← back

MarketPulse AI

research infrastructure

A stock and crypto prediction engine combining k-NN, Linear Regression, and LSTM neural networks with VADER sentiment analysis. Built as a modular system with a clean separation between data layer, model engine, and interface, plus a walk-forward backtesting framework. Started as a university semester project, now also serving as the research infrastructure behind an upcoming paper comparing Prophet vs. LSTM+Prophet hybrid models for financial time-series forecasting.

Educational/research project — predictions are not financial advice.

Models

• k-NN (naive + enhanced) — classifies next-day direction from return patterns; enhanced adds volume, RSI, volatility, MACD
• Linear Regression (naive + enhanced) — predicts next-day return, direction from sign, confidence via sigmoid
• LSTM — recurrent network for sequential patterns, with quick/standard/cluster training presets and early stopping
• Sentiment — all models predict from price first, then VADER sentiment shifts the probability post-hoc

Backtesting

Walk-forward testing with simulated P/L, configurable trading fees and stop-loss, a buy-and-hold benchmark, and risk metrics (max drawdown, Sharpe, Sortino, yearly rolling performance).

Web GUI

A FastAPI backend with a React 19 + TypeScript + Vite frontend. Dashboard, Predict, and Settings pages are functional; Backtest, Training, and a News vs. No-News Analysis page (built for the paper) are in progress.

Testing & CI

103 pytest tests plus a 13-test dependency-free smoke test. GitHub Actions runs ruff (lint), mypy (typecheck), and pytest with Codecov coverage on a Python 3.12/3.13 matrix, backed by pre-commit hooks locally.