Stock Price Prediction Using LSTM

The Problem

Stock price movement is notoriously hard to predict from raw historical prices alone — a lot of naive forecasting models effectively just lag the actual price by one time step and call it a prediction. The goal was to build something that goes beyond that: pulling in technical indicators and daily news sentiment as auxiliary signals, then packaging the whole thing as a pipeline that keeps itself current rather than going stale the day after training.

Approach

1. 01Ran a daily cron job (weekdays) that pulled fresh OHLC data via yfinance and computed a full technical-indicator set — SMA, MACD, RSI, Bollinger Bands, Stochastic Oscillator, ATR, and OBV.
2. 02In parallel, pulled the day's news headlines via NewsAPI and scored each one with VADER sentiment analysis.
3. 03Merged the two streams by trading day, rolling weekend and holiday sentiment forward onto the next trading day so no signal was silently dropped.
4. 04Retrained a Bidirectional LSTM with an attention mechanism on 16 features across 120-day sequences, using Huber loss for robustness to price outliers.
5. 05Served predictions two ways: a FastAPI /predict endpoint (containerized with Docker, deployed on AWS EC2) for programmatic access, and a Streamlit dashboard for interactive use.

Visuals

Results

Reflection

The architecture ended up mattering as much as the data — moving from a plain LSTM to a bidirectional model with attention, and handling the weekend/holiday sentiment gap explicitly, both came from hitting real edge cases while building this rather than being obvious upfront. If I revisited this, I'd backtest the model against an actual trading strategy rather than point-accuracy metrics alone, since a lower MAE doesn't always translate into profitable decisions.