Sentimentsdriven portföljoptimering : En utvärdering av Rädsla & Girighet med XGBoost och Black-Littermans Allokeringsmodell

Investor sentiment plays a crucial role in driving sudden market regime shifts. The idea of exiting markets before major downturns and re-entering ahead of rallies is appealing – but can it be achieved through a systematic approach? This study explores the use of Fear and Greed Indices as proxies for investor sentiment, combined with ETF data for four major asset classes: stocks (SPY), bonds (BND), gold (GLD), and cryptocurrencies (HODL.PA). We employ the machine learning algorithm XGBoost to uncover patterns that may predict short-term market movements. These predictions are then used as inputs, or "views," in the Black-Litterman model to optimize portfolio weights. The model was evaluated using different configurations, with both pre fine-tuned and fine-tuned hyperparameters, using training sizes 80% of the data. Our results show that the sentiment-driven portfolio was able to outperform a traditional 60/40 portfolio (60% stocks, 40% bonds) in several scenarios – but only before hyperparameter tuning. Once tuning was applied, the model’s performance declined and consistently underperformed relative to the 60/40 benchmark. These findings suggest that incorporating sentiment through machine learning offers potential, especially in simpler model configurations. However, the results also highlight the risk of overfitting and the challenge of generalizing in non-stationary financial environments. Overall, the approach shows promise, but does not provide a reliable improvement over traditional portfolio strategies.