Criscent Birungi

About me

I am a doctoral candidate in Mathematics and Statistics, specializing in Mathematical Finance. Currently, I am working on developing and implementing deep learning methods and machine learning-based recommender systems for investments and optimal annuitization—more broadly, portfolio management with AI. My research focuses on mathematical finance, computational finance, statistical machine learning, statistical theory, econometrics, and time series analysis. I am also working on developing pricing and hedging strategies using deep learning for financial derivatives in the commodities and energy markets. More broadly, I develop machine learning tools to tackle challenges in finance, insurance, statistics, economics, and climate change

Outputs

Recommending the right securities at the right time: A global system based on optimized decision-tree algorithms for dynamic factor solutions

Identifying influential investment variables is a complex process, especially when considering the numerous global companies involved. This paper presents a dynamic stock recommendation system trained on a diverse dataset of fundamental and technical factors. We develop and implement deep learning algorithms to extract the most important variables for return prediction. Due to computational constraints, we focused on evaluating machine learning models that strike a balance between performance and efficiency, ultimately selecting two powerful gradient boosting models: LightGBM and XGBoost. Both LightGBM and XGBoost proved effective for cross-sectional stock ranking, with XGBoost demonstrating superior profitability. To optimize portfolio construction, we implemented a long/short trading strategy designed to identify a sub-universe of securities best suited to the nature of the strategy we want to develop. This strategy involved taking long positions in the fifth quintile, which represents the top-performing stocks ranked monthly, and short positions in the first quintile, reflecting the bottom-performing stocks expected to underperform. This strategy was further evaluated by analyzing conditional performance and return convexity across diverse market conditions. Results demonstrate that our machine learning models are effective for cross-sectional stock ranking, with technical factors dominating in the short term. In contrast, fundamental factors provide more stable signals over time. The XGBoost-based long-short strategy significantly outperforms all other strategies, achieving notable metrics including 18.50% cumulative returns, 22.10% average annual returns, and a 2.10 Risk-Return Ratio. A key finding is its superior risk management, as evidenced by a significantly lower maximum drawdown of -14.15% and substantial diversification benefits. The observed return convexity suggests the strategy maintains resilience in both favourable and adverse market conditions. By combining fundamental and technical factors, our recommender system consistently delivers strong forecasting accuracy and profitability. It effectively uncovers new factors and combinations suited to evolving market environments, offering considerable diversification benefits despite high monthly turnover (around 52% to 58%). Understanding the strategy’s behavior under different conditions enables investors to make more informed predictions aligned with their desired risk and return profiles.