Can Machine Learning improve Demand Planning?

It’s an exciting time to be working in the space of Integrated Business Planning.  Emerging technologies such as Big Data and Artificial Intelligence (AI) present amazing possibilities to transform the practice of operational planning for businesses.  McKinsey forecasts "AI has the potential to deliver additional global economic activity of around $13 trillion by 2030, or about 16 percent higher cumulative GDP compared with today.”  Yet, even McKinsey is right to point out that only with careful and appropriate application of AI can this growth be realized.  

So the opportunity is there, but how to take advantage of it?  Well, there is no one-size-fits-all answer to this question.  Each business is unique in how it will be able to best leverage it.  I would like to illustrate an example of how I believe most businesses can take advantage and begin the journey into AI.  We’re going to focus on the potential use of Machine Learning (ML) to enhance a Demand Planning system.

First, let’s review a couple of definitions, as it can get quite confusing.  Artificial Intelligence is a broad term that covers the development of machines to perform human tasks, tasks that require “intelligence;” not simple 1+1=2, but complex tasks that require perception, nuance, and decision-making.  There are several categories within AI, and lots of potential here.  Machine Learning is a subset of AI that focuses on enabling a machine to adapt and “learn” based on different inputs.   Essentially, feeding the machine with vast amounts of data and allowing it to develop the best algorithm.  It is through Machine Learning that Demand Planning can be greatly improved.

Now, let’s talk about Demand Planning.  We break Demand Planning down into two categories:  Quantitative and Qualitative.  

Quantitative is forecast driven by data only, normally historical sales data.  If your organization uses Statistical Forecasting, then you’re using Quantitative forecasting methods.  Statistical Forecasting uses historical sales or shipment data to project what future demands may be.  The advantages of quantitative forecasting are that vast amounts of data can be churned to generate demand at a very detailed level, usually SKU and/or customer level, with minimal organizational effort.  The disadvantages of this forecasting method are accuracy, data cleansing, and quite frankly, a dumb forecast since it’s only a reflection of what was done in the past.  New algorithms, such as Demand Sensing, are enabling organizations to incorporate other inputs—not just sales history—and significantly improving the short-term forecast accuracy.  Yet, it is still limited to sales orders or other downstream Point of Sales data. 

Qualitative demand planning is forecast-driven by human input, using either intuition or other market inputs to formulate a sales forecast.  The advantage of qualitative forecasting is that inputs not directly associated with sales can be incorporated into the forecast.  For example, if your business trends with U.S. GDP, forecast changes in GDP can easily be incorporated into a qualitative forecast.  The disadvantages of this forecasting method are the organization effort required to maintain it and inherent bias in the forecast.  There is a considerable amount of people’s time put into generating a qualitative forecast.

The best practice in the field of demand planning is to combine both methods.  Use Quantitative Forecasting to generate the details: SKU, Customer, Week; then use Qualitative Forecasting to adjust the aggregate: Product Line, Market, Year.  I like to use the ocean as an analogy.  The level of the water changes on a second-to-second basis.  Mostly due to waves, but the tide also plays a part.  The mechanisms that drive waves (the wind) and tide (the moon) are different, and they both play a role.  So if we wanted to predict the water level, we would want to use a combination of methods; one that used the wind, and another that used the moon.  Either one by themselves can work, but predictors get better when both are used.

OK, so now that we’ve explained a bit of Demand Planning 101.  What does this have to do with Machine Learning?

Well, the opportunity that Machine Learning presents is the ability to replace a Qualitative Forecasting method with algorithms that use other, more intuitive inputs.  These are the same types of inputs that would be used in a Marketing Forecasting, but without the disadvantages of organization resource loads and bias.  The waves can continue to be forecasted using sales data, and the tide can now be forecasted algorithmically using external market information.  Organization effort can be committed to discussing options and action rather than used to generate forecast data.

How do you do it?  Well, there’s no shortcut.  It’s a journey.

Build a roadmap. Start where you are today and build a roadmap.  If you are purely qualitative, then start with building a statistical forecasting base that you can adjust qualitatively.  If you are purely quantitative, then engage with your commercial organization on a monthly basis to get some intuition into the forecast.  They key is to combine the two and understand how much each one factors into demand.

Deploy Machine Learning. Once you’ve established the use of both, then begin with using Machine Learning outputs to vet the Qualitative outputs.  Use at least two different versions of the Qualitative demand output (Marketing & ML). If you have the possibility of creating additional forecast, such as Sales or Customer Forecast, then do it.  Just like any consensus planning process, discuss the differences and propose a consensus to advance for S&OP/IBP.

Then it’s up to your organization whether to continue with or discontinue manual Qualitative Forecasting methods.  Remember, the goal is a better forecast and better business plans.  

The ultimate tease:  When forecast is generated algorithmically, then it can be updated perpetually and adjusted ad-hoc.  When combined with automated Supply Planning, scenarios can be generated in mass. Whenever market conditions change or different conditions want to be modeled, Integrated Business Planning can be immediately executed. Ultimately, decisions can be taken based on all of the possible scenarios rather than a single plan.