Article

Leverage machine learning to assess the impact of promotions on demand

By Scott Fenwick,
Leverage machine learning to assess the impact of promotions on demand

Before the modern advancements in math, data science and computing technology, forecasting methods was a time-consuming process that was often unreliable. And, for the modern demand planner and marketer, not knowing which promotional campaign contributed to sales was a problem. As a result, we resorted to our intuition and judgment, aka “in my experience…” and that opened the door to unconscious biases.

Over time, the problem has only gotten worse.  To begin with, it takes an already tricky problem and makes it exponentially more complicated, with the proliferation of promotional data. When hyper-promotional activities of campaigns are added, resulting in a heavily overlapping sea of promotions in dozens of channels, spanning a mixture of media types, durations and incentives, the future appears as clear as mud.

This can be an insurmountable problem to solve, but it is solvable. Untangling and decoupling promotional demand from baseline demand patterns, all while there are potentially overlapped promotional events is the first step. By separating promotional demand from baseline demand patterns, the predictive model becomes more effective and more accurate.  We can then apply unsupervised machine learning across this enhanced, massive data set to gain new insights. So, we can confidently learn – with probabilistic [RN1] guarantees – the possible association and impact that each individual promotional event has on demand of a particular SKU.

In other words, Manhattan Associates has built a solution that has the exact capability that demand planners and marketers need. Our software provides clients with the ability to optimally learn deep insights and predictive capabilities from extensive, very complex datasets without the client needing to program it to do so.  Our solution utilizes advanced data science and machine learning algorithms to analyze and explicitly learn the relationships and associations that each individual event has with each other, and with observed demand responses that have been spread across this interconnected graph of billions of demand/promotional-event combinations.

With these methods, we can now reliably decouple and accurately assess the impact that each individual promotional event had on demand, and therefore build better demand models, which in turn will yield better more accurate predictive models for both promotional forecasting and baseline forecasting purposes.

When we avoid our instinct to oversimplify the solution and not insert human subjectivity, and actually embrace the massive complexity and volumes of data by leveraging advances in computing, data science, and machine learning, we become scientific predictors of the future.  We begin to have clarity of what’s to come with a higher degree of confidence in knowing that predicting the future isn’t an exact science, but a scientific approach that models out the variations of countless scenarios, allowing us to apply the most useful one.


[RN1]“Probabilistic guarantee” is the language the data scientist uses, it’s specific to a graphical model to express dependence structure. I think changing it to problematic would change the message.

Deliver On Your Promise to Customers