average MAPE 10 percent higher than the combination benchmarked methods.
- The second most accurate method was a combination of seven statistical methods and one ML algorithm with the weights for the averaging being calculated by a ML algorithm that was trained to minimize the forecast error.
- The two most accurate methods achieved an amazing success in specifying the 95 percent prediction intervals correctly.
- The six pure ML methods performed poorly. None of them were more accurate than the combination benchmark, and only one was more accurate than a Naïve model.
For more information regarding the 2018 M4 Competition, read the International Journal of Forecasting article, The M4 Competition: Results, findings, conclusion and way forward.
What does all this mean?
The bottom line is: Beware of the hype around using machine learning to automatically build statistical forecasts. Most traditional forecasting technology can only automatically build models up/down a business hierarchy using basic time series methods (moving averages, exponential smoothing) let alone use ML. Al uses business rules, not ML to update and adjust models. Most use best fit to history to determine the best model, rather than in-sample/out-of-sample testing, which has been the proven method to determine how well a model will forecast future demand for a product.
According to the recent results of the M4 competition, we must accept that all forecasting approaches and individual methods have advantages and disadvantages. Therefore, you must consider exploiting such advantages while avoiding or minimizing the shortcomings of each method. Moving forward, the best result is the utilization of a stacked model approach that combines traditional statistical and ML models, as well as a weighted combined statistical model approach (Spyros Makridakis, 2018).