Fin Forecast App

• Front-end (Flask) to select instrument and horizon
• Back-end (SQLite via SQLAlchemy) to store historical data and predictions
• Forecasting (Python) with ARIMA (traditional), Moving Average baseline, and LSTM (neural)
• Visualization (Plotly): OHLC candlesticks + forecast overlay
• Unit tests for critical utilities
• Reproducibility via requirements.txt

This implementation follows the assignment spec (front-end, back-end, ML, candlesticks, tests).

Quickstart

How it works

Architecture (high-level)

[Flask UI] --(form: ticker, horizon, model)--> [Controller]
    \-> [Data Loader (yfinance)] --(OHLC)-> [SQLite: prices]
    \-> [Models: ARIMA | MA | LSTM] --(forecast)-> [SQLite: forecasts]
    \-> [Plotly] candlestick(OHLC) + overlay(forecast) -> HTML

Database

• prices (ticker, datetime, open, high, low, close, volume, interval)
• forecasts (ticker, created_at, horizon_steps, model_name, yhat_json, interval, train_start, train_end, metrics_json)

Models

• ARIMA (statsmodels) on Close prices; automatic (p,d,q) simple sweep
• MA baseline: rolling mean used for naive forecast
• LSTM (PyTorch): small seq2seq-style iterative one-step predictor trained on Close

Notes

• If yfinance fetch fails, app will fall back to a bundled sample CSV for AAPL (1h data) to keep demo working.
• Forecast horizons supported: 1h, 3h, 24h, 72h.
• Intervals are chosen based on horizon: 15m for <= 24h, else 1h.

Tests

Screenshots

• When you run the app and submit a ticker+horizon, you'll see a candlestick chart with the forecast line overlaid and a metrics table (RMSE/MAE/MAPE).

Academic Integrity

• All libraries used are open-source.
• No proprietary LLMs or paid APIs are used.