Gold Price Forecasting Using Tuned Gated Recurrent Units: Comparing Random Search and Bayesian Optimization

Gold price forecasting has long been a significant area of empirical and academic research, given gold's role as a safe-haven asset and a hedge against market instability for investors and central banks worldwide. This paper predicts today's gold closing price based on the previous ten-day open, high, low, and closing (OHLC) prices. The Gated Recurrent Unit (GRU) was utilized for this prediction. However, GRU, like many other deep learning models, has numerous hyperparameters, the tuning of which directly impacts its performance. To address this, two common hyperparameter tuning methods, namely random search (RS) and Bayesian Optimization (BO), were employed. To determine the superior tuning method, RMSE, MAE, MAPE, and R۲, as well as tuning time, were compared using the non-parametric Mann-Whitney U test. Statistical analysis indicates that with ۹۵% confidence, there is no statistically significant difference in any of the evaluated metrics between the two tuning methods. Only with approximately ۹۰% confidence can it be stated that Bayesian Optimization tunes the GRU more rapidly. In terms of performance metrics, the best parameter setting was achieved through random search, resulting in MAPE = ۱.۷۹% and R² = ۹۹.۰۷%. To the best of my knowledge, no comprehensive study to date has compared RS and BO tuning strategies in GRU-based gold price forecasting using multi-day OHLC data and a statistical method, highlighting the novelty of this research.

Keywords:

gold price , forecasting , gated recurrent unit (GRU) , hyperparameter tuning , random search , Bayesian Optimization