The potentials of AI are endless for demand forecasting

AI has a bit of an image problem. From gender bias to Brexit, it feels like there’s little AI hasn’t been accused of. So, when I read this article on Essential Retail, I was pleasantly surprised to see it getting some of the positive press it deserves. Interesting analogies aside – I’m not sure the dinosaurs would have evolved quicker if given due warning, though I must admit to never having read the Origin of the Species - I completely agree with the author’s sentiment; implemented effectively, AI has the potential to completely transform the retail industry. Despite this, there are a few points on which we disagree, particularly when it comes to demand forecasting.

The author correctly points out that demand forecasting is one of the areas already benefiting most from advances in machine learning. However, they then go on to cite better customer experience through improved availability as the primary benefit. At Atheon, we’ve been fortunate enough to work with a number of clients to improve their forecasting processes, and have seen first hand the numerous benefits an AI-driven approach can deliver. While improved availability is undoubtedly a big plus, it really is just the tip of the iceberg.

Availability is rightly a big priority for retailers, but the heavy focus placed upon it often causes inefficiencies elsewhere as uncertainty over upcoming demand leads to increased stockholding to protect against unforeseen spikes. Owing to this, the biggest wins we’ve seen have actually come further up the chain in the form of lower stock on hand and reduced logistics costs. When there is a high degree of confidence in the forecast, buffer stock levels can be lowered, reducing working capital and freeing up valuable space. Beyond stockholding, better forecasts can deliver a wealth of savings across the value chain; improved transport planning, optimised depot labour schedules and even better deals with suppliers who no longer have to react to unexpected last-minute orders. All of this, of course, is based on the assumption that AI-based forecasts are more accurate than their traditional counterparts, which brings me to my second point.

I don’t agree with the assertion that the main advantage of AI-powered forecasting is the speed at which the forecasts are produced; in my experience, the real benefit is the increased accuracy of the forecasts themselves. Traditional methods generally begin with a set of predefined rules which, when applied to data, output a forecast. Unfortunately, as humans, we don’t have the cognitive capabilities required to process all of the complexity present in the world around us, and so by necessity these rule sets are grossly simplified. Deep learning methods turn this process around, learning the rules from huge volumes of historic data, and then using these to make predictions about the present. This reversal, alongside the processing power that modern computers provide, means a huge number of influencing factors and their various interdependencies can be considered and learned. Thanks to this, the resulting forecasts account for much more of the inherent complexity, better represent the real world, and are often significantly more accurate.

The advantages don’t end there; for example, how demand for a product responds to different factors will change over its lifecycle. When forecasting with traditional methods, these changes often have to be identified manually before a new, different rule set is applied. Learning from the data, AI-driven forecasting methods are better placed to pick up on these subtle changes in behaviour, and the forecasts generated will respond accordingly. On top of this, as mentioned by the original author, the increased automation present in the process frees up man-hours, allowing those who would previously have been tasked with creating the forecasts to focus instead on the few exceptional instances which require human attention.

In conclusion, AI presents a far broader opportunity for demand forecasting than increasing the speed of forecast production. Deep learning algorithms will consider many more variables, work in much more detail and learn from past outliers in a way which traditional methods simply can’t address. The results will be greater accuracy, faster execution (yes) leading to near real-time re-forecasting and the inclusion of AI-forecasting as a primary driver of stock ordering, inventory management and supply chain execution.