Classificação de retornos utilizando dados de alta frequência no mercado de bitcoins

Abstract

In the cryptocurrency market, Bitcoin stands out as the most accepted traded in the world. However, due to its high volatility, the prediction of price behaviors, in special, the trend classification, becomes a challenge and valuable to investors. In this sense, in recent years, some studies have been proposed based on machine learning techniques to classify reliable trading signals from the Bitcoin return series. Thus, the present work explores Bitcoin's trend behaviors, seeking to classify its returns using high-frequency data from limit order book in terms of US dollar (BTCUSD) and euro (BTCEUR). The proposed methodology seeks to accurately predict returns, creating opportunities that better support trading strategies. Therefore, from the Bitcoin historical series, technical indicators (commonly used in financial markets), market variables and the own series of returns for different time intervals were extracted. It is worth mentioning that the data have a frequency of at most one second for each update of the entire order book, thus characterizing the high-frequency. Subsequently, these data are submitted to the inputs of trend classifiers based on Artificial Neural Networks and XGBoost, which were trained and validated in three months of trading, from January to March 2019, composing approximately 5 million market updates for each currency base. In a first analysis, the behaviors of the models for each base currency were observed, it is possible to note that the market for BTCUSD proves to be more efficient related to BTCEUR. Consequently, the machine learning models obtained more reliable and stable results over time for BTCUSD. Based on this initial result, sought to validate the predictive models for different training and validation configurations. In this second analysis, it was noted that the trend classification is more accurate for short intervals (between 1 and 3 minutes) in both base currencies, which is justified due to the high volatility in Bitcoin prices in short intervals. In a third analysis, it was observed that as we increased the number of training days for classifiers, in some cases, was noted a gradual loss in classification (between 1% and 2%). Still, in a fourth analysis, one-vs.-rest was applied to each trend behavior, there is a considerable performance increase in the classification in relation to the base case (between 1% and 7%)