Comparison of different Artificial Neural Networks for Bitcoin price forecasting

Overview

• This study investigates how the length of input sequences affects the accuracy of predicting cryptocurrency returns using Artificial Neural Networks (ANNs).
• The researchers use the Mean Absolute Error (MAE) as a threshold to exclude small returns and focus on evaluating the accuracy of predicted returns above this threshold.
• They compare four different sequence lengths: 168 hours (7 days), 72 hours (3 days), 24 hours, and 12 hours, each with a 2-hour return prediction interval.
• The findings highlight how sequence length impacts prediction accuracy and emphasize the potential for optimizing sequence configurations in financial forecasting models.

Plain English Explanation

The researchers wanted to understand how the length of the data input sequence fed into an Artificial Neural Network (ANN) model affects the accuracy of predicting cryptocurrency returns. They used a measure called Mean Absolute Error (MAE) as a threshold, excluding small returns below this threshold and focusing on evaluating the accuracy of larger predicted returns.

The team compared four different sequence lengths: 168 hours (7 days), 72 hours (3 days), 24 hours, and 12 hours. In each case, the model was used to predict returns 2 hours into the future. The results showed that the length of the input sequence had a significant impact on the prediction accuracy. This suggests that optimizing the sequence configuration could be an important consideration when developing financial forecasting models using ANNs.

Technical Explanation

The researchers designed an experiment to investigate the impact of varying the length of input sequences on the accuracy of predicting cryptocurrency returns using Artificial Neural Networks (ANNs). They used the Mean Absolute Error (MAE) as a threshold to exclude small returns, focusing their evaluation on the accuracy of predicted returns that exceeded this threshold.

Four different sequence lengths were compared: 168 hours (7 days), 72 hours (3 days), 24 hours, and 12 hours. Each sequence was used to predict returns 2 hours into the future. The researchers evaluated the prediction accuracy using the MAE metric, comparing the results across the different sequence lengths.

The findings reveal that the length of the input sequence has a significant influence on the prediction accuracy. This suggests that optimizing the sequence configuration could be a valuable approach for enhancing the performance of financial forecasting models using ANNs. The researchers highlight the potential for this type of time series analysis to improve the accuracy of cryptocurrency return predictions.

Critical Analysis

The paper provides a solid experimental design and rigorous analysis to investigate the impact of sequence length on the accuracy of cryptocurrency return predictions using ANNs. However, the researchers acknowledge that their study is limited to a single dataset and a specific prediction interval of 2 hours.

Further research could explore the generalizability of these findings by evaluating the approach on a broader range of cryptocurrency data and considering different prediction intervals. Additionally, the paper does not delve into the potential reasons why certain sequence lengths may be more effective than others, which could be an interesting area for future investigation.

Overall, the study offers valuable insights into the importance of optimizing sequence configurations in ANN-based financial forecasting models, though additional research is needed to fully understand the nuances and broader implications of this approach.

Conclusion

This study demonstrates the significant impact that the length of input sequences can have on the accuracy of predicting cryptocurrency returns using Artificial Neural Networks (ANNs). By excluding small returns and focusing on evaluating the accuracy of larger predicted returns, the researchers were able to highlight the potential for optimizing sequence configurations to enhance the performance of financial forecasting models.

The findings underscore the need for careful consideration of input sequence lengths when developing ANN-based models for cryptocurrency price prediction and other financial time series analysis tasks. While further research is needed to fully explore the generalizability and underlying mechanisms, this study provides valuable insights for practitioners and researchers in the field of financial technology and machine learning.