Claim Missing Document
Check
Articles

Found 2 Documents
Search

Stock Price Prediction Using Backpropagation ANN: Case Study of ADMR (2023–2025) Muhammad Khozin; Zainal Abidin; Totok Chamidy
Journal of Information System and Informatics Vol 7 No 4 (2025): December
Publisher : Asosiasi Doktor Sistem Informasi Indonesia

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.63158/journalisi.v7i4.1347

Abstract

This study develops an Artificial Neural Network (ANN) backpropagation model for predicting stock prices using ADMR stock data from 2023 to 2025, obtained from Yahoo Finance. Given the inherent volatility and unpredictability of stock prices, accurate forecasting plays a crucial role in investment decision-making. ANN models are particularly effective for capturing complex, non-linear relationships and patterns in financial data, which traditional statistical models may fail to address. In this research, various configurations were tested by adjusting the number of hidden neurons (5, 10, and 15) and learning rates (0.1, 0.3, and 0.5). The optimal model architecture was found to be 3-10-1, consisting of three input neurons, ten hidden neurons, and one output neuron, which achieved the best prediction performance with a Mean Absolute Percentage Error (MAPE) of 2.26%. This model was trained with a learning rate of 0.3 and completed in 915 iterations. However, the model's predictive capabilities are constrained by its reliance on historical stock prices alone, excluding external factors such as macroeconomic indicators, market sentiment, or trading volume, which may improve its generalization and overall accuracy. Future work could integrate these variables for better robustness and predictive power.
Ekstraksi fitur tekstur kerusakan situs warisan budaya pascabencana menggunakan deep features efficientnet-B7 Muhammad Hatif; Sri Harini; Ririen Kusumawati; Zainal Abidin; Muhammad Ainul Yaqin; Agung Teguh Wibowo Almais
JURNAL ELTEK Vol. 24 No. 2 (2026): October 2026
Publisher : Politeknik Negeri Malang

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.33795/eltek.v24i2.9886

Abstract

Situs warisan budaya sangat rentan mengalami kerusakan akibat bencana alam, sehingga memerlukan metode inspeksi yang cepat, akurat, dan objektif untuk mendukung upaya konservasi. Penelitian ini mengusulkan pendekatan ekstraksi dan karakterisasi fitur tekstur otomatis melalui pemanfaatan fitur mendalam (deep feature extraction) menggunakan arsitektur EfficientNet-B7 sebagai landasan untuk deteksi kerusakan. Metode ini memanfaatkan model pre-trained untuk mengekstraksi peta fitur dari lapisan awal block2a_expand_activation, yang peka terhadap detail tekstural seperti tepi, retak, dan variasi permukaan. Dari 144 saluran fitur yang dihasilkan, dihitung tiga deskriptor statistik sederhana per saluran—Kontras (standar deviasi), Energi (rata-rata kuadrat aktivasi), dan Homogenitas (turunan kontras)—sebagai deskriptor tekstur kuantitatif (bukan fitur GLCM klasik). Sebagai studi awal (proof of concept), pendekatan ini diujicobakan pada satu citra kerusakan plafon masjid pascabencana dan berhasil mengkarakterisasi variasi tekstur antar-saluran, menghasilkan 432 fitur numerik (3 statistik × 144 saluran). Secara konseptual, dibandingkan metode tradisional seperti Gray Level Co-occurrence Matrix (GLCM), pendekatan ini berpotensi lebih unggul dalam hal otomasi, kedalaman fitur, dan skalabilitas karena tidak memerlukan parameterisasi manual, meskipun perbandingan empiris langsung masih perlu dilakukan. Meskipun membutuhkan sumber daya komputasi yang signifikan, metode yang diusulkan menawarkan solusi yang efisien dan andal untuk analisis kerusakan otomatis, yang dapat mendukung prioritisasi dan pemulihan situs warisan budaya pascabencana. Abstract Cultural heritage sites are highly vulnerable to damage from natural disasters, necessitating fast, accurate, and objective inspection methods to support conservation efforts. This study proposes an automated approach for texture feature extraction and characterization through deep feature extraction using the EfficientNet-B7 architecture, serving as a foundation for damage detection. This method utilizes a pre-trained model to extract feature maps from the initial block2a_expand_activation layer, which is sensitive to textural details such as edges, cracks, and surface variations. From the 144 feature channels generated, three simple per-channel statistical descriptors were computed—Contrast (standard deviation), Energy (mean squared activation), and Homogeneity (derived from contrast)—as quantitative texture descriptors. As a proof of concept, the approach was tested on a single mosque ceiling damage image and successfully characterized inter-channel texture variations, producing 432 numerical features (3 statistics × 144 channels). Conceptually, compared to traditional methods such as Gray Level Co-occurrence Matrix (GLCM), this approach is potentially superior in terms of automation, feature depth, and scalability, as it does not require manual parameterization, although a direct empirical comparison remains to be conducted. Although requiring significant computational resources, the proposed method offers an efficient and reliable solution for automated damage analysis, which can support the prioritization and restoration of cultural heritage sites after disasters.