Nifty50 Portfolio Optimizer

AI-powered Decision Support System (DSS) for constructing optimized Nifty 50 portfolios using LSTM price forecasts and sentiment signals (FinBERT + momentum fallback). Provides an interactive Streamlit dashboard to explore forecasts, optimized allocations and investment plans.

---

Key features

• Per-ticker LSTM forecasting using multivariate inputs (price, volume, technicals, sentiment).
• Sentiment pipeline with FinBERT (news) and a momentum-based fallback when news/FinBERT unavailable.
• Covariance matrix calculation and constraint-aware portfolio optimization (Sharpe ratio maximization).
• Interactive Streamlit UI: refresh data, set constraints (num stocks, target return, max risk), visualize allocations and per-stock charts.
• Export investment plan as CSV.

---

Architecture (high level)

• Data ingestion: yfinance (historical OHLCV) and optional news.
• Feature engine: returns, volatility, moving averages, RSI and sentiment alignment.
• Modeling: per-ticker LSTM models + scalers, multi-day rolling forecasts.
• Risk engine: historical returns → covariance matrix.
• Optimizer: Monte‑Carlo / constrained search → best Sharpe ratio portfolio.
• Presentation: Streamlit app reads serialized artifact (portfolio_model_data.pkl) and displays results.

Simple diagram (paste into mermaid.live to render)

flowchart TD
  Data["Price & News Data (yfinance)"]
  ETL["ETL / Preprocessing"]
  Sentiment["Sentiment (FinBERT ⇄ Momentum)"]
  LSTM["LSTM Trainer & Predictor"]
  Risk["Covariance Matrix Calculation"]
  Optim["Optimizer (Sharpe Max)"]
  UI["Streamlit Dashboard"]
  
  Data --> ETL --> Sentiment --> LSTM --> Risk --> Optim --> UI
  UI -->|user control| ETL

Loading

---

Quick start

1. Clone repository

   git clone https://github.com/your-org/nifty50-portfolio-optimizer.git
   cd nifty50-portfolio-optimizer

2. Create Python environment (recommended)

   python -m venv .venv
   source .venv/bin/activate   # macOS / Linux
   .venv\Scripts\activate      # Windows
   pip install -r requirements.txt

   Important packages: streamlit, pandas, numpy, yfinance, plotly, scikit-learn, tensorflow, transformers, torch (optional if FinBERT used).

3. Prepare data / initial artifact

   • If portfolio_model_data.pkl is not present, either:
     • Run the provided notebook / pipeline to generate it, or
     • Launch the app and click "Refresh Market Data" (note: full refresh trains LSTMs and can be compute intensive)

4. Run the app

   streamlit run app.py

---

Usage notes

• "Refresh Market Data" runs a full ETL → sentiment → train → predict → optimize pipeline. This is CPU/GPU and time intensive.
• FinBERT loading requires network access and PyTorch; the app falls back to technical/momentum-based sentiment if unavailable.
• Models and scalers are saved inside portfolio_model_data.pkl. For production, replace with a model registry (MLflow / S3).

---

Files & important modules

• app.py — Streamlit UI & orchestration
• portfolio_functions.py — optimization utilities and portfolio metric calculations
• portfolio_model_data.pkl — serialized artifact (created by refresh or notebook)
• notebook/ — optional training / data-prep notebooks (if present)
• README.md — this document

---

Evaluation & experimentation

• Evaluate LSTM forecasts with MSE/MAE on holdout windows.
• Backtest optimized portfolios vs benchmark (Nifty 50) for realized returns, Sharpe ratio, max drawdown.
• Ablation experiments: FinBERT vs momentum-only vs no-sentiment features.

---

Development notes

• For faster iteration, reduce number of tickers or training epochs.
• To disable FinBERT entirely, run refresh_data with the FinBERT-loading block commented or ensure HuggingFace snapshot is unavailable.
• Consider decoupling heavy tasks into background jobs (Airflow / Kubernetes) for production.

---

License & authors

Authors: Aritra Sarkar, Satish Prem Anand, Dhiranjit Daimary
License: MIT

---

Contact

For issues or contributions, open an issue or PR on the project repository.