📈 PortfolioAI — GenAI-Driven Portfolio Optimisation for Indian Markets

BTech Minor Project 2024–25
AI-powered NSE portfolio optimisation combining Black-Litterman model with FinBERT sentiment analysis

---

🎯 Overview

PortfolioAI is a full-stack portfolio optimisation system that helps Indian retail investors intelligently allocate capital across 16 Nifty-50 stocks. It combines:

• Black-Litterman Model — Market-implied equilibrium returns as prior, updated by LLM sentiment views
• FinBERT Sentiment — Pre-trained financial NLP model scores live news headlines per stock
• Macro Regime Overlay — Nifty 50 moving averages + VIX fear index scales equity exposure
• CVaR Optimisation — Minimises worst-case tail losses for conservative profiles
• Streamlit Dashboard — Product-level UI with fresh investment and portfolio rebalancing flows

---

🏗️ Architecture

data_collector.py      ← M1: Downloads NSE prices, Nifty 50, market caps
        ↓
sentiment_engine.py    ← M2: Fetches news (Serper API) → FinBERT scoring
        ↓
feature_builder.py     ← M3: BL prior μ, covariance Σ, LLM views Q, posterior μ_BL
        ↓
optimizer.py           ← M4: EfficientFrontier / CVaR / walk-forward backtest
        ↓
dashboard.py           ← M5: Streamlit product dashboard
macro_overlay.py       ← L3: Nifty regime detector + VIX overlay

---

🚀 Quick Start

1. Clone & set up environment

git clone https://github.com/Jevik-R/portfolio-ai.git
cd portfolio-ai

python3 -m venv venv
source venv/bin/activate          # Windows: venv\Scripts\activate
pip install -r requirements.txt

2. Configure API keys

Create a .env file (see .env.example):

SERPER_API_KEY=your_serper_key      # https://serper.dev — 2500 free queries/month
OPENAI_API_KEY=your_openai_key      # Optional — for GPT rationale & narratives

Or set SERPER_API_KEY directly in sentiment_engine.py line 27.

3. Run the pipeline

# Step 1 — Download 7 years of NSE market data
python data_collector.py

# Step 2 — Fetch live news & score with FinBERT (~5 min, downloads model once)
python sentiment_engine.py

# Step 3 — Launch the dashboard
streamlit run dashboard.py

Open http://localhost:8501 in your browser.

---

📊 Stock Universe — 16 NSE Large-Caps

Sector	Stocks
Technology	TCS, Infosys, Wipro, HCL Technologies
Finance	HDFC Bank, ICICI Bank, SBI, Kotak Mahindra Bank
Healthcare	Sun Pharma, Dr. Reddy's
Consumer/FMCG	Hindustan Unilever, ITC
Energy	Reliance Industries, ONGC
Infrastructure	Larsen & Toubro
Telecom	Bharti Airtel

---

🧠 Technical Methodology

Black-Litterman Model

μ_prior  =  δ × Σ × w_mkt          (market-implied equilibrium returns)
Q_i      =  μ_prior_i
           + α_dynamic × sentiment_score × σ_i × 0.5    (FinBERT tilt)
           + β × earnings_surprise × σ_i                 (earnings signal)
μ_BL     =  BL posterior via Idzorek confidence weighting

FinBERT Sentiment Pipeline

1. Fetch 20 headlines per stock via Serper API (Google News)
2. Score each headline with ProsusAI/finbert → {positive, negative, neutral}
3. Confidence-weighted aggregation with source credibility weights
   • Reuters / Bloomberg: 1.5× | Seeking Alpha: 0.7× | ET / Mint: 1.4×
4. Dynamic alpha: α = 0.10 + (conf − 0.5) × 0.15 − std × 0.20

Risk Profiles

Profile	Objective	Max Weight
Conservative	Min CVaR (5% tail)	20%
Moderate	Min Volatility / Max Sharpe	30%
Aggressive	Max Sharpe	40%

Macro Overlay (Level 3)

Regime scale:  Bull 100% | Neutral 90% | Bear 70%   (Nifty 50 vs MA50/MA200)
VIX scale:     Normal 100% | Elevated 85-92% | High 75% | Extreme 50%
Combined:      max(40%, regime_scale × vix_scale)

---

📁 Project Structure

portfolio-ai/
├── data_collector.py      # M1 — NSE data download
├── sentiment_engine.py    # M2 — FinBERT news sentiment
├── feature_builder.py     # M3 — BL feature engineering
├── optimizer.py           # M4 — Portfolio optimisation + backtest
├── dashboard.py           # M5 — Streamlit dashboard
├── macro_overlay.py       # L3 — Market regime + VIX overlay
├── data/
│   ├── market_caps.csv    # NSE market capitalisations (BL weights)
│   ├── sentiment_scores.csv  # Latest FinBERT scores
│   └── backtest_metrics.csv  # Walk-forward backtest results
├── requirements.txt
└── README.md

---

🖥️ Dashboard Features

💰 Fresh Investment Flow

• Enter ₹ amount → Black-Litterman optimisation → allocation with exact shares
• Donut chart coloured by sector | Horizontal bar chart by sentiment signal
• AI-generated rationale per stock (template or GPT-3.5)

🔄 Portfolio Rebalancer

• Enter current holdings (₹) → BUY / SELL / HOLD plan
• Cash flow summary | Transaction cost estimate (0.1% brokerage)
• Current vs target allocation comparison

📉 Backtest Tab

• Walk-forward backtest: June 2018 → present (bi-weekly rebalancing)
• BL + Sentiment vs Pure Quant vs Nifty 50 cumulative return chart
• Look-ahead bias warning (sentiment is live-only)

🌍 Macro Panel

• Live Nifty 50 regime badge (🟢 Bull / 🟡 Neutral / 🔴 Bear)
• Live VIX fear gauge | Combined equity scale | Cash buffer %
• Sector sentiment heatmap | Analyst consensus scores

---

⚙️ Configuration

Parameter	File	Default	Description
DELTA	feature_builder.py	2.5	Market risk aversion
TAU	feature_builder.py	0.025	BL prior confidence
VIEW_SENSITIVITY	feature_builder.py	0.5	Sentiment → return magnitude
MAX_HEADLINES	sentiment_engine.py	20	News per stock
RISK_FREE_INR	optimizer.py	0.065	Indian G-Sec yield
LOOKBACK	optimizer.py	252	Covariance window (trading days)

---

📦 Dependencies

Package	Purpose
yfinance	NSE price data & market caps
pypfopt	Black-Litterman, EfficientFrontier, CVaR
transformers	FinBERT model (ProsusAI/finbert)
torch	FinBERT inference backend
streamlit	Web dashboard
plotly	Interactive charts
requests	Serper API news fetching

---

⚠️ Disclaimer

This tool is for educational purposes only and does not constitute financial advice.
NSE equity investments are subject to market risk. Brokerage fees, STT, and capital gains tax (LTCG/STCG) apply.
Consult a SEBI-registered investment advisor (RIA) before investing.

---

👨‍💻 Author

Built as a BTech Minor Project
Powered by: FinBERT · Black-Litterman · PyPortfolioOpt · Streamlit