Woraphot Watthusin

This study presents a Graph-Based Adaptive Fusion Network (GAFN) for forecasting the closing prices of Bitcoin, Ethereum, Litecoin, and XRP at 7-, 14-, and 30-day horizons. GAFN combines a BiLSTM encoder, a graph attention network constructed from training-period return correlations, and target-conditioned adaptive fusion. It predicts standardized future log returns as intermediate outputs, which are converted to closing-price forecasts for evaluation. The experiments used 3131 aligned daily observations from Yahoo Finance covering 9 November 2017 to 5 June 2026. The data were divided chronologically into training, validation, and confirmatory test partitions. The final 156 observations, covering 1 January to 5 June 2026, yielded 97 forecast origins after sequence generation. Feature and target standardization and graph construction used the training partition only; model selection and residual scaling used the validation partition. GAFN was compared with Persistence, recurrent, fusion, attention-based, and graph-based models using asset-level NRMSE and an unweighted macro-average across the four assets. Each trainable model was evaluated over five random seeds. Several simpler models achieved lower macro-averaged NRMSE than GAFN at 7 days. At 14 days, GAFN tied with LSTM at the reported precision. At the 30-day horizon, GAFN achieved a macro-averaged test NRMSE of 0.03997 ± 0.00019, compared with 0.03979 for Persistence and 0.04062 for Concatenation Fusion. GAFN had the lowest mean NRMSE among the trainable models at this horizon, but the observed differences were not statistically significant after adjustment for multiple comparisons. The ablation results also showed no consistent improvement from any individual component across all three horizons. GAFN was therefore competitive in this experiment, but the results do not establish superiority beyond the tested assets and period.

This study presents a comparative evaluation of machine learning models for domain adaptation in multiclass sentiment classification. While sentiment analysis aims to categorize opinions as positive, neutral, or negative, adapting models across domains remains a significant challenge due to differences in vocabulary, writing style, and sentiment expression. Models trained on a specific domain often fail to generalize effectively to others. To solve this problem, we evaluate how well six models-logistic regression, support vector machine (SVM) with a linear kernel, random forest, convolutional neural network (CNN), long short-term memory (LSTM), and BERT-perform on sentiment data from books, beauty & personal care, and automotive categories. The evaluation uses Amazon review data and measures performance via accuracy, F1 score, and Area Under the ROC Curve (AUC). Results indicate that BERT consistently outperforms all other models due to its attention-based transformer architecture, which captures nuanced contextual information across diverse domains. CNN and LSTM models also perform well, particularly in domain-specific settings, with CNN excelling in extracting local features and LSTM in modeling sequential relationships. Traditional models, such as logistic regression and SVM, show limitations in generalizability, while random forest demonstrates stable yet moderate performance. These findings highlight the strengths and trade-offs of each approach for effective cross-domain sentiment classification.