We aim to identify an effective approach to constructing robust predictive models. It is well-recognized in the academic community that, in addition to single models, hybrid models have also garnered significant attention. However, the selection of models to incorporate into a hybrid framework and the methodology for constructing such models remain topics of ongoing discussion.
This study compares the predictive performance of multiple common machine learning models, including those beyond the models already discussed above, to identify those suitable for ETH price prediction and explore the possibility of constructing hybrid models. We selected ETH price data from Binance for the past 30 days (From November 13, 2024, to December 12, 2024) at a one-minute interval as the training set, totaling 42,897 records. The training data remained consistent across all models to ensure a fair comparison.
To validate the predictive capabilities of the models, we used the most recent 1,000 one-minute data points for inference validation. Specifically, we used the first 999 data points for inference prediction and compared the predicted results with the latest real data point. This process was repeated 100 times, and line graphs were generated to visually display the prediction results. Additionally, we calculated the Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Mean Absolute Error (MAE) for each model as quantitative evaluation metrics.
This study compared a total of seven common machine learning models, including:
- Ordinary Least Squares Model (OLS)
- Kernel Ridge Regression Model
- Linear Regression Model
- Bayesian Ridge Regression Model
- Autoregressive Integrated Moving Average Model (ARIMA)
- Autoregressive Model (AR)
- Support Vector Regression Model (SVR)
Validation
| Model name | MSE | RMSE | MAE |
|---|---|---|---|
| Ordinary Least Squares (OLS) | 3.343782637 | 1.828601279 | 1.307054565 |
| Kernel Ridge Regression | 3.343814207 | 1.828609911 | 1.307063295 |
| Linear Regression | 3.344748424 | 1.828865338 | 1.307083291 |
| Bayesian Ridge Regression | 3.345167898 | 1.828980016 | 1.307235138 |
| Autoregressive Integrated Moving Average (ARIMA) | 3.456699 | 1.859219998 | 1.3409 |
| Autoregressive (AR) | 3.558250784 | 1.886332628 | 1.37225902 |
| Support Vector Regression (SVR) | 8.90060058 | 2.983387434 | 2.440409737 |
According to the table, the experimental results showed that the Support Vector Regression Model had the worst predictive performance. The other six models basically aligned with the real values in terms of prediction trends, with similar predictive performances. These findings suggest that regression models (such as OLS, Kernel Ridge Regression, Linear Regression, and Bayesian Ridge Regression) exhibit better predictive capabilities when historical data is relatively abundant.
Furthermore, we considered the application scenarios of time series models. When historical data is scarce, time series models may have an advantage. In this study, ARIMA had slightly lower errors compared to other time series models. Therefore, when selecting a time series method, the ARIMA model may be a more suitable choice.
In summary, this study compared the predictive performances of multiple machine learning models to identify those suitable for ETH price prediction and initially explored the possibility of constructing hybrid models.
Future research can further investigate the methods for constructing hybrid models and how to select the optimal combination of prediction models under different levels of data abundance.
The comparison between the predicted values and actual values, as well as the observed trends for all the models mentioned in this study, are presented as follows:
Ordinary Least Squares Model (OLS)
Kernel Ridge Regression Model
Linear Regression Model
Bayesian Ridge Regression Model
Autoregressive Integrated Moving Average Model (ARIMA)
Autoregressive Model (AR)
Support Vector Regression Model (SVR)
