Bitcoin and cryptocurrencies : comte-leftist hybrid explanations and time-series classification
The “Global Race For AI” has driven the pursuit of a strategy known as “AI for society”. One of the key outcomes of this strategy was the General Data Protection Regulation (GDPR), an European regulation enforced on May 28, 2018, which established the “right to explana- tion”. This regulation signif...
| Author: | |
|---|---|
| Format: | master thesis |
| Status: | Published version |
| Publication Date: | 2025 |
| Country: | Brasil |
| Institution: | Universidade Federal de Pernambuco (UFPE) |
| Repository: | Repositório Institucional da UFPE |
| Language: | English |
| OAI Identifier: | oai:repositorio.ufpe.br:123456789/64973 |
| Online Access: | https://repositorio.ufpe.br/handle/123456789/64973 |
| Access Level: | Open access |
| Keyword: | Explainable AI Time-Series Classification Hybrid Explanations. |
| Summary: | The “Global Race For AI” has driven the pursuit of a strategy known as “AI for society”. One of the key outcomes of this strategy was the General Data Protection Regulation (GDPR), an European regulation enforced on May 28, 2018, which established the “right to explana- tion”. This regulation significantly contributed to the rise of Explainable AI (XAI). Amidst this wave of technological innovation, the market around digital assets, commonly known as the cryptocurrency market has benefited from research into Artificial Intelligence (AI) and Machine Learning (ML) based trading systems. However, these systems often rely on black-box mod- els, making explainability crucial. In this context, this work applies Machine Learning models specifically designed for Time-Series Classification (TSC) and proposes a novel hybrid method that provides time-series-based explanations. After collecting Bitcoin and cryptocurrency data from a crypto exchange, the data is processed and trained using ML tabular models, ML TSC models, and Deep Learning (DL) models. The study evaluates uncertainty, performance, and explainability through a hybrid explainability model, which merges COMTE (a counterfactual TSC explanation method) and LEFTIST (a time-point-based method that provides feature importance for each timestep). The results show that the Multiple Representations Sequence Miner (MRSQM) TSC model achieved a strong performance, while ML tabular models did not differ significantly from TSC models. DL models, however, performed poorly, particularly in the second experiment. Uncertainty analysis revealed notable differences in uncertainty estimation, and the COMTE-LEFTIST hybrid explainability model successfully provided hybrid explana- tions. The hybrid model performed particularly well in the first experiment, which focused on univariate time-series data, while the second experiment, involving multiple time-series in a tabular format, presented additional challenges. In conclusion, this is among the first works to apply TSC methods to Bitcoin and other cryptocurrencies, while also proposing a novel hybrid explainability approach for TSC, encouraging further research and development in the field. |
|---|