ADS Capstone Chronicles Revised

Table 1: Ridge Regression Model Performance Metrics

6.1.4 Hyperparameter Tuning and

was identified as the best-performing model during the evaluation phase.

Cross-Validation Hyperparameter tuning and cross-validation enhanced the model ’ s precision for Ridge Regression. A range of alpha values was explored, and TimeSeriesSplit was employed for cross-validation to ensure temporal integrity. The consistent discovery of the best alpha, 0.01, for all product IDs underscored the success of the tuning process in optimally configuring the model for each cryptocurrency, thereby enhancing its predictive accuracy. 6.1.5 Final Evaluation The tuned Ridge Regression model was then evaluated on the test set. The model exhibited impressive performance, with predicted values closely matching the actual percentage change values across all product IDs. This evaluation confirmed Ridge Regression ’ s efficacy and reliability for forecasting needs. The results showed the Ridge Regression model provided a balanced approach to handling multicollinearity and making accurate predictions. With this model, proceeding to the forecasting phase was done with confidence. 6.2 Model Forecasting The goal of the forecasting model was to predict the percentage change for the next seven days following the end of the train and test datasets, providing a real-time assessment of the model ’ s forecasting capabilities. This phase involved using the tuned Ridge Regression model, which

6.2.1 Data Engineering for Forecasting The journey began with meticulous data engineering, which is crucial to ensuring the robustness of forecasting efforts. This involved creating lag features for the percentage change ‘ pct_change ’ over the past seven days, a technique recommended for capturing temporal dependencies and trends in time series data. This step is vital as it allows the model to consider the recent historical context when making predictions. 6.2.2 Using Benchmark Datasets For the utmost accuracy in forecasting, a separate dataset, ‘ last_week, ’ containing the actual percentage change values for the same seven days that were aimed to be forecasted, was used. This dataset, chosen for its reliability, served as a benchmark, providing a robust measure for evaluating the model ’ s performance. 6.2.3 Forecast Evaluation To assess the model's accuracy, the forecasted percentage changes were compared to the actual values from the ‘last_week’ dataset. This comparison revealed that the model performed well for the first six days across all unique product IDs, except for BTC-USD, which showed discrepancies.

83