AI-driven predictions of industrial metal prices on the London metal exchange

Abstract

The battery manufacturer Northvolt aims to reduce inventory risks. In an increasingly competitive industrial metals market, protecting inventory value by employing hedging strategies is a key component of lowering costs. Using artificial intelligence for reliable predictions of short term prices is an interesting prospect for improving such strategies. This study uses data from the London Metal Exchange, including cash and threemonth futures contracts, along with stock volume, to predict next-day cash prices for six industrial metals. These predictions are compared to a baseline random walk model using the metrics mean squared error (MSE), mean absolute error (MAE), R2 , and directional accuracy. Predictions are made using an Echo State Network (ESN), with optimized parameters chosen by a Genetic Algorithm (GA). The network is trained offline with ridge regression and online with stochastic gradient descent. The performance of the ESNGA setup is validated using chaotic systems (Mackey-Glass and Lorenz equations) before being applied to metals futures data. The data is split into four different sets: a warmup set necessary for ESNs, a training set used for offline training, a validation set to measure GA performance, and a test set of unseen data to ensure the network generalizes well. The GA optimization significantly reduced prediction errors in the Mackey-Glass and Lorenz systems, with MSE values improving from 1.2E-8 to 5.3E-12 and from 3.3E-2 to 8.8E-7, respectively, on the validation set. For metals futures data, the GA enhanced ESN performance, outperforming the random walk across all metrics on the validation set for all six metals. However, slight modifications were necessary to achieve superior performance on the unseen test set. Superior results were achieved for four metals across all metrics, while the results for the remaining two metals were mixed. The GA effectively optimizes ESN parameters for systems with clear, deterministic dynamics but encounters challenges when applied to stochastic futures data, particularly due to the risk of overfitting the validation set. While the ESN-GA method does not consistently outperform the random walk on unseen test data, it demonstrates significant potential. Further exploration with alternative configurations, additional data types, and more robust validation techniques is warranted to enhance its practical applicability.