An Improved Machine Leaning, Wavelet-ARIMA Approach for Forecasting Spodumene Prices

Abstract

The thesis addresses the urgent need for reliable forecasting of spodumene, a hard-rock source of lithium whose demand is surging with the global transition to electric vehicles and energy storage. Because public price data only span 2018-2025 and are characterized by sharp booms and corrections, the development of a a leakage-safe framework that combines classical time-series models (ARIMA), discrete wavelet transforms and machine learning learners (Elastic Net, XGBoost and LSTM). The dataset was sourced from China Spodumene Li2O6% CIF (Asian metal limited) index from Bloomberg which had 379 observations from 14th Jan 2018–28th Jul 2025 which was aligned to a Sunday index, log-transformed, differenced and augmented with calendar and lag features. A chronological train–validation–test split and rolling-origin cross-validation are used to assess two forecasting horizons 2 weeks (short-term operational) and 12 weeks (medium-term planning). Naive persistence and ARIMA models serve as transparent baselines wavelet enhanced, and residual hybrids test whether routing frequency components to appropriate learners improves accuracy. Performance is judged via MAE, RMSE and MAPE, with residual whiteness and parity plots ensuring model adequacy. Results show that at the two-week horizon, wavelet-based hybrids particularly those combining elastic net or gradient-boosted trees with ARIMA reduce test MAE by roughly one-fifth relative to wavelet–ARIMA and improve robustness over Naive and ARIMA benchmarks. Stand-alone XGBoost also performs well, while deep LSTM models overfit the short series. At the twelve-week horizon, errors grow markedly across all models wavelet–ARIMA performs poorly and hybrid gains narrow, reflecting multi-step uncertainty and data scarcity. The study concludes that frequency-aware, machine-learning–augmented models offer practical value for short-horizon lithium procurement and hedging, whereas medium-term forecasts remain uncertain. It recommends future research incorporating exogenous market drivers and adaptive learning to strengthen medium-horizon reliability.