Bitcoin Volatility Forecasting Software

Forecasting Software The goal of the software is to identify price volatility and, consequently, the positive and negative fluctuations that Bitcoin will have over the next 7 days. This data will be used to find a historical time series from the past that had percentage fluctuations (not based on price levels) as similar as possible. Once the most similar 7-day historical series to the predicted future series is found, the date of the historical series should be displayed on the screen. The entire software must be fully "commented" and explained in every function and calculation, clarifying the practicality and utility of each element within the model and highlighting the added value it provides. The software must include a frontend to visualize the period (between two dates) that most closely resembles the one analyzed by the forecasting. It should also provide visual representation, including charts that display the forecasting period and the identified historical period on a trading graph. When calculating the predicted volatility, it is important to account for both positive and negative volatility. This distinction will be crucial for identifying a historical period that most closely matches the forecasting period. Data Requirements The software will be based on the BTC-USD price history from the past 5 years. The data should be continuously updated, and a daily granularity is recommended. It is preferable to directly implement data retrieval using yfinance, avoiding reliance on a CSV file for management. Feature Engineering The extracted features should include: • ATR (Average True Range) • Bollinger Bands • Moving averages • Logarithmic returns (Additional suggestions and implementations are welcome.) For this specific part, training an LSTM model is advised. Modeling 1. Train a GARCH(1,1) or GARCH(2,1) model on the logarithmic returns time series. 2. Use the predicted volatility as an input feature for the LSTM model. 3. Create time sequences for the LSTM model using a sliding window technique. 4. Incorporate GARCH predictions as an additional feature. • LSTM Layer: Include one or more layers with dropout to reduce overfitting. • Dense Layer: Output the positive and negative volatility for the next 10 days. Output • Predicted volatility for the next 10 days. • Use GARCH predictions as a foundational feature and allow LSTM to capture complex and non-linear patterns. Phase 3: Matching with DTW (Dynamic Time Warping) Preparation • Compare the predicted volatility sequence (LSTM + GARCH) with historical volatility. (Ensure the stationarity of the compared data.) • Use a distance metric to identify similar sequences. DTW Process • Calculate the DTW distance between the predicted sequence and historical sequences. • Identify historical sequences with the minimum distance. Output • Return the dates or historical periods with similar volatility patterns. • Visualize the match to validate the output. Phase 4: Model Evaluation Forecasting • Use MSE (Mean Squared Error) to evaluate the model's predictions. Matching • Evaluate the accuracy of DTW matches using a similarity metric based on distance (e.g., Normalized DTW Distance). Validation • Perform Time Series Cross-Validation to ensure the model generalizes well. Adaptive Model Features An auto-adaptive model implementation is encouraged, allowing the system to autonomously select and improve forecasting techniques and the search for similar historical series/patterns. This could include: • Adjusting the temporal length of the series. • Utilizing smaller datasets. • Refining parameters for models like GARCH, etc. The objective is to have software that improves its analysis over time through self-training as more data becomes available. Finally, 7 days after the forecast, verify it against actual volatility to assess the software's performance.