jf's repositories
Pruning-DeepNeuralNetwork
Eliminated parameters with two pruning methods (weight and units) sparsity (k%) with minimal loss in performance.
Optimal-RegressionModel-HyperParameters-Flask-Azure-Docker
Seek the optimal regression model while optimizing model's hyperparameters using Tree-Structured Parzen Estimator Approach (TPE) that applies the Bayes Rule and evaluating performance with RMSE.
PortfolioVarianceBW
Collected Bridgewater Associate's 13F SEC Filing for 2021Q1 top 25 holdings based on percentage weight. Calculated factors covariance, factor exposures, idiosyncratic variances, rebalanced weights for Portfolio Variance.
DNNHyperparameterTuning
Optimize DNN hyperparameters using TPE-Bayesian optimization based on kernel fitting with 0.9625 ROC-AUC. The model uses metric values through search process probabilistic to converge to the optimal combination of hyperparameters.
BERT-NLP-Deployment-FastAPI-Docker
FastAPI BERT deployment, Logistic Regression, and Multinomial NaiveBayes to predict NLP sentiment analysis with an Evaluation Accuracy of 94.15%, 89.31%, 84.77, respectively.
RandomForest-Optimal-HyperParameter
RandomForest Model to classify binary target values (pos&neg) return with calculated features (RSI, MA, MACD, etc). Tuning hyperparameters with Bayesian Optimization, GridSearchCV, and RandomSearchCV.
BGRU-BLSTM-GloVe-fastText-NLP
Compared model's performance using Bidirectional GRU and LSTM with GloVe and fastText pre-trained word embeddings on sentiment dataset. BGRU-GloVe Embeddings outperformed with an accuracy of 90.99%.
KMeansDbscanPCA
Two Unsupervised Learning clustering models (KMeans & DBSCAN) and PCA for reduction dimensionality. Applied techniques to find the optimal hyperparameters and visualized the outputs.
EDA-FeatureEngineering-FeatureSelection
Different ways of Data Collection, Exploratory Data Analysis (EDA), Feature Engineering, and Feature Importance on 4 datasets. Seek patterns by visualization, test hypothesis, check assumptions, etc
DeepNeuralNetworks
Predicted a binary classification problem with an Artificial Neural Network and outputted 0.8733 ROC-AUC. Implemented scheduler to reduce learning rate, used criterion as a Binary Cross-Entropy and Adam Optimizer
XGBoost-OptimalHyperParameters-AWS-SageMaker
XGBoost Classifier to predict binary target for vehicle insurance with a 0.9085 ROC-AUC by optimizing hyperparameters with Bayesian Optimization. Reduced overfitting by implementing Stratified K-Fold.
LogisticRegressionPyTorch
Predict binary values using a Log-Reg with PyTorch and from scratch. Used Optuna to find the optimal model (SVM, Decision-Tree, Log-Reg) and seek optimal hyper-parameters.
KNearestNeighbors
K-Nearest Neighbors from scratch with different Distance Metrics and an Accuracy: 96.67%. Distance Metrics [Euclidean, Minkowski, Manhattan, Hamming] to find the optimal accuracy.
MonteCarloPortfolioOptimization
Optimal Portfolio by assigning random weights to each SPDR Sector ETF with Returns 44.52%, Risk: 0.20, Sharpe Ratio 2.84. Iterating n-times to increase the accuracy for portfolio optimization and evaluating portfolio's Sharpe-Ratio.
Multi-LinearRegressionModel
Find the relationship between multiple independent variables to a dependent variable (GDP Growth %). Validated assumptions and tested hypothesis Variance Inflation Factor, Breusch–Pagan, Ljung-Box & Anderson-Darling Test
CointegratedPairsTrading
Verify if BTC and ETH pair of crypto are cointegrated. Calculating the hedge ratio by fitting a linear regression and check if the spread is stationary using Augmented Dickey-Fuller Test