There's been a lot of hype about using machine learning for forecasting. And rightfully so, given the advancements in data collection, storage, and processing along with technology improvements, such as super computers and more powerful software. There's no reason why machine learning can't be used as another forecasting method among the array of existing forecasting methods.
At SAS we're fortunate to have some of the top domain experts in forecasting, predictive analytics and machine learning. As the leaders in predictive analytics and machine learning, according to the recent The Forrester Wave™: Predictive Analytics and Machine Learning Solutions, Q1 2017, SAS domain experts were working with multilayer neural networks 30 years ago.
Guess what? Multilayer neural networks are now called deep learning. It's new. It's hot! Actually, it's new spin on an old language.
To be fair, I found myself taken in too. So much so that I wrote a blog post summarizing the current thinking behind machine learning-based forecasting, and how it could be potentially applied to demand forecasting.
The question is: Are there any enterprise demand management solutions that use machine learning-based forecasting, today? As far as I know, the answer is no. That said, several software providers include machine learning as an addition to their capabilities.
To be honest, there's no single tool — mathematical equation or algorithm — designated just for forecasting. All ERP demand management solutions use “best fit” selection based on how well a model fits to the historical demand using basic time series methods (e.g., moving averages and exponential smoothing—non-seasonal and seasonal).
Some standalone forecasting software packages include ARIMA models and multiple linear regression. SAS® Forecast Server and SAS® Demand-Driven Planning and Optimization: Forecast Analyst Workbench use the same expert system that's built on a multiple patented automatic large scale hierarchical forecasting platform. Used at all levels of the business, these packages contain all categories of models, including moving averages, exponential smoothing, dynamic regression, ARIMA(X), unobserved component models, combined models, and others.
This type of forecasting system requires minimal human intervention as the system does all the heavy lifting from a modeling standpoint, and uses not only model fit, but can also incorporate in-sample/out-of-sample analysis to determine the appropriate model automatically. Given its capabilities, it could be mistaken as being a machine learning-based forecasting solution. It automatically builds the models based on patterns in the data, and it requires minimal human intervention.
Where does artificial intelligence fit in?
Expert systems are an integral part of artificial intelligence, also known as symbolic artificial intelligence. This type of AI, that's designed to think like a human, is differentiated from sub-symbolic artificial intelligence, which includes the newly revived research area of neural networks and deep learning.
So, it would be absolutely fine to say that automatic large scale hierarchical forecasting systems like SAS Forecast Server and SAS Demand-Driven Planning and Optimization: Forecast Analyst Workbench use AI. However, expert systems aren't really a part of machine learning, unless you have a decision tree learn the rules of the expert system from the data. Then, it would indeed be a so-called "meta learning" approach. The most successful form of symbolic AI are expert systems, which use a network of production rules, like SAS Forecast Server or the Forecast Analyst Workbench.
Production rules connect symbols in a relationship similar to an if-then statement. The expert system processes the rules to make deductions and to determine what additional information it needs, (i.e. what questions to ask, using human-readable symbols). In other words, they run the same way every time. They are completely deterministic. They are mathematical models in which outcomes are precisely determined through known relationships among conditions and events, without any room for random variation. In such models, a given input will always produce the same output, such as in a known chemical reaction. Given the same data, they give the same answer every time.
SAS has been offering machine learning algorithms for the past 40 years. While we haven’t integrated these algorithms into our forecasting and demand planning platforms, we have applied machine learning for other relevant customer needs over the years.
Now, the recently launched SAS® Visual Forecasting, which runs on SAS® Viya, allows customers to add their own code — any code that can be generated with machine learning algorithms or any other algorithms. Those same machine learning algorithms can also be added to the open model repositories in SAS Forecast Server or the Forecast Analyst Workbench to complement existing forecasting algorithms and to enhance forecast accuracy across your entire product portfolio.
My advice: before you embark on more advanced modeling algorithms like machine learning, make sure you have exhausted all the traditional time series methods including those that can incorporate causal factors.
To quote one of my mentors, Dr. Oral Capps from Texas A&M University:
An ordinary least squares regression algorithm may be a linear modeling approach, but many times it works in situations that you would think it normally would not.
Learn more about the opportunities for machine learning in the whitepaper, The Evolution of Analytics.