Welcome to the Optiver Trading at the Close. This project focuses on forecasting the 60-second weighted average price (WAP) change against the 60-second synthetic stock index change, utilizing advanced machine learning and deep learning techniques.
- Overview
- Data Analysis
- Feature Engineering
- Modeling and Training
- Model Evaluation
- Deep Learning Model
- Data Preprocessing
- API Testing
The project involves intricate data analysis, preprocessing, feature engineering, and the deployment of sophisticated models like XGBoost and CNN-BiLSTM-ATN for time series prediction in the financial domain.
The initial phase focuses on exploratory data analysis (EDA) to understand the intricacies of the dataset, which comprises over 5 million entries and 17 columns, including time-related data, order book metrics, and market metrics.
In-depth feature engineering was conducted to enhance the model's predictive power. This includes:
- Imbalance Features: Calculating imbalance direction, book imbalance, and price imbalance.
- Order Book Features: Including bid-ask spread, mid-price movement, normalized spread, and size ratio.
- Order Flow and Volume Features: Tracking volume changes and daily volatility.
- Pricing Features: Focusing on reference to WAP ratio, overnight returns, and more.
- Technical Indicators: Applying MACD, RSI, RoC13, Bollinger Bands, and others.
- Transformation Techniques: Utilizing linear interpolation, linear regression, and various transformations for missing value imputation and feature optimization.
- Parameters: Learning rate, max depth, min child weight, colsample by tree, subsample.
- Bayesian Hyperparameter Tuning for enhanced performance.
- Model Evaluation using Mean Absolute Error (MAE).
- Architecture: Combines convolutional layers, bidirectional LSTM layers, and attention mechanisms.
- Training: Utilizes Adam optimizer with a learning rate of 0.001 and TPU execution for efficiency.
- Layers: Includes Conv1D, BatchNormalization, Dropout, Bidirectional LSTM, and Dense layers.
- Hyperparameters: Batch size dynamically adjusted for TPU execution, dropout rates for regularization.
- Specifics: The model was trained with 82 input dimensions and 54 time steps. LSTM units were set to 32. The model was compiled with a mean squared error loss function and evaluated using MAE.
- The XGBoost model showed significant performance, optimized through Bayesian Hyperparameter Tuning, achieving a Mean Absolute Error (MAE) of 2.46999 on the test set and about 3.0 during training, indicating its effectiveness in time series forecasting.
- Architecture: Integrates convolutional layers for feature extraction, bidirectional LSTM layers for capturing temporal dependencies, and attention mechanisms for focusing on relevant time steps.
- Training Environment: Utilizes Tensor Processing Units (TPUs) for efficient and faster training processes.
- Model Serialization: The trained models are saved for future use, ensuring reproducibility and deployment readiness.
- Data scaling and transformation using MinMaxScaler.
- Outlier detection and handling.
- Lag features and rolling window statistics for capturing temporal trends.
- Data partitioning into training, validation, and test sets.
During the API testing phase, a mock loop was established to handle incoming test data. The test data was processed and prepared for prediction using either the Neural or XGBoost models, depending on the Modelling flag.
- Neural Model Processing: Test data is scaled using a saved MinMaxScaler model. Predictions are generated using the CNN-BiLSTM-ATN model, with reshaped data to match the model's input shape.
- XGBoost Model Processing: Features are prepared, and predictions are made using the XGBoost model.
- Dynamic Data Update: The training dataset is dynamically updated with new predictions and sorted for consistency.
This approach ensures that the model stays up-to-date with the latest market trends and adapts its predictions accordingly.