Hue University Journal of Science: Techniques and Technology

Short-Term Bitcoin Forecasting Across Multiple Horizons: The Effectiveness of GRU Networks

2026-06-14T15:55:59+07:00

This study proposed a GRU-based framework for short-term multi-horizon Bitcoin forecasting and validated its effectiveness through a systematic comparison with four state-of-the-art baseline architectures: Long Short-Term Memory (LSTM), Bidirectional LSTM (Bi-LSTM), one-dimensional Convolutional Neural Networks (1D-CNN), and Transformer networks over multiple time horizons. Based on daily Bitcoin price data spanning from September 17, 2014, to June 11, 2025, the models were systematically evaluated under three forecasting horizons—3-day, 5-day, and 7-day—using various input window sizes. Model performance was quantitatively assessed through five standard metrics: Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE) and the Coefficient of Determination (R²). Experimental results demonstrate that the GRU model consistently outperforms the other architectures in terms of predictive accuracy, temporal stability, and computational efficiency across all tested configurations. Notably, this study provides one of the first comprehensive benchmarks of these deep learning models under consistent experimental settings, offering practical guidance for selecting forecasting architectures in highly volatile cryptocurrency markets.

A multi-horizon CNN-GRU architecture for gold price forecasting

2026-03-02T06:49:04+07:00

Precise price forecasting in gold markets constitutes a fundamental objective for financial risk management systems operating within volatile environments. Conventional econometric frameworks and standalone deep learning models frequently exhibit insufficient capacity to capture the non-linear dependencies and chaotic dynamics that characterize these financial sequences. This research establishes a hybrid neural network architecture that integrates one-dimensional convolutional neural network (1D-CNN) layers with gated recurrent units (GRU) to synergize localized feature extraction with long-term temporal sequence modeling. The methodology incorporates a specific sliding window preprocessing algorithm to transform multivariate time-series data into structured three-dimensional tensors, thereby preserving critical cross-feature interactions for the learning process. Empirical validation utilizing price records from 2000 to 2025 confirms that the proposed Hybrid CNN-GRU framework consistently achieves predictive accuracy superior to the standalone 1D-CNN, the Hybrid CNN-Long Short-Term Memory (LSTM), and the Hybrid CNN-Dual GRU benchmarks across all experimental window configurations. Specifically, in the ten-day window configuration, the model attained a maximum coefficient of determination (R²) of 0.9962 ± 0.0008, a mean absolute error (MAE) of 0.0178 ± 0.0018, and a mean squared error (MSE) of 0.0006 ± 0.0001. These findings confirm that fusing convolutional pattern mining with streamlined recurrent gating mechanisms significantly enhances predictive precision while optimizing computational efficiency for high-dimensional financial datasets.

Impact of skeleton data quality and tracking-based refinement on human action recognition

2026-06-21T14:26:28+07:00

Skeleton-based human action recognition has attracted increasing attention because it provides a concise structural representation of human motion and is less affected by environmental factors. However, for real-deployment scenarios, the skeleton data are usually extracted automatically using human pose estimation models, resulting in joint localization errors, occlusion of missing joints, and temporal inconsistencies that affect action recognition performance. Here we systematically analyze how skeleton data quality affects the performance of a state-of-the-art Transformer-based AcTv2 model used in three input configurations: ground-truth skeletons, OpenPose-extracted skeletons, and tracking-refined skeletons. Using JHMDB datasets, we show that the performance degrades profoundly on automatically extracted skeleton data, and that our tracking-based refinement improves recognition accuracy of actions with high levels of occlusion. The results also show that an imbalance in structural distribution between action groups can also lead to instability in learning motion representations.

Personalized recommendation of tourism points of interest using classification-based machine learning

2026-07-21T08:52:39+07:00

This study proposes a two-stage personalized recommendation framework for tourism Points of Interest (POIs) based on classification techniques. Unlike conventional approaches that require detailed user–item rating matrices or fine-grained POI-level data, our model predicts each user’s most suitable POI category by leveraging demographic attributes, temporal features, and aggregated preference scores. Once the POI Category is identified, the system retrieves specific destinations from a structured POI database. Using data collected from 1,000 tourists in Hue City, Vietnam, we implement and compare five classification algorithms (Decision Tree, Random Forest, K-Nearest Neighbors, Support Vector Machine, and Naïve Bayes). Experimental results demonstrate that Random Forest achieves the highest predictive accuracy and lowest error metrics (MAE, RMSE, MSE), with strong agreement measured by the Kappa statistic. The originality of this research lies in: (i) reframing POI recommendation as a classification-based task suitable for data-scarce environments; (ii) providing empirical evidence from an emerging tourism destination; and (iii) conducting a statistical comparison of multiple algorithms using non-parametric tests (Friedman and Nemenyi). The findings lay a methodological and practical foundation for developing smart tourism systems that deliver effective personalization even with limited rating data.

Evaluating the absolute efficiency of the TCS-172 gamma scintillation survey meter for 662 keV and 1250 keV gamma rays emitted from point sources

2026-08-19T16:21:07+07:00

In this study, McGregor’s analytical method was applied to calculate the absolute efficiency of the TCS-172 GSSM equipped with a 1 in. ´ 1 in. NaI(Tl) detector. Calculations were performed for 662 keV and 1250 keV gamma rays emitted from Cs-137 and Co-60 point sources, respectively. The calculated efficiencies were compared with experimental measurements conducted at different source-to-detector distances. The calculated data were fitted over 0.5–5.0 cm and evaluated at the experimental distances. The differences ranged from −7% to 6% for 662 keV and −9% to 6% for 1250 keV, with a maximum absolute difference of 9%. The results showed that the absolute efficiency decreases with increasing distance and is higher at 662 keV than at 1250 keV due to the greater interaction probability of lower-energy gamma rays in NaI(Tl). Good agreement was observed between the calculated and measured efficiencies, confirming the applicability of McGregor’s analytical method to the TCS-172 detector. The proposed approach provides a simple and rapid tool for estimating the absolute efficiency of portable NaI(Tl)-based gamma survey meters.

Calculate the average macroscopic total cross-section of Be, C, Fe, and Pb for neutrons from a Cf-252 source based on neutron energy group model

2026-06-23T13:18:34+07:00

This study presents an analytical calculation method based on neutron energy group models for calculating the spectrum-averaged macroscopic total cross-sections of Be, C, Fe, and Pb for the Cf-252 fission neutron spectrum using neutron energy group models. Calculations were performed with 11-group and 22-group structures based on microscopic cross-section data from a nuclear data library. The calculated macroscopic total cross-sections obtained with the 11-group model were 0.3170, 0.1689, 0.2542, and 0.2107 cm⁻¹ for Be, C, Fe, and Pb, respectively, while the corresponding values from the 22-group model were 0.3041, 0.1684, 0.2507, and 0.1939 cm⁻¹. Comparison with published MCNP results showed deviations ranging from -2.2% to 2.4% for the 11-group model and from -6.2% to -1.2% for the 22-group model. In contrast, calculations based on a single neutron energy equal to the average energy of the Cf-252 spectrum produced significantly larger errors. The results demonstrate that multi-group treatment of the neutron spectrum is essential for reliable cross-section calculations and that the 11-group model provides a practical combination of accuracy and computational simplicity for educational and shielding applications.