Performance Optimisation of Regression-Based Machine Learning Models to Predict Mineral Land Prices for Tokenization

The tokenization of in-ground mineral land involves the digital conversion of physical mineral assets by issuing security tokens registered on a blockchain. This transformative approach offers the potential to fractionalize mineral land ownership, enhancing market liquidity, accessibility, and transparency. However, the accuracy of mineral land price prediction is paramount for its success. Historically, traditional methods such as geostatistical techniques, geometric modelling, kriging, and Inverse Distance Weighting were employed for price prediction. Yet, geological complexities and unpredictable external factors have posed formidable challenges. Hence, it is critical that mining companies have the capability to accurately predict the mineral land price to evaluate the feasibility of their mining operations and provide a fair market valuation to critical stakeholders. To address these challenges, this paper empirically evaluates three machine learning algorithms: multiple linear regression, gradient boosting models, and deep neural network models (DNN) for mineral land price prediction in the context of tokenization. Leveraging a dataset from a Western Australian mineral deposit with Gold, Iron, and Nickel commodities across 45 mine sites, the results demonstrate that the DNN approach, when fine-tuned with the Keras Tuner method, outperforms other models, yielding an optimized Root Mean Square Error (RMSE) of 5154.42.