Skip to content

README.md

Latest commit

 

History

History
128 lines (91 loc) · 4.64 KB

README.md

File metadata and controls

128 lines (91 loc) · 4.64 KB

AlphaPredictor

Screenshot (473)

AlphaPredictor is a stock analytics web app built with Streamlit. It downloads 5 years of historical price data, engineers return-based features, trains three machine learning models, and predicts the next-day closing price — all with an interactive dashboard.

Features

  • Fetches 5 years of historical market data via yfinance
  • Trains three independent models:
    • Linear Regression — fast baseline
    • Random Forest Regressor — ensemble tree model
    • XGBoost Regressor — gradient-boosted trees
  • Predicts next-day closing price as a 3-model ensemble average
  • Confidence score derived from how closely the three models agree
  • Bullish / Bearish trend signal (▲ / ▼)
  • Model Comparison table — MAE, RMSE, and MAPE on a held-out test set
  • Actual vs Predicted chart — visual comparison of all three models on the test period
  • Historical price chart with 20-day and 50-day moving averages (Plotly)
  • Modular codebase, clean error handling, and type-annotated throughout

Project Structure

AlphaPredictor/
├── app.py                 # Streamlit entry point
├── modules/
│   ├── data_loader.py     # Fetch & clean historical stock data (yfinance)
│   ├── model.py           # Feature engineering + LR / RF / XGBoost training
│   └── visualization.py   # Historical price chart (Plotly)
├── data/                  # Optional folder for cached CSVs
├── requirements.txt
└── README.md

Installation

python -m venv .venv
# Windows
.venv\Scripts\activate
# macOS / Linux
source .venv/bin/activate

pip install -r requirements.txt

Run the App

streamlit run app.py

Open the URL shown in the terminal (usually http://localhost:8501).

How It Works

  1. Enter a ticker such as AAPL, TSLA, or MSFT and click Predict →.
  2. The app downloads 5 years of daily OHLCV data.
  3. Feature engineering converts raw prices to stationary daily returns, then builds:
    • Three lag features (lag_1, lag_2, lag_3)
    • Two rolling-average features (ma_5, ma_10)
  4. Train / test split — oldest 80% for training, newest 20% for held-out evaluation (no shuffle — time-series integrity is preserved).
  5. Each model is evaluated on the test window → MAE, RMSE, MAPE reported.
  6. All three models are retrained on 100% of data before predicting tomorrow, so the most recent market behaviour is captured.
  7. Tomorrow's predicted return is converted back to a price: price = latest_close × (1 + predicted_return).
  8. The ensemble prediction is the simple average of all three model prices.
  9. Confidence is estimated from the standard deviation of the three predicted returns — lower spread = higher confidence.

Output Explained

Field Description
Latest Close Most recent closing price
Predicted Next Close Ensemble (average of LR + RF + XGB) price forecast
Expected Change % difference between ensemble prediction and latest close
Confidence 0–100 score; higher = more agreement between models
Signal ▲ Bullish if prediction > latest close, ▼ Bearish otherwise

Model Comparison Table

After prediction, the app shows a per-model breakdown on the held-out test set:

Model MAE RMSE MAPE (%)
Linear Regression — — —
Random Forest — — —
XGBoost — — —

Design Decisions

  • Returns instead of raw prices — keeps the target stationary; Random Forest and XGBoost can extrapolate correctly even at all-time highs.
  • Two-phase training — eval phase on the test split gives honest error metrics; final phase retrains on all data before forecasting.
  • No data leakage — the latest row (for tomorrow's forecast) is captured before dropna() removes it from the training set.
  • MultiIndex compatibility — newer yfinance versions return a MultiIndex column structure; columns are flattened automatically after download.

Ideas for Extension

  • Add RSI, MACD, Bollinger Bands as features
  • Experiment with LSTM or Transformer-based models
  • Add ticker autocomplete / search
  • Cache downloaded data with st.cache_data to speed up repeated runs
  • Deploy to Streamlit Community Cloud

Disclaimer

This project is for educational purposes only and is not financial advice. Stock markets are inherently unpredictable; never make investment decisions based solely on model output.