StockPredictionEnsemble

Stock prediction algorithms that feature deep learning and ensemble techniques. Data consisted of stock prices,economic search-terms of google, technical indicators from feature engineering, stock sentiment, stock fundamentals, and alternative economic data. The ensemble algorithm features preprocessing of data using matrix profiles, fractional differentiation, vector time-embeddings, regime-shift indices, and outlier detection. Additional techniques such as meta-labeling and threshold-based predictions improved the f1 scores above 70% on a multitude of different assets. The meta model consisted of a bayesian LSTM(Pytorch-based), and the ensemble comprised of tree classifers, SVM, KNN, neural networks, gradient boosters, logistic regression, and FB Prophet. To reduce computational burden, feature importance analysis and dimensionality reduction(PCA,KPCA) were employed. Finally, portfolio allocation was calculated via the Black-Litterman model provided weights from the ensemble. Of course, hyperparameter tuning was performed using optuna for an optimal version for use. Data is sourced from quandl, yahoo finance,pandas data-reader, google trends and finbox.