Título: Autoencoder Neural Network Approaches for Anomaly Detection in IBOVESPA Stock Market Index
Autores: Lucas Takara, Viviana Mariani and Leandro Coelho.
Resumo:
Anomalies are patterns in data that do not conform to a well-defined notion of normal behavior. Anomaly detection has been applied to many problems such as bank fraud, fault detection, noise reduction, among many others. Some approaches to detect anomalies include classical statistical econometric methods such as AutoRegressive Moving Average (ARMA) and AutoRegressive Integrated Moving Average (ARIMA) approaches. More recently, with the progress of artificial intelligence and more specifically, machine learning, new algorithms such as one-class support vector machines, isolation forest, gradient boosting, and deep neural networks were applied to such tasks. This paper focuses on propose an anomaly detection framework for the Índice da Bolsa de Valores de São Paulo (IBOVESPA). It is a major stock market index that tracks the performance of around 50 most liquid stocks traded on the São Paulo Stock Exchange in Brazil. Exploring unsupervised autoencoder neural network algorithms, we compare the long short-term autoencoder, bidirectional long short-term autoencoder, and convolutional autoencoder models, aiming to explore the performance of these architectures for anomaly detection. Due to the ability of autoencoders to learn a compressed representation of their respective input, we train these models with standard data by minimizing the mean absolute error (MAE) loss function and evaluate them with anomalous inputs. We set a reconstruction error threshold, and in case that the reconstruction error of the test data sample is beyond it, anomalies are detected. Our results show that these models perform quite well and can be applied to real stock market data.
Palavras-chave:
Anomaly Detection, Stock Markets, Deep Learning, Autoencoders, Financial Signal Processing.
Páginas: 8
Código DOI: 10.21528/CBIC2021-37
Artigo em pdf: CBIC_2021_paper_37.pdf
Arquivo BibTeX: CBIC_2021_37.bib