Strategic Evaluation and Implementation of Machine Learning Models for Predicting Price Volatility in the Steel Commodity Market: A Comparative Study with the Foreign Exchange Market for Commercial Decision-Making
TL;DR
LSTM was the most accurate forecasting model, whereas DQN was the most effective model for converting market information into profitable and risk-controlled commercial decisions and to evaluate their usefulness for commercial decision-making.
Abstract
This study aimed to implement and strategically compare machine-learning and deep-learning models for predicting price volatility in the Iranian steel commodity market and the EUR/USD foreign-exchange market and to evaluate their usefulness for commercial decision-making. This applied-developmental study used a descriptive-analytical design based on secondary time-series data from January 2019 to January 2024. The foreign-exchange dataset included hourly and daily EUR/USD opening, high, low, closing, and volume data extracted from MetaTrader 5. The steel dataset included daily and weekly prices of steel billets and sheets traded on the Iran Mercantile Exchange, together with the unofficial exchange rate, global 62% iron-ore prices, and seasonal demand indicators. After outlier correction, linear interpolation, logarithmic differencing, and feature engineering, Random Forest, XGBoost, Long Short-Term Memory, and Deep Q-Network models were implemented. Model validation was conducted through walk-forward validation using a three-year training window and a one-year out-of-sample testing window. Predictive and strategic performance were assessed through forecasting errors, directional accuracy, Sharpe ratio, maximum drawdown, win rate, and profit factor. LSTM significantly outperformed Random Forest and XGBoost in both markets, producing the lowest forecasting errors and the highest directional accuracy and explanatory power. In the steel market, LSTM achieved an RMSE of 0.0062, directional accuracy of 70.80%, and R² of 0.77. However, DQN generated the strongest strategic performance, with an annualized return of 28.90%, Sharpe ratio of 1.86, maximum drawdown of 8.70%, and profit factor of 1.84. Bootstrap comparisons confirmed that DQN significantly outperformed LSTM in risk-adjusted trading performance. The steel-market models generally exceeded the EUR/USD models because fundamental variables improved predictability. LSTM was the most accurate forecasting model, whereas DQN was the most effective model for converting market information into profitable and risk-controlled commercial decisions.