Predicting Bull and Bear Markets: A Deep Learning and Linear Regression Study in Cryptocurrencies

• João Paulo Costa e Souza UnB
• Rodolfo I. Meneguette USP
• Vinícius P. Gonçalves UnB
• Fábio L. L. de Mendonça UnB
• Francisco Airton Silva UFPI
• Geraldo P. Rocha Filho UnB / UESB

Resumo

Despite the growing popularity and increasingly widespread use of cryptocurrencies in contemporary financial markets, understanding market trends and predicting their future movements is a formidable challenge in financial analysis. Regardless of the numerous tools and techniques available, accurately forecasting market behavior remains complex due to financial data’s inherent volatility and noise. This work proposes a novel hybrid model for market trend classification and trend slope regression to address this gap. The proposed approach leverages technical indicators, autoencoders, a BiLSTM-CNN architecture, and the use of linear regression to label the trends, enhancing prediction accuracy. The core innovation lies in the integration of these advanced techniques to extract meaningful and useful patterns from the data. This methodology allows the model to better capture the complexities of financial time series, thus improving its predictive capabilities. As a proof of concept, the proposed model was evaluated using historical Bitcoin data spanning two years, with performance metrics such as accuracy, precision, recall for trend classification, and mean squared error (MSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) for trend slope regression. The results are compelling, demonstrating that the combination of linear regression labeling and the architecture proposed achieves a remarkable level of accuracy both in-sample and out-of-sample. This accuracy significantly surpasses the 50% random benchmark in the predictions of bull and bear markets. Ultimately, this research presents a promising approach to enhancing the precision of trend predictions in the volatile and unpredictable cryptocurrency market.