top of page
Image by SJ Objio
Title: Different Techniques for Predicting Mineral Product Prices
Organization

Vale - University of New South Wales

Period

June 2021 - June 2023

Status

Completed

Est. Budget

USD 0.9M

Est. Business Benefit

USD 7.0M

Indicative Value-to-cost

7.8x

Keywords

Mineral product prices, Chaos Theory, Machine Learning, nonlinear dynamics, price forecasting, market intelligence

Benefit basis

Estimated value from improved price-risk insight, better long-term planning assumptions, improved timing of commercial decisions, and reduced forecast-error exposure.

Business Challenges 
Mineral product prices are difficult to forecast because time-dependent market behaviours, macroeconomic conditions, supply-demand shocks, inventory cycles, energy cost, currency effects, geopolitical events, and market sentiment influence them. Traditional econometric, stochastic-Gaussian, and basic time-series methods may not capture the nonlinear and dynamic behaviour of mineral-product markets. The project needed a method that could represent temporal dynamics while maintaining economic interpretability.
Expanded Technical Solution
The project investigated the integration of Chaos Theory (CT) and Machine Learning (ML) for mineral-product price prediction. Chaos Theory was used to characterise system dynamics through time-delay analysis, embedding dimension, phase-space reconstruction, and nonlinear behaviour. Machine Learning was used to identify hidden predictive patterns and forecast price behaviour. The combined approach aimed to overcome limitations of using CT or ML alone by connecting dynamic-system representation, variable influence, temporal relationships, and predictive modelling.
Technical Work Packages/Methods
  • Prepared mineral-product price series and related explanatory variables for time-series and nonlinear-dynamics analysis.
  • Used time-delay embedding, embedding-dimension estimation, phase-space reconstruction, recurrence analysis, and nonlinear-dynamics diagnostics.
  • Applied ML models such as Support Vector Regression, Random Forest, gradient boosting, recurrent neural networks, and ensemble forecasting methods.
  • Compared CT-only, ML-only, and hybrid CT-ML approaches using walk-forward testing and error/robustness metrics.
  • Added interpretability analysis to check whether selected variables and hidden patterns could be linked to economic logic.
Key Deliverables
  • Hybrid CT-ML methodology for mineral-product price forecasting.
  • Benchmark comparison against econometric, stochastic, and standard time-series methods.
  • Technical framework for long-term price-trend prediction and market behaviour interpretation.
  • Guidance on model limitations, uncertainty, and economic rationality.
Results and Value Created
The project clarified the technical limitations of traditional price forecasting and demonstrated how nonlinear dynamics and machine learning can be combined to improve representation of mineral-product market behaviour. It created a more realistic foundation for long-term trend forecasting and price-risk analysis. The estimated benefit reflects improved planning assumptions, reduced exposure to forecast error, better commercial timing, and improved long-term scenario planning.
bottom of page