COMPARISON OF CNN-GRU AND GRU MODELS FOR GOLD PRICE PREDICTION

Fahlefi, Muhammad Reza (2026) COMPARISON OF CNN-GRU AND GRU MODELS FOR GOLD PRICE PREDICTION. Undergraduate thesis, UPN Veteran Jawa Timur.

	Text (Cover) 22081010223.-cover.pdf Download (942kB)
	Text (Bab 1) 22081010223.-bab1.pdf Download (149kB)
	Text (Bab 2) 22081010223.-bab2.pdf Restricted to Repository staff only until 19 July 2028. Download (346kB) \| Request a copy
	Text (Bab 3) 22081010223.-bab3.pdf Restricted to Repository staff only until 19 July 2028. Download (879kB) \| Request a copy
	Text (Bab 4) 22081010223.-bab4.pdf Restricted to Repository staff only until 19 July 2028. Download (763kB) \| Request a copy
	Text (Bab 5) 22081010223.-bab5.pdf Download (145kB)
	Text (Daftar Pustaka) 22081010223.-daftarpustaka.pdf Download (160kB)

Abstract

As a globally traded commodity, gold price movements are shaped by a wide range of macroeconomic variables, posing significant challenges for conventional linear prediction techniques. This research implements and evaluates two deep learning architectures the Convolutional Neural Network, Gated Recurrent Unit (CNN-GRU) and the standalone GRU for forecasting gold prices using XAU/USD time series data. Historical gold price data in the XAU/USD pair was obtained from Investing.com covering the period from 2016 to 2026, consisting of 2,450 daily observations. Data preparation encompassed several sequential steps: removal of anomalies and missing values, derivation of stationary input features guided by statistical testing and importance scoring, Min-Max Scaling to normalize value ranges, and organization of samples into overlapping sequences via a sliding window of five trading days. Two final input features were established, namely high_low_pct_diff and rolling_std_14_ret_diff, with log_return as the target variable. The CNN-GRU model was constructed with one Conv1D layer (32 filters, kernel size 2, ReLU activation) as a feature extractor, followed by one GRU layer (64 units) for temporal modeling, and a Dense layer as the output. The GRU model was built with an identical architecture but without the CNN layer. Experiments were conducted under three data splitting scenarios: 70:15:15, 80:10:10, and 60:20:20. Evaluation results show that the CNN-GRU model consistently produced lower RMSE, MAE, and MAPE values than GRU across all scenarios. In the best-performing scenario (60:20:20), CNN-GRU achieved an RMSE of 37.0115 USD, MAE of 26.0068 USD, and MAPE of 0.8373%, while GRU yielded an RMSE of 37.1107 USD, MAE of 26.1016 USD, and MAPE of 0.8403%. Statistical validation through the Diebold-Mariano Test further confirmed CNN-GRU's advantage, with p-values consistently below the 0.05 threshold under both Squared Error and Absolute Error loss functions across all experimental scenarios. These findings confirm that the addition of a convolutional layer as a feature extractor provides a measurable and statistically significant improvement in gold price prediction accuracy based on time series data.

Item Type:

Thesis (Undergraduate)

Contributors:

Contribution	Contributors	NIDN/NIDK	Email
Thesis advisor	Puspaningrum, Eva Yulia	NIDN0005078908	evapuspaningrum.if@upnjatim.ac.id
Thesis advisor	Anggraeny, Fetty Tri	NIDN0711028201	fettyanggraeny.if@upnjatim.ac.id

Subjects:

T Technology > T Technology (General) > T385 Computer Graphics

Divisions:

Faculty of Computer Science > Departemen of Informatics

Depositing User:

Muhammad Reza Fahlefi

Date Deposited:

20 Jul 2026 03:17

Last Modified:

20 Jul 2026 05:12

URI:

https://repository.upnjatim.ac.id/id/eprint/56268