Stock Forecast & Backtest

XGBoost stock forecasting with walk-forward backtesting. Built to measure model performance honestly, since a common failure point in finance ML projects is data leakage that inflates backtest results.

What it does

• Fetches 1-5 years of OHLCV data via yfinance
• Engineers 34 features (lag returns, rolling MAs, RSI, MACD, volatility, volume ratio)
• Trains XGBoost with walk-forward temporal splitting (no random k-fold, no lookahead)
• Reports MAPE, RMSE (return basis), and directional accuracy vs 50% random baseline
• Serves predictions through a FastAPI layer with in-memory model caching (~85ms p50)
• Stores forecast history per user in MongoDB (indexed for O(log n) history queries)
• Auth via Google OAuth

Leakage fix

Standard k-fold assigns rows randomly to folds, so a row from day 300 can appear in training while day 250 is in validation. This is invalid for time series data.

Walk-forward helps, but there's a subtler issue: the last horizon rows of each training fold have targets (close[t+horizon]/close[t] - 1) computed from prices that fall inside the test window. The fix is to end training at train_end - horizon, so no training label touches a test-period price.

With this fix applied, directional accuracy dropped from ~64% to ~57% and RMSE increased slightly. The difference was leakage.

Stack

Layer	Tech
Frontend	Vue 3, Tailwind CSS, Chart.js
Backend	Node.js, Express, Passport.js
ML service	FastAPI, XGBoost, yfinance, pandas
Database	MongoDB (Mongoose)
Auth	Google OAuth 2.0

Running locally

1. ML service

cd ml-service
pip install -r requirements.txt
uvicorn main:app --reload --port 8000

2. Backend

cd backend
npm install
cp .env.example .env   # fill in MongoDB URI + Google OAuth credentials
npm run dev

3. Frontend

cd frontend
npm install
npm run dev            # http://localhost:5173

Smoke test (no auth needed):

curl -X POST http://localhost:8000/train \
  -H "Content-Type: application/json" \
  -d '{"ticker":"AAPL","start_date":"2020-01-01","end_date":"2025-01-01","horizon":5}'

Metrics

All metrics are computed on returns, not prices. Price-based RMSE conflates model skill with price scale (AAPL at $190 vs a $10 stock). Return-based RMSE is scale-invariant and cross-ticker comparable.

Directional accuracy is shown with a +Xpp vs random label relative to the 50% random baseline. The number is smaller after fixing leakage, which is expected.

Limitations

• No transaction cost simulation; directional accuracy alone does not imply profitability
• In-memory model cache resets on service restart; no persistence layer for trained models
• Regime changes (e.g. 2020 crash) can invalidate a model trained on bull-market data
• For production you'd move to PostgreSQL and add model versioning