Application of LSTM recurrent neural networks for bitcoin price prediction

The growing interest in cryptocurrencies such as Bitcoin highlights the need for effective predictive models in this volatile market. This study developed and trained a model based on the Long Short-Term Memory (LSTM) Recurrent Neural Network architecture to forecast Bitcoin values with a low error percentage. The results confirmed the effectiveness of the LSTM model in predicting Bitcoin prices, demonstrating its ability to handle the high volatility characteristics of this market. Hypotheses regarding the efficiency of shorter versus longer lookback periods and the influence of data volume on model performance were tested. The experiments showed that increasing the volume of data used in training significantly increases the accuracy of predictions, evidenced by the lower error rates (MAPE and RMSE) obtained with larger data volumes. However, a saturation point was observed, after which further increases in data volume did not result in significant improvements. Regarding the lookback period, the results indicated that 30-day periods presented with the best performance, with lower forecast errors. Very short or very long lookback periods tend to increase the error, which highlights the importance of proper window selection for this type of model. Statistical analysis confirmed the significant influence of data volume and lookback period on model performance, although the interaction between these factors did not show statistical significance.