Hugging Face's logo

Spaces:
Brajmovech
/
IRIS-AI_DEMO
Running

App Files Community
IRIS-AI_DEMO / README.md
Brajmovech's picture
Brajmovech
Add API key rotation for NewsAPI and Webz
f7ce805
preview code
|
raw
history blame contribute delete
7.46 kB
metadata 
title: IRIS-AI DEMO
emoji: 📈
colorFrom: blue
colorTo: indigo
sdk: docker
app_file: app.py
app_port: 7860
pinned: false

IRIS: Vision in Volatility

CP3405 Team Project

Project Description

IRIS is an AI-driven decision support tool designed to filter market noise. We identify short-term divergences between market sentiment and price movement to flag risk and validate sentiment before a trade is made.

Key Features:

  • Sentiment Analysis: Synthesizing news and social media sentiment.
  • Risk Detection: Flagging when sentiment and price action diverge.
  • Focus Market: US Tech Sector (e.g. AAPL, TSLA, and AMD).

Our Goal

Our vision is not to build a "Magic Crystal Ball" that gives Buy/Sell advice.

Instead, our goal is to build a decision support tool that helps retail traders manage volatility logically. We aim to act as a "Check Engine Light," alerting users to irrational market behavior so they can pause and reassess their risk.

The Scrum Team (Sprint 1&2)

Role Name
Product Owner Mayur Ramesh
Scrum Master Minzhi Liu
Data Scientist Erdong Xiao
AI Engineer Shen Chen
Engineer IC (FE/BE) Jun Bin Yap
UI/UX Designer Jinyu Xie

Disclaimer

IRIS provides decision support flags based on historical data and sentiment analysis. It does not provide financial advice.

Team Runtime Data (Conflict-Free)

To avoid report/cache merge conflicts across different developer machines:

  • By default, runtime outputs now go to runtime_data/ (or demo_guests_data/ when DEMO_MODE=true).
  • Tracked data/ files are no longer the default write target.
  • You can override with environment variables:
    • IRIS_USE_REPO_DATA=true to intentionally write into data/
    • IRIS_DATA_DIR=<path> to use a custom machine-local folder

This keeps generated reports, sessions, feedback logs, and yfinance caches machine-local by default.

FinBERT Setup

FinBERT sentiment analysis is enabled by default for IRIS.

  • Dependencies are already declared in requirements.txt:
    • torch
    • transformers
  • On first startup, IRIS loads the ProsusAI/finbert model using the normal Hugging Face cache for the machine.
  • Optional environment overrides:
    • IRIS_ENABLE_FINBERT=false to disable FinBERT explicitly
    • IRIS_FINBERT_MODEL_ID=<hugging-face-model-id> to swap the model
    • IRIS_FINBERT_CACHE_DIR=<path> to pin the model cache to a project-local folder such as runtime_data/huggingface/transformers

If FinBERT cannot load, IRIS now logs the exact reason and continues running with sentiment fallback behavior instead of failing app startup.

News API and Webz.io Key Fallback

IRIS now supports multiple API keys for NewsAPI and Webz.io headline collection failover.

  • NewsAPI:
    • NEWS_API_KEY=your_primary_key
    • NEWS_API_KEYS=key_one,key_two,key_three
  • Webz.io:
    • WEBZ_API_KEY=your_primary_key
    • WEBZ_API_KEYS=key_one,key_two,key_three
  • If both plural and legacy single-key vars are set, IRIS uses the plural list first and then appends the legacy single key if it is different.
  • When NewsAPI reports a rate/quota exhaustion error, IRIS retries the same request with the next configured NewsAPI key.
  • When Webz reports a rate-limit, quota, or token-limit style error, IRIS retries the same request with the next configured Webz key.

Historical Reports and Trend Accuracy

Use these commands to keep the pipeline reproducible across machines.

  • Generate fresh runtime reports/session summary:
    • python run_daily.py --once --tickers AAPL,MSFT
  • Read latest session summary from active runtime dir:
    • runtime_data/sessions/latest_session_summary.json (default)
    • demo_guests_data/sessions/latest_session_summary.json (DEMO_MODE=true)
  • Generate trend-accuracy chart using active runtime reports:
    • python trend_accuracy_comparison_chart.py
  • Compare using historical archived reports in tracked data/:
    • python trend_accuracy_comparison_chart.py --data-dir data --output runtime_data/charts/<YYYY-MM-DD>/trend_accuracy_from_repo_data.png

Notes:

  • Accuracy requires at least two market sessions per ticker in the selected dataset.
  • The script auto-falls back to data/LLM reports for LLM baseline files when runtime dirs do not contain them.

Current Work Progress

  • Live Demo (Hugging Face Space):
  • Latest Updates (March 26, 2026):
    • Finalized headlines panel UX fixes: restored stable scroll behavior, tightened spacing/density, and added dynamic height sync so the right-column headlines card always aligns with left-column dashboard cards.
    • Added frontend height re-sync triggers after dashboard render, async LLM insights render, and browser resize for consistent layout across ticker/data states.
    • Synced frontend fixes to both local and server projects with cache-version updates for immediate deployment refresh.
  • Backend & REST API (app.py):
    • Set up a Flask server with CORS support.
    • Implemented core API endpoints (/api/analyze, /api/chart, /api/session-summary/latest) to securely interface the frontend dashboard with the backend AI engines.
    • LLM Integration: Aggregates predictions from ChatGPT 5.2, DeepSeek V3, and Gemini V3 Pro to provide comparative "LLM Insights".
    • Pipeline Efficiency: Removed redundant news-processing work in /api/llm-predict to reduce repeated backend load.
    • Reliability: Hardened LLM prediction handling for unavailable models and failure edge cases.
  • AI & Risk Engine (iris_mvp.py):
    • Market Data: Integrated yfinance to pull historical OHLCV data.
    • Trend Prediction: Built a "Crystal Ball" feature using scikit-learn's Ridge regression to predict next-session prices and identify Uptrends/Downtrends.
    • Sentiment Analysis: Integrated the transformers library using the ProsusAI/finbert model to analyze the sentiment of recent market headlines.
    • Strict News Filtering: Implemented Webz.io and NewsAPI aggregation with strict Regex boundaries and noise reduction patterns to ensure high-quality financial context.
    • Coverage Expansion: Added Google News RSS ingestion and candidate scoring to improve headline coverage quality.
  • Automated Schedulers (run_daily.py & generate_llm_reports.py):
    • Developed timezone-aware automation scripts to trigger bulk ticker analyses exactly at the US market open (09:00 ET).
    • Implemented automated LLM forecasting scripts to build baseline reference data.
  • Interactive Frontend Dashboard (index.html & app.js):
    • Built a sleek, responsive UI to query tickers dynamically.
    • Integrated TradingView Lightweight Charts featuring real-time interactive Candlestick and Volume Histogram overlays.
    • Added dynamic UI widgets for the "Check Engine Light" (Risk Indicator), Market Prediction summaries, LLM Insights, and a Sentiment Meter with recent headline rendering.
    • Added LLM timeout/retry UX flow and smoother loading-state handling for model cards.
    • Improved ticker input validation to support special market symbols.