딥러닝 기반 시계열 분석 모델의 불확실성 정량화 비교 연구

Abstract

With the advancement of artificial intelligence, machine learning, and deep learning, their applications in various industries, particularly finance, have increased. However, interpreting predictions from deep learning models poses challenges, especially in finance where result interpretation is important. This study aims to determine the uncertainty of stable deep learning models, despite changes in the model like dropout, to establish standards for reliable models and identify those detecting anomal data through model uncertainty. In the experiment, the traditional statistical model ARIMA and deep learning models mainly used for time series analysis, CNN, LSTM, MLP, and CNN-LSTM. Uncertainty was measured from a Bayesian perspective using MC Dropout. The experimental results confirmed that deep learning models performed better than statistical models for various patterns of time series data. It was observed that, even if the performance was not the best, LSTM based models exhibited low uncertainty, indicating stability. Consequently, this study highlights the importance of considering uncertainty along with accuracy in model selection. Moreover, it was confirmed that models with higher uncertainty are suitable for anomaly detection, making CNN based models particularly fitting for this purpose.