Project 07 / 10 · Forecast & ML
Amazon Stock Forecasting
Multi-model forecasting dashboard on AWS EC2 — Seq2Seq LSTM cut MAPE from 17.45% to 8.86% with technical-indicator fusion and backtested signals.
01 / Context
The problem
Instead of treating stock prediction as a visually convincing curve-fit, this project tests whether deep learning or classical models can beat simple baselines under chronological, leakage-resistant evaluation.
02 / System workflow
From input to output
- 01yfinance OHLCV
- 0211 engineered signals
- 03Chronological split
- 04Four-model benchmark
- 05Risk dashboard
03 / Architecture
Inside the system
Daily market data feeds technical indicators including RSI, MACD, Bollinger width, moving-average ratios, volume ratio, and rolling volatility.
Linear Regression, one-step LSTM, Seq2Seq LSTM, and XGBoost are compared with rolling walk-forward validation across AMZN, MSFT, and GOOGL.
MLflow tracks 18 runs while Streamlit combines forecasts with SHAP, risk statistics, anomaly detection, backtests, sentiment, and peer comparisons.
04 / Engineering decisions
What shaped the build
Decision 01
Leakage-resistant evaluation
Every split is chronological, scalers fit only on training data, and results are scored against naive or majority baselines.
Decision 02
Predict the actionable target
Returns and direction are evaluated separately from absolute price, which is easily flattered by carry-forward behavior.
Decision 03
Document failed experiments
Bidirectional LSTM, FinBERT features, autoregressive rollout, and larger cloud tuning runs are reported with their failure reasons.
05 / Evidence
Measured, not implied
56.7%
5-day direction
High-confidence XGBoost predictions with 69% coverage.
+3.7pp
Baseline edge
Five-day high-confidence result over majority class.
7,300+
Trading days
Historical OHLCV records used in the current benchmark.
06 / Delivery
What shipped
- 01GitHub Actions refreshes the market CSV after trading hours, with a local fallback for API rate limits.
- 02Streamlit caching keeps LSTM inference, SHAP, walk-forward validation, and risk calculations responsive.
- 03The current application is deployed on Streamlit Cloud with resource-aware XGBoost settings.
Constraint documented
No model beats the naive previous-price baseline on MAPE. The useful result is the modest directional edge and the disciplined evidence showing where the models do and do not generalize.