FinDocGPT - AI for Financial Document Analysis & Investment Strategy

An AI-powered platform for financial document analysis and investment strategy development, built for the AkashX.ai challenge at HackNation 2025.

🏆 Selected as one of 15 global teams for the Hack-Nation VC Track 2025
(out of 2,800+ applications and 1,500 hackathon participants).

🎯 Challenge Overview

This project implements a comprehensive 3-stage AI system for financial analysis, directly mapped to the AkashX.ai FinDocGPT challenge:

Stage 1: Insights & Analysis (Document Q&A) ✅

• Process financial documents from FinanceBench dataset
• Natural language Q&A interface with RAG for document analysis
• Sentiment analysis for overall and key sections

Stage 2: Financial Forecasting ✅

• Predict near‑term price movements using historical data from yfinance
• Interactive chart with time window selection, error metric, and forecast vs. history

Stage 3: Investment Strategy & Decision-Making ✅

• Generate actionable Buy/Sell/Hold with confidence and risk
• Combines document sentiment, recent news sentiment, and forecast signal
• Returns transparency details: used documents, per-doc breakdown, component scores

🏗️ Architecture

             ┌───────────────────────────┐
             │       React Frontend      │
             │         (TypeScript)      │
             └───────────────────────────┘
                           ▲
                           │  REST/JSON
                           │
             ┌───────────────────────────┐
             │       FastAPI Backend     │
             │           (Python)        │
             └───────────────────────────┘
              ▲                          ▲
              │ Ingest PDFs              │ Market Data
              │                          │
┌───────────────────────────┐   ┌───────────────────────────┐
│       FinanceBench        │   │          yfinance         │
│       Dataset (PDFs)      │   │      (Yahoo Finance)      │
└───────────────────────────┘   └───────────────────────────┘

🚀 Quick Start

Prerequisites

• Python 3.10+
• Node.js 18+

1) Start Backend

1. Set up environment variables

   • Copy the example file and edit it with your own values:

     cp .env.example .env

   • Open .env and set your keys:
     • NEWS_API_KEY → Enables news sentiment analysis. Get from newsapi.org.
     • OPENAI_API_KEY → Get from platform.openai.com.
     • PORT → (optional) Server port (default: 5001).

2. Create a conda environment (optional)

   conda create -n hack-nation python=3.10 -y
   

3. Run the start script This automatically activates the conda env, installs all dependencies and starts frontend/backend server.

   ./start.sh
   

Backend runs at http://localhost:5001 (FastAPI docs at /docs). App runs at http://localhost:3000.

📚 Key Routes

• / — Dashboard
• /document/:id — Document Analysis (Q&A, PDF, Sentiment)
• /forecasting — Forecasting (RF/Prophet/LSTM)
• /investment-strategy — Investment Strategy (Stage 3)

🧠 APIs (selected)

• GET /api/documents — List available documents (FinanceBench)
• GET /api/documents/{id} — Document metadata
• GET /api/documents/{id}/text — Extracted document text
• GET /api/documents/{id}/sentiment — TextBlob sentiment for document
• GET /api/documents/{id}/pdf — Serve the PDF inline
• GET /api/companies/{company}/news-sentiment — Aggregated recent news sentiment (requires NEWS_API_KEY)
• POST /api/forecast — Body: { ticker, period, horizon, model } → forecasts + MAE
• POST /api/investment-recommendation — Body: { company|ticker, document_ids? } → action/confidence/risk + transparency
• GET /api/lookup?q= — Resolve ticker/company candidates (for UX)

Notes

• News sentiment requires NEWS_API_KEY; otherwise that component is omitted gracefully.
• PDFs are served inline via /api/documents/:id/pdf.

📊 Dataset

The project uses the bundled FinanceBench dataset containing:

• 300+ Financial Documents: 10-K, 10-Q, 8-K filings, earnings reports
• Major Companies: Apple, Microsoft, Amazon, Google, Tesla, and more
• Q&A Examples: Pre-made questions and expert answers
• Multiple Years: 2015-2023 data coverage
• Diverse Sectors: Technology, Healthcare, Finance, Consumer, etc.

Sample Companies

• Technology: Apple, Microsoft, AMD, Intel, Adobe, Oracle
• Consumer: Amazon, Costco, Nike, McDonald's, Coca-Cola
• Healthcare: Johnson & Johnson, Pfizer, CVS Health
• Finance: JPMorgan, American Express, PayPal
• And many more...

🛠️ Features

Stage 1 — Document Q&A

• Document Library: Browse 300+ financial documents
• Smart Search: Filter by company, sector, document type
• Document Viewer: Read full document text
• Q&A Chat Interface: Ask questions about any document
• Financial Metrics Extraction: Automatic revenue, profit analysis via Q&A
  • Note: For newly opened documents, the first query may take longer or fail as the system generates the vector database for retrieval. This is a one-time process per document. Wait about a minute, then try again.
• Sentiment Analysis: Market sentiment from financial text
• Responsive Design: Works on desktop and mobile

Stage 2 — Forecasting

• Historical price ingestion and model selection (LSTM/RF/Prophet)
• MAE surfaced in UI; interactive brush/zoom; dashed forecast line; directional coloring

Stage 3 — Investment Strategy

• Synthesizes document sentiment, news sentiment, and forecast
• Returns action (BUY/SELL/HOLD), confidence, risk level, target price (when applicable)
• Transparency: used_documents and doc_breakdown to show which docs influenced the decision

🔧 Technical Stack

Frontend

• React 18 with TypeScript
• React Router for navigation
• Axios for API communication
• Lucide React for icons
• Custom CSS with modern design system

Backend

• FastAPI web framework
• CORS for cross-origin requests
• PyMuPDF for PDF text extraction
• LangChain for document processing
• NLTK for sentiment analysis

Key Libraries

• pymupdf: PDF text extraction
• langchain: Document chunking and processing
• nltk: Natural language processing
• pdfplumber: Enhanced PDF table extraction

📁 Project Structure

hack-nation-2025/
├── backend/
│   ├── main.py               # FastAPI app entry
│   └── util/                 # PDF, retrieval, sentiment, forecast utils
├── datasets/
│   └── financebench/         # Document metadata and PDFs
├── frontend/                 # React + TypeScript app
│   ├── src/pages/            # Pages: Dashboard, DocumentAnalysis, Forecasting, InvestmentStrategy
│   ├── src/services/api.ts   # API client
│   └── public/
├── requirements.txt          # Python deps
├── start.sh                  # Dev helper to launch backend
└── README.md

🎨 User Interface

Dashboard

• Company Overview: Browse financial documents by company
• Sector Filtering: Filter by industry sector
• Document Types: 10-K, 10-Q, 8-K, Earnings reports
• Search Functionality: Find specific companies or documents

Document Analysis

• Dual-Pane Interface: Chat Q&A + Document viewer
• Real-time Q&A: Ask questions and get AI responses
• Financial Metrics: Automatic extraction of key numbers
• Confidence Scores: AI confidence in responses
• Quick Questions: Pre-built common questions

Demo Guide

1. Document Analysis: Open /document/:id from Dashboard; ask questions; view sentiment and PDF side-by-side.
2. Forecasting: Open /forecasting; enter ticker (e.g., AAPL), choose model and horizon; run forecast; inspect MAE and overlay.
3. Investment Strategy: Open /investment-strategy; enter ticker/company (e.g., Apple or AAPL); review action, confidence, risk; see which documents were used and the sentiment-over-time chart.

🧪 Try These Prompts

Try these sample questions:

Financial Metrics

• "What is Apple's revenue for 2022?"
• "How much debt does Microsoft have?"
• "What are Amazon's key risk factors?"

Business Analysis

• "What is Tesla's main business?"
• "How does Google make money?"
• "What are Netflix's competitive advantages?"

Performance Questions

• "How is McDonald's cash flow?"
• "What drove Apple's growth in 2022?"
• "What are Intel's biggest challenges?"

🔮 Future Enhancements

Stage 1: Financial Document Analysis and Q&A

• Improved Metric Extraction: Use pattern matching and keyword search (reports are partially structured), prompt engineering or a combination of both
• Improved RAG: Employ different kinds of document indices for better content retrieval (e.g., LLM-generated summaries)

Stage 2: Advanced Forecasting

• Machine Learning Models: Use more sophisticated models and include financial metrics from the reports as features
• Risk Modeling: VaR, Monte Carlo simulations
• Uncertainty Estimates on Forecasting: Use an ensemble of different prediction models and model accuracy from historical data

Stage 3: Investment Strategy

• Portfolio Theory: Modern portfolio optimization
• ESG Integration: Environmental, Social, Governance factors
• Backtesting: Historical strategy performance
• Real-time Alerts: Market change notifications

📝 API Documentation

See “APIs (selected)” above and live docs at http://localhost:5001/docs.

Authors

• Lukas Probst
• Christopher von Klitzing

📄 License

This project is developed for the AkashX.ai HackNation 2025 challenge.

---

Sponsored by AkashX.ai | HackNation 2025 Challenge

— Building the future of AI-powered financial analysis —