Artificial Intelligence Driven Portfolio Recommendation System: Leveraging Sentiment Analysis Towards Improved Performance

By leveraging sophisticated algorithms with the right inputs, academia already demonstrated representative accuracies and returns in predictive models for financial markets. With the adherence to internet and media sharing platforms increasing, investors gain faster access to public information, leading to widespread discussions on social media that can actively impact market conditions. This work explores the integration of sentiment analysis from news articles and social media into portfolio recommendation systems. By combining both sentiment analysis with established technical indicators, this approach aims to enhance the financial performance of portfolio recommendations, reflecting market sentiment and trends. For each of the top 100 companies selected, the algorithm selects the best-performing technical indicators, sentiment types, and machine learning models. The results demonstrate a high selection rate across all technical indicators employed, and a pattern of predominantly choosing Support Vector Machine and Extreme Gradient Boosting models, jointly accounting for 89.29% of selections. Including sentiment data significantly enhances model performance, with models incorporating any form of sentiment being chosen in 72.45% of the test samples. Sentiment from news articles alone accounted for 30.06%, social media sentiment alone accounted for 20.68%, and the combined inclusion of both sentiment sources for 21.71%. The portfolios constructed by the system surpassed benchmark returns, with the top-performing strategy achieving an annualized mean return of 29.3%, compared to 19.65% from the Standard & Poor's 500 benchmark and 17.01% from an equally weighted benchmark of the 100 stocks, and an annualized Sharpe Ratio of 164.46, compared to 138.29 and 140.15 from both benchmarks.