From Supply Chains to Demand Chains: A Machine Learning Approach to Managing Complex Supply Chains
In the first of a series of online research seminars, Sasin’s Distinguished Professor of Digital Business Strategy & Machine Learning, Ravi Aron, recently offered us a teasing, whirlwind glimpse into his world. In a 90-minute talk, From Supply Chains to Demand Chains: A Machine Learning Approach to Managing Complex Supply Chains, he introduced an illustrative case which he presented with beguiling simplicity. By the end of his talk, he had led everyone through a steep learning curve culminating in a profound theoretical construct.
The model around which he chose to frame his presentation seemed straightforward enough: how to improve the long term viability of the health sector supply chain in Bangkok. Professor Ravi identified this challenge as an example of a wicked [as in ’resisting resolution,’ not ‘evil’] problem. The difficult and contradictory requirements posed by the numbers of variables involved, and their timing, seemed pre-destined to point us towards an inevitable conundrum.
Nonetheless, Professor Ravi felt confident enough to test the hypothesis that machine learning techniques, appropriately supplemented by human and IoT inputs, can assist in the rationalized and efficient systemwide ordering and stocking of medical supplies.
All supply chains are sensitive to stockouts and excessive inventories, which can be disruptive in the one case, and wasteful in the other. Overstocking can encourage pilfering while shortages can delay performance. This is the case regardless of the industry. The stakes are highest in the medical community where errors can cost not just physical loss, but loss of life.
Drilling down, Professor Ravi was able to identify a high occurrence rate of inventory shortages and overages among all hospitals, large and small. Moreover, shortages at the hospital level correlate with shortages and oversupplies at the dealer levels, as well. (Between the dealers and manufacturers are the Aggregators, who add another level of complexity.)
Decentralized storage and consumption characterizes the medical supply chain. Many products are subject to price controls which militates against pricing as a demand-dampening option. Moreover, hospitals do not cross-communicate well, and the wide range of procedures performed add still further to the predictive miasma. Human sampling-based needs forecasting is therefore hit or miss.
Come to stock control, the accuracy of predictions can be improved by machine learning techniques. They can smooth over gaps in data and generate forecasts from complex, noisy inputs. Machine learning techniques are unphased by large volumes of rapidly churning data regardless of hierarchy or heterogeneity. On the other hand, while machine learning techniques can generate good predictors, they cannot resolve associations between activities and outcomes in real time. Rather, they yield snapshot predictions based on data available at any given moment. Professor Ravi calls this the black box limitation and points to structural modeling and econometric attenuation to help create practical management-quality guidance.
Our challenge, he concluded, is to find a way to perfect the admixture of econometrics, machine learning, and IoT tracking to predict ongoing consumption of medical supplies, and other ‘straightforward’ concerns as well, no doubt. He used a two-stage modeling technique where in the first stage he generated machine learning predictions and in the second stage he created variables out of these predictions and combined them with an econometric model to explain the factors that would lead to the optimal functioning of complex supply chains.