How lldelduca/btc_predictions_denoisingautoencoders works
Project Type and Purpose
• Type: Data-science/machine-learning pipeline contained in a Jupyter notebook.
• Purpose:
– Train and compare denoising autoencoders (DAE) versus standard autoencoders on MNIST for feature extraction and noise reduction.
– Use those extracted features (plus raw financial indicators) to build and evaluate Bitcoin price predictors (linear regression, SVR, LSTM).
– Benchmark against a simple moving average and run statistical tests on residuals.
File Structure
• Only file: Full_Notebook.ipynb
– Sections inside notebook correspond to distinct functional stages: Data Acquisition, Preprocessing, Model Definitions, Training/Validation, Prediction, Evaluation, Visualization.
README Insights
• Data sources: MNIST dataset, custom web scrapers, TA-Lib, Yahoo Finance YQL API.
• Models: Stacked Autoencoder (SAE), Stacked Denoising Autoencoder (SDAE), Linear Regression, SVR, LSTM.
• Hyperparameter tuning: Optuna.
• Evaluation metrics: MAE, RMSE, MAPE; statistical tests (Durbin-Watson, White’s, Jarque-Bera).
• Visualization: plots of reconstructions, residuals, kernel densities, prediction vs. benchmark.
Main Components and Their Interactions
a. Data Acquisition
– MNIST loader (PyTorch)
– Financial data scrapers (TA-Lib, Yahoo API)
b. Data Preprocessing & Feature Engineering
– Train/validation/test splits
– Noise injection (masking) for DAE
– Feature scaling, lagged price creation
c. Encoding Stage
– SAE training (encoder/decoder)
– SDAE training (with noise)
– Output: reduced-dimension feature vectors
d. Feature Selection & Combination
– Compare SAE, SDAE, and raw feature sets
– Select robust features based on reconstruction performance
e. Predictive Modeling
– Linear Regression (with statistical validation)
– Support Vector Regression
– LSTM network
– Hyperparameter tuning via Optuna
f. Benchmarking & Evaluation
– Simple Moving Average baseline
– Compute MAE, RMSE, MAPE
– Conduct Durbin-Watson, White’s, Jarque-Bera tests
g. Visualization & Reporting
– Reconstruction examples (original/noisy/reconstructed)
– Prediction vs. actual price plots
– Residual and kernel density plots
Architectural Patterns & Design Principles
• Pipeline pattern: clear separation between data ingestion, preprocessing, feature extraction, modeling, evaluation, and visualization.
• Modularity inside notebook: each section implements one stage.
• Reuse of encoders for two purposes (feature reduction + denoising).
• Experimentation loop: swap feature sets and model types.
Technologies & Frameworks
– Jupyter Notebook – orchestration layer
– PyTorch – SAE & SDAE
– Scikit-Learn – regression, SVR, preprocessing
– Keras or PyTorch – LSTM (depending on implementation)
– Optuna – hyperparameter tuning
– TA-Lib, yfinance – feature scraping
– pandas, NumPy, matplotlib/seaborn – data manipulation & plotting