GNN + LLM Portfolio Optimization | M.Sc. Thesis; GraphSAGE, FinBERT, Mistral-7B, PyTorch
Built a PyTorch pipeline combining GraphSAGE stock-relationship embeddings with FinBERT news sentiment, alongside a separate LLM-based scorer (Mistral-7B) that ranked stocks from prompted news summaries. Designed and ran backtests comparing four strategies Markowitz baseline, GNN-only, GNN + FinBERT, and the LLM direct scorer and found news-coverage quality, not model architecture, was the main constraint on the LLM scorer's returns.