단기 NVDA 예측을 위한 AI 기반 Quant 대시보드

A full-stack quantitative analysis and real-time prediction tool for NVIDIA ($NVDA) stock. It combines neural networks for time-series forecasting with Natural Language Processing (NLP) for sentiment analysis, delivering insights through a reactive, custom Cyberpunk UI. The main goal is to build an end-to-end AI pipeline that doesn't solely rely on historical prices (technical analysis) but actively integrates the news factor (fundamental/sentiment analysis). The system provides an automated, short-term estimation of the stock's price direction (UP/DOWN) for the next active market hour. The FastAPI backend acts as the data orchestrator: - Market Data (Prices): Fetched in real-time via the yfinance library, pulling hourly candlesticks (Open, High, Low, Close, Volume). - Sentiment Data (News): Calls News APIs to fetch the latest headlines regarding NVIDIA and the broader AI/Semiconductor sector. - Data Merging: The two data streams are merged using strict chronological sorting. Missing sentiment values (NaN) are forward-filled to ensure every stock timestamp has a corresponding sentiment score. The core prediction engine is powered by a Long Short-Term Memory (LSTM) neural network, implemented in PyTorch. - Feature Engineering: The model processes 2 primary features: the Close Price and a Sentiment Score. The sentiment score is generated by passing news headlines through FinBERT (an NLP model specifically fine-tuned on financial text). - Scaling: Before inference, raw data undergoes MinMax Scaling using parameters saved during the training phase ( scaler_params.npy ). - LSTM Mechanism: The network ingests a rolling "window" (Sequence Length) of the last 10 hours ( SEQ_LENGTH=10 ). It analyzes the sequential pattern of these 10 data points and outputs a scaled prediction, which is then inversely transformed back to a real USD value. - Prediction Horizon: The model strictly forecasts 1 hour ahead (Next-Hour Forecast). It focuses on micro-trends rather than long-term investing. - Market State Awareness: The backend is fully aware of NASDAQ trading hours. If the market is closed (Weekends, After-Hours, Pre-Market), the system accurately calculates and displays the next valid market open hour (e.g., Monday 09:30 AM EST) instead of blindly predicting the next chronological hour. - AI / Machine Learning: PyTorch, LSTM, FinBERT (HuggingFace) - Backend: Python, FastAPI, Pandas, yfinance, Uvicorn - Frontend: Vue.js 3, Vite, Chart.js, TailwindCSS - Architecture: RESTful API, Reactive UI, State Management 1. Install requirements pip install -r requirements.txt - Create a .env file in thenvda_predictor directory and add yourNEWS_API_KEY :NEWS_API_KEY=your_api_key_here - Then start the backend server: uvicorn main:app --reload