Supply Chain Management: Machine Learning and Demand Forecasting

Tuesday, January 28, 2014

Machine Learning and Demand Forecasting

Demand forecasting is an essential part of inventory management. Many organizations use methods based on past experience, either of the organization or the manager responsible for that prediction. Often, even basics like time-series models are used inappropriately.[1] This all amounts to a lot of money spent on infrastructure, technical systems and human capital to keep what can be a relatively inefficient supply chain running.

However, in the world of ‘Big Data’ and advanced analytics, many wonder whether larger forecasting models couldn’t be made using machine learning techniques. The hypothesis is that particularly for industries with high variability and volatility of demand, advanced machine learning and large data sets can do what current forecasters cannot. (Or, at least what current forecasting methods cannot do very well.)

The chart below describes 4 different machine learning techniques and in which situation they are most applicable[2]:

Overall, many of the uses above correspond to the fashion industry. As outlined by Nenni, Giustiniano and Pirolo,[3] the fashion industry is marked by:

· High impulse shopping
· High volatility
· Low predictability
· Large product variety
· Large variance in demand

This suggests supply chains of fashion companies could be improved by machine learning techniques, as opposed to just modifying the supply chain to be more responsive and agile. However, as far as I can tell very few companies are significantly using these techniques.

Certainly, one reason for this is lack of adequate human capital. Machine learning is a specialized skill and employing a team of engineers for the task is expensive. Additionally, computation time of many of the aforementioned methods is high- at least now, perhaps too high to consider operationally practical, especially in a highly volatile industry like fashion. Finally, translating a model’s data into an actionable forecast can prove difficult.

Additionally, I would bring up a more philosophical point: while machine learning is certainly cool- and has some really interesting real-world applications- many people and many organizations are not quite ready to turn everything over to complicated black boxes and algorithms. Despite statistical viability, humans often believe their own experiences to be more accurate, and handing multi-million dollar decisions to a statistical method is easily dismissed.

What do you think: Is it only a matter of time before advanced machine learning is integrated into supply chains within high-volatility industries? Or is the computational time and uncertainty too much to be feasible in the near future when simple methods (like a highly responsive supply chain) seem to work?