Understand your decision rather than your model prescription: Towards explainable deep learning approaches for commodity procurement

Abstract

Hedging against price increases is particularly important in times of significant market uncertainty and price volatility. For commodity procuring firms, futures contracts are a widespread means of financially hedging price risks. Recently, digital data-driven decision-support approaches have been developed, with deep learning-based methods achieving outstanding results in capturing non-linear relationships between external features and price trends. Digital procurement systems leverage targeted purchasing decisions of these optimization models, yet the decision-mechanisms are opaque. We employ a prescriptive deep-learning approach modeling hedging decisions as a multi-label time series classification problem. We backtest the performance in two evaluation periods, i. e., 2018–2020 and 2021–2023, for natural gas, crude oil, nickel, and copper. The approach performs well in the first evaluation period of the testing framework yet fails to capture market disruptions (pandemic, geopolitical developments) in the second evaluation period, yielding significant hedging losses or degenerating into a simple hand-to-mouth procurement policy. We employ explainable artificial intelligence to analyze the performance for both periods. The results show that the models cannot handle market regime switches and disruptive events within the considered feature set.