The Role of Sparse Training and Evolutionary Optimization in Volatility Forecasting Machine Learning Models
| dc.contributor.advisor | Muñoz-Elguezábal, Juan F. | |
| dc.contributor.author | Arriaza-Alonzo, Diego F. | |
| dc.date.accessioned | 2024-06-26T18:05:01Z | |
| dc.date.available | 2024-06-26T18:05:01Z | |
| dc.date.issued | 2024-05 | |
| dc.description.abstract | ETH is the native cryptocurrency of the Ethereum network, one of the most prominent blockchains for its intelligent contracts and diverse ecosystem of decentralized projects. In this research it is studied the problem of ETH/USDT 10 min short-term volatility forecasting by exploiting volatility history, order book data and public trades data 30 minutes prior. On the one hand, order book data consists of buy and sell orders over time and, on the other hand, public trades are orders executed. It is possible to calculate features from both sources that can be used as predictors for models. For the first experiment GARCH(1,1), LSTM with one layer of 100 neurons and an Encoder-Decoder with the Encoder with one LSTM of 100 neurons and the Decoder with one LSTM of 100 neurons are the models selected for volatility predictions. For the second experiment, GARCH is excluded due to poor performance on the first experiment. 10 T-Folds-SV were created omitting 50 minutes between Training and Validation sets to avoid leakage and by KL Divergence only five folds were selected that have the characteristic of being different from each other and provide unique information. With this experiment the RAM consumption was significantly reduced and the results were similar to the first one. Hyperparameter Optimization with less data is now possible and is done by Genetic Algorithms. After three generations of 750 models for both LSTM and Encoder-Decoder it was possible to find the best hyperparameter values and the LSTM best model outperformed its counterparty. | |
| dc.identifier.citation | Arriaza-Alonzo, D. F. (2024). The Role of Sparse Training and Evolutionary Optimization in Volatility Forecasting Machine Learning Models. Trabajo de obtención de grado, Maestría en Ciencia de Datos. Tlaquepaque, Jalisco: ITESO. | |
| dc.identifier.uri | https://hdl.handle.net/11117/10968 | |
| dc.language.iso | eng | |
| dc.publisher | ITESO | |
| dc.rights.uri | https://creativecommons.org/licenses/by-nc/4.0/deed.es | |
| dc.subject | ETH/USDT | |
| dc.subject | Cryptocurrency | |
| dc.subject | LSTM | |
| dc.subject | Encoder-Decoder | |
| dc.subject | Genetic Algorithm | |
| dc.title | The Role of Sparse Training and Evolutionary Optimization in Volatility Forecasting Machine Learning Models | |
| dc.title.alternative | El rol del entrenamiento disperso y la optimización evolutiva en el pronóstico de la volatilidad con modelos de Machine Learning | |
| dc.type | info:eu-repo/semantics/masterThesis | |
| dc.type.version | info:eu-repo/semantics/acceptedVersion |
Archivos
Bloque original
1 - 1 de 1
No hay miniatura disponible
- Nombre:
- Paper_Iteso_final.pdf
- Tamaño:
- 4.49 MB
- Formato:
- Adobe Portable Document Format
- Descripción:
- ETH is the native cryptocurrency of the Ethereum network, one of the most prominent blockchains for its intelligent contracts and diverse ecosystem of decentralized projects. In this research it is studied the problem of ETH/USDT 10 min short-term volatility forecasting by exploiting volatility history, order book data and public trades data 30 minutes prior. On the one hand, order book data consists of buy and sell orders over time and, on the other hand, public trades are orders executed. It is possible to calculate features from both sources that can be used as predictors for models. For the first experiment GARCH(1,1), LSTM with one layer of 100 neurons and an Encoder-Decoder with the Encoder with one LSTM of 100 neurons and the Decoder with one LSTM of 100 neurons are the models selected for volatility predictions. For the second experiment, GARCH is excluded due to poor performance on the first experiment. 10 T-Folds-SV were created omitting 50 minutes between Training and Validation sets to avoid leakage and by KL Divergence only five folds were selected that have the characteristic of being different from each other and provide unique information. With this experiment the RAM consumption was significantly reduced and the results were similar to the first one. Hyperparameter Optimization with less data is now possible and is done by Genetic Algorithms. After three generations of 750 models for both LSTM and Encoder-Decoder it was possible to find the best hyperparameter values and the LSTM best model outperformed its counterparty.