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

---

Predicting Next Day's Market Price

via: Long Short-Term Memory (LSTM)

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

---

Download data using Alpha Vantage API and plot original data

Data Preparation: Training vs. Validation

• 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

DataLoader

• 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.

Long Short-Term Memory (LSTM) deep learning framework:

• 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

Predicting Next Day's Close Price

• using Tensor via PyTorch for unseen data
• tensor: matrix of data -> turning unseen data into a matrix
• actual future price:

References:

• Tokenist: How are Stock Prices Determined