rahulbandekar-hub/ghana-pension-portfolio-ai

# Ghana Pension Portfolio AI A team capstone project (MIT Sloan): a deep-learning portfolio optimizer for Ghana pension funds, built for and with a Ghanaian investment fund partner. Combines structured market data with LLM-generated news sentiment, trains a Temporal Fusion Transformer to forecast asset returns, and produces monthly rebalancing weights that satisfy Ghana's NPRA (National Pension Regulatory Authority) allocation limits. **Note on data:** this repo contains the modeling code only. The underlying structured market data, news corpus, and benchmark fund performance data were provided by the project's industry partner under a data-use agreement and are not included or reproducible from this repo. Company names have been genericized in the writeup and code accordingly. ## The problem Ghana's pension funds are legally capped in how they can allocate capital — a hard ceiling on equities, mutual funds/ETFs, alternatives, and a floor pushed toward government fixed income. Within those constraints, fund managers still need to decide *which* assets to overweight each month. Our project asked: does adding LLM-derived news sentiment to a standard structured-data forecasting pipeline actually improve on that decision, or is it noise? ## Pipeline ```mermaid flowchart LR A[Ghanaian financial news] -->|FinBERT| B[Daily sentiment scores, 35 assets] C[Structured market data equities, FI yields, macro] --> D[Merge + scale] B --> D D --> E[Temporal Fusion Transformer] E --> F[Monthly forecasts] F --> G[NPRA-constrained weight optimizer] G --> H[Backtested vs LR / LSTM / equal-weight / benchmark fund] ``` **Step 1 — Sentiment generation:** FinBERT (`ProsusAI/finbert`) scores Ghanaian financial news articles daily, mapped to 35 tradable assets by category (banking, energy, consumer goods, etc.), producing a daily sentiment score per asset from 2021–2025. **Step 2 — Forecasting:** structured market data (equity prices, T-bill/bond yields, macro indicators) is merged w …