Stock Projections

Creating Machine Learning Models to Predict Future Stock Prices, Using Alpha Vantage API

Stock Prices:

Supply of Stock: shares outstanding, or total amount of shares held by investors issued by a company
Value directly linked to market capitalization (total valuation of a company, or (shares outstanding * current share price))
Markets run by humans, thus irrational emotion/bias also factors into price
Market Capitalization of a publicly traded firm directly linked to stock price
High amount of buyers compared to sellers -> price increase
High amount of sellers and few buyers -> price decreases
Primary Markets: buying at IPO
Secondary Markets: Investors selling stocks to one another (aka stock market), prices change rapidly
Earnings Calls, Annual Meetings, M&A (suspected or finalized), Media Coverage, Price Trends/Patterns, and Inflation all major contributors

Deep Learning Framework
Recurrent Neural Network
Supervised Machine Learning
- use labeled datasets to train algorithms to classify data/predict outcomes

generate training and validation data sets by normalizing via a Z-Score and splitting the data over a predetermined window_size
training data: section of data to train LSTM problem on
validation data: section of separate data to test LSTM's accuracy
z-score: measurement of data point's relationship to the mean -> normalizer

via PyTorch and requires Dataset object
Automatic batching through parallelization of the data loading process, leveraging GPU during training.
Boosts speed and saves memory in the program.

Input Modulation Gate = linear_1: maps input values into high dimensional feature space
- transforming features for LSTM layer so non-linear and zero-mean
Input Gate = lstm: learns data in sequence
Output Gate = linear_2: produce predicted value based on LSTM's output
Forget Gate = Dropout: randomly selected neurons ignored during training
- prevents overfitting
Model Training:
- Loss Function: Mean Square Error of difference between predicted and actual values
- Bakcpropagation: backtracking to improve predictions
- Adam Optimizer: update parameters based on learning rate
- StepLR Scheduler: reduce learning rate during training
Model Evaluation:
- Predict on training data and plot predicted vs. actual data for validation dataset
- Zoomed-In LSTM Model