Welcome

• nettey-boevi-gilles-b.koumou.1@ulaval.ca   

• nettey.boevi.gilles.koumou@usherbrooke.ca

Publications

• Weight bound constraints in mean-variance models: a robust control theory foundation via machine learning  Quantitative Finance. 2024

• Hierarchical Risk Budgeting Journal of Financial Data Science. 2024

• Unifying Portfolio Diversification Measures Using Rao’s Quadratic Entropy (with Benoît Carmichael et Kevin Moran).  Journal of Quantitative Economics. 2023

• Risk budgeting using a generalized diversity index. Journal of Asset Management. 2023

• Coherent Diversification Measures in Portfolio Theory : An Axiomatic Foundation (with Georges Dionne) Risks. 2022

• Diversification and Portfolio Theory : A Review.  Financial Markets and Portfolio Management. 2020

• Mean-Variance Model And Investors’ Diversification Attitude : A Theoretical Revisit.  Finance Research Letters. 2020

• Rao’s Quadratic Entropy and Maximum Diversification Indexation (with Benoît Carmichael and Kevin Moran). Quantitative Finance. 2018

Working Papers

1. The RQE-CAPM: New insights about the pricing of idiosyncratic risk (with Benoît Carmichael et Kevin Moran)

Abstract: We use an equivalent form of Markowitz's mean-variance utility function, based on Rao's Quadratic Entropy (RQE), to update the standard capital asset pricing model (CAPM), both in the presence and in the absence of a risk-free asset. The resulting new equilibrium, which we denote RQE-CAPM, offers important new insights about the pricing of risk. Notably, it reveals that the reason for which the standard CAPM does not price idiosyncratic risk is not only because the market portfolio is law of large numbers diversified but also because the model implicitly assumes agents' total risk aversion and their correlation diversification risk preference balance each other exactly. We then demonstrate that idiosyncratic risk is priced in a general RQE-CAPM where agents' total risk aversion and their correlation diversification risk preference coefficients are not necessary equal. Our general RQE-CAPM therefore offers a unifying way of thinking about the pricing of idiosyncratic risk, including cases where such risk is negatively priced, and is relevant for the literature assessing the idiosyncratic risk puzzle. It also provides a natural theoretical underpinning for the empirical tests of the CAPM or the pricing of idiosyncratic risk performed in some existence studies.

2. Large-Scale Mean-Variance Optimization and Chunking Algorithm

Abstract: We propose a new, highly effective and easy-to-implement algorithm for solving large-scale mean-variance optimization problems --- with weight upper bound constraints and without short sales --- when the size of mean-variance portfolios is much smaller than the number of assets, which is almost always the case. Our novel algorithm is built on the novel representation of mean-variance models in terms of the support vector data description --- an unsupervised machine learning algorithm designed for a one-class classification problem --- and the chunking algorithm, a decomposition algorithm for support vector machine. 

3.  Enhancing ESG News Annotation: Leveraging GPT for the analysis of ESG News and Events (with Keven Bluteau and Frank Coggins)

Abstract: This paper studies whether a general-purpose large language model (GPT-3.5) can recover standard ESG concepts from raw news text and whether the resulting signals are priced. Using Wall Street Journal articles from 2000 to 2022 for S&P 500 firms, we build a firm-level database of ESG media events and, in a zero-shot setting, prompt GPT-3.5 to classify each item by relevance, novelty, materiality, sentiment, and severity. Materiality is assessed both directly, from the full news narrative, and indirectly, via ESG topics mapped to SASB-style key issues. We validate GPT-based labels against manual annotations, ProQuest metadata, and interpretable text models. Embedding these labels in standard asset-pricing tests, we show that novel, severe, and material negative ESG events are associated with economically large and often persistent stock-price declines, whereas nonmaterial, follow-up, and most positive news have little or only transitory impact. Direct, full-text materiality explains return patterns more consistently than the indirect, topicbased approach. Finally, GPT-based controversy measures are only weakly correlated with MSCI pillar scores, indicating that LLM-derived news signals capture an eventdriven dimension of ESG risk that complements traditional ESG ratings.

4. ESG Risk Beyond Aggregate Scores: Forecasting Severe ESG Events with Granular Data and Machine Learning (with Keven Bluteau and Frank Coggins)

Abstract: Environmental, Social, and Governance (ESG) risks are central to investment risk management, yet many investors still rely on aggregate E, S, and G scores. We develop a machine-learning framework that forecasts an S&P 500 firm’s three-year probability of a new severe ESG event, using severity measures derived from ESG-related Wall Street Journal news article. Predictors include aggregate and granular ESG information from three major providers, firm characteristics, and a media-coverage index. We compare logistic benchmarks based on sector indicators, media coverage, and aggregate pillar scores

with extreme gradient boosting (XGB) and other machine-learning models that leverage the full predictor set. At the pillar level, once media coverage is included, pillar scores add little predictive power; XGB delivers the largest gains for E events, smaller gains for S events, and no gains for G events. At the key-issue level and for higher severity thresholds, XGB yields substantial improvements relative to the benchmarks. Feature-importance results indicate that these gains are driven by economically interpretable signals, including emissions and resource-use measures, labor and product-safety indicators, ownership and compensation structures, and proxies for liquidity and the information environment. Overall, these results suggest that ESG event risk is too heterogeneous to be summarized by pillar scores, and granular machin learning models better predict events most likely to become economically material.