Tuning Deep Learning for Predicting Aluminum Prices Under Different Sampling: Bayesian Optimization Versus Random...

Area of Science:

• Computational finance
• Machine learning applications
• Econometrics

Background:

• Aluminum prices exhibit complex, non-linear dynamics.
• Accurate price forecasting is crucial for market participants.
• Traditional models may struggle with the intricate behavior of commodity markets.

Purpose of the Study:

• To implement and compare deep learning models for aluminum price forecasting.
• To evaluate the efficacy of Bayesian Optimization (BO) against Random Search (RS) for model tuning.
• To identify the optimal model for predicting aluminum spot prices across various time horizons.

Main Methods:

• Utilized Long Short-Term Memory (LSTM) and Deep Feedforward Neural Networks (FFNN) as deep learning models.
• Employed Support Vector Regression (SVR) as a baseline comparative model.
• Tuned models using Bayesian Optimization (BO) and Random Search (RS) on daily, weekly, and monthly data.

Main Results:

• The LSTM model optimized with Bayesian Optimization (LSTM-BO) demonstrated superior performance.
• LSTM-BO consistently achieved the lowest Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE), alongside the highest R-squared (R²).
• Models tuned with Bayesian Optimization generally outperformed those tuned with Random Search across all tested configurations.

Conclusions:

• LSTM-BO is highly effective for aluminum spot price forecasting.
• Bayesian Optimization is a more efficient tuning method than Random Search for these predictive models.
• Deep learning, particularly LSTM with BO, offers significant advantages for capturing complex financial market behavior.