超参数
计算机科学
股票市场
人工神经网络
证券交易所
计量经济学
印度尼西亚语
综合指数
时间序列
股票市场指数
金融市场
索引(排版)
金融危机
股市崩盘
市场时机
人工智能
鉴定(生物学)
机器学习
汇率
滞后
撞车
辍学(神经网络)
系列(地层学)
深度学习
库存(枪支)
循环神经网络
市场数据
新兴市场
长记忆
数据挖掘
回归
作者
Bagus Sartono,Tahniah Salsabila Elenaputri,Yenni Angraini,Gerry Alfa Dito
摘要
The prediction of stock market crises is critical for safeguarding investors and enabling regulatory authorities to implement timely interventions. This study investigates the predictive capability of long short‐term memory (LSTM) neural networks for modeling the returns of the Indonesian Composite Index (IDX Composite) and detecting potential market crashes in advance. LSTM architectures are well‐suited for financial time series due to their gated cell structure, which selectively retains or discards sequential information, enabling the capture of long‐term dependencies and nonlinear temporal patterns. We propose an LSTM with a multiple input–multiple output (MIMO) strategy to incorporate diverse predictors, including the IDX composite, global market indices, the VIX index, and exchange rates. The model is trained using daily return data from 2002 to 2023. A comprehensive hyperparameter search was conducted, optimizing the number of input/output periods, learning rate, dropout, batch size, and number of neurons. The best‐performing configuration uses two lag periods of input and one‐step‐ahead output, optimized with the Adam algorithm (learning rate = 0.0001, dropout = 0.1, 150 epochs, 32 neurons, batch size = 32). Experimental results demonstrate that the proposed MIMO‐LSTM model can detect market crash signals one day before occurrence, achieving an accuracy of 78%. These findings highlight the potential of deep learning–based sequential modeling for early crisis detection in emerging financial markets.
科研通智能强力驱动
Strongly Powered by AbleSci AI