Claim Missing Document
Check
Articles

Found 2 Documents
Search

Hybrid Relevance and Sentiment Classification of Indonesian Gold Tweets Using Machine Learning for Market Risk Signal Extraction Kamalia, Antika Zahrotul; Indra, Indra; Wibowo, Arief; Riwurohi, Jan Everhard; Hassan, Shiza
International Journal of Advances in Data and Information Systems Vol. 7 No. 1 (2026): April 2026 - International Journal of Advances in Data and Information Systems
Publisher : Indonesian Scientific Journal

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.59395/ijadis.v7i1.1517

Abstract

This study proposes a hybrid relevance–sentiment classification framework to analyze public opinion on physical Antam gold from Indonesian Twitter data and to support exploratory market-risk signal extraction. Tweets were collected during February–November 2025, after preprocessing and text-normalized deduplication, 1,271 unique tweets were retained. The approach combines weak supervision (rule-/lexicon-based silver labels) with TF-IDF-based machine learning in two stages: (1) relevance classification to separate tweets genuinely discussing physical Antam gold from non-relevant contexts (e.g., ANTM stock/capital-market discussions), and (2) two-class sentiment classification (positive vs negative) applied to relevance-filtered tweets. Random Forest achieved the strongest relevance performance (Accuracy = 0.984; macro-F1 = 0.943; 5-fold CV macro-F1 = 0.928 ± 0.033). For sentiment classification, performance was moderate and close across models; the most stable model under cross-validation (Logistic Regression/Naive Bayes) was used for downstream aggregation. Sentiment outputs were aggregated into a monthly sentiment index for descriptive comparison with gold prices; the observed association was weak, indicating that the index is better interpreted as a risk-perception proxy rather than a direct price predictor.
Comparative evaluation of deep learning models for dried corn price prediction in east java Kamalia, Antika Zahrotul; Latansa, Choiriyatun Nisa; Rozikin, Zaenur; Herlianto, Hemdani Rahendra; Hassan, Shiza
Journal of Soft Computing Exploration Vol. 7 No. 2 (2026): June 2026
Publisher : SHM Publisher

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.52465/joscex.v7i2.48

Abstract

Forecasting dry shelled corn prices was important for supporting decision-making by farmers, traders, feed industries, and local governments. This study comparatively evaluated several deep learning models, namely Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Convolutional Neural Network 1D (CNN1D), Temporal Convolutional Network (TCN), and Transformer, for predicting dry shelled corn prices in East Java. Classical benchmark models, namely naïve, drift, and simple exponential smoothing (SES), were also incorporated into the experimental design. Using daily price data from 2020 to 2024, a 30-day lookback window, and multivariate features derived from price movements, calendar variables, and rolling statistics, model performance was assessed using MAE, RMSE, MAPE, sMAPE, and . The results showed that the naïve baseline achieved the best overall performance on the 2024 test set, while TCN was the strongest among the evaluated deep learning models. TCN obtained RMSE of 176.95 and of 0.6895, whereas the naïve baseline achieved RMSE of 20.06 and of 0.9960. Overall, all deep learning models were outperformed by the naïve persistence benchmark, indicating that greater model complexity did not automatically improve forecasting accuracy on this highly persistent price series.