Predicting Stock Price Direction

Overview

This repository contains my solution to a StrataScratch challenge, where I predict the daily price direction of Amazon.com, Inc. (AMZN) stock. The solution, implemented in classification.ipynb, demonstrates a binary classification approach to determine whether the closing price will be higher or lower than the opening price for the next trading day.

Although the question and solution focus on AMZN, the methodology can be generalized to other tickers by pulling stock data via the yfinance API in Python.

Dataset

The dataset consists of historical stock price data from 1997 to 2020, which has been split into:

• Training Data: 1997–2016 (AMZN_train.csv)
• Validation Data: 2016–2018 (AMZN_val.csv)
• Testing Data: 2018–2020 (AMZN_test.csv)

Each dataset has the following columns:

Column	Description
Date	Trading date (YYYY-MM-DD format)
Open	Stock price at market open
High	Highest price of the day
Low	Lowest price of the day
Close	Stock price at market close
Adj Close	Adjusted closing price after corporate actions
Volume	Number of shares traded

Objective

The objective is to train and evaluate a predictive model that takes historical stock data as input and outputs a binary prediction:

• 1 (Up) → Closing price is higher than the opening price.
• 0 (Down) → Closing price is lower than the opening price.

We focus on classification metrics rather than predicting exact stock prices, as directional accuracy is more relevant for trading decisions.

Approach

1. Data Preprocessing

   • Load data from CSV files
   • Handle missing values and data formatting
   • Feature engineering (e.g., moving averages, volatility measures)

2. Exploratory Data Analysis (EDA)

   • Visualizing stock trends and patterns
   • Checking data distributions and correlations

3. Model Selection & Training

   • Implementing machine learning models (e.g., Logistic Regression, Random Forest, XGBoost)
   • Tuning hyperparameters for optimal performance
   • Evaluating models using AUC (Area Under Curve), accuracy, and other classification metrics

4. Evaluation & Testing

   • Assessing model performance on validation and test data
   • Interpreting feature importance

Key Insights

• The stock market is highly volatile, making short-term predictions challenging.
• Rather than predicting exact prices, binary classification models help identify trends.
• A model achieving AUC > 0.515 is considered sufficient for this challenge.

Practical Considerations

• No external data sources were used beyond the provided dataset.
• The focus is on code structure, interpretability, and reproducibility rather than maximizing model accuracy.
• This project was designed to be completed within 3 hours, balancing efficiency and depth of analysis.

Running the Notebook

Clone this repository and open classification.ipynb in Jupyter Notebook

Final Findings

The final model, an SVC (Support Vector Classifier) with a linear kernel (C=100), was trained using a StandardScaler preprocessing step to normalize the feature data. The model was trained on the combined train-test dataset and evaluated on the validation set.

Performance Metrics

• Accuracy: 80.9%
• AUC (Area Under ROC Curve): 0.91

Key Insights

• The model achieved a high accuracy of 80.9%, indicating strong predictive performance in classifying whether the next day's closing price will be higher or lower than the opening price.
• The AUC score of 0.91 suggests that the model is very effective at distinguishing between upward and downward price movements.
• Using SVM with a linear kernel proved to be a robust approach for this classification problem, capturing relevant patterns in stock price data.
• Feature scaling via StandardScaler played a crucial role in optimizing model performance.

Limitations & Considerations

• The model is only trained on historical price data and does not incorporate external factors such as news sentiment, macroeconomic indicators, or fundamental data.
• While the accuracy is high, stock market conditions change dynamically, meaning model performance may vary in real-world trading scenarios.
• Further hyperparameter tuning or feature engineering (e.g., incorporating technical indicators) could improve model robustness and generalization.

Next Steps

• Experimenting with ensemble models (e.g., Random Forest, XGBoost) to compare performance.
• Exploring deep learning models (e.g., LSTMs, GRUs) for time-series forecasting.
• Adding technical indicators (RSI, MACD, Bollinger Bands) to enhance feature set.

Contact

If you have any questions or would like further details about this project or my experience, feel free to reach out:

• Email: jordan.c.l.wright@gmail.com