Painting the black box white: a fundamentalist-based trading strategy using interpretable trees

• Autor(a) principal: Possatto, André Bina
• Data de Publicação: 2020
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Institucional do FGV (FGV Repositório Digital)
• Texto Completo: https://hdl.handle.net/10438/29461
• topic: Estratégias long-short; Machine Learning; Stock returns prediction; Model interpretation; Long-short trading strategy; Aprendizado de máquina; Previsão de retorno de ações; Interpretação de modelos; Economia; Aprendizado do computador; Ações (Finanças) - Preços - Previsão; Investimentos - Análise; Mercado financeiro - Brasil

Difficulty understanding how a black box model makes predictions has undermined machine learning’s success in financial markets, according to a recent article from Bloomberg (2019b). Our work shows how model-agnostic methods to interpret machine learning predictions turn these models more transparent to a human investor. We benchmark three tree-based algorithms between themselves, creating long-short investment strategies with independent models for each leg and using only fundamentalist analysis. We then apply the models to the Brazilian stock market (Bovespa) and achieve an out-of-sample expected annual return of 26.4% with a Sharpe ratio of 0.50. Ensembles between the long and short legs improve this result for a Sharpe ratio of up to 1.26, comparable to other works in the literature reported by Avramov et al. (2019) when considering real-world constraints. Our strategy has low asset turnover and transaction costs do not explain the results. All models achieve positive risk premiums and two are statistically significant. Interpretation shows differences on the key predictors for over- and underperformance, with the first focusing on price-to-value and the second on size and liquidity. Local interpretation is discussed in the case of Magazine Luiza, showing how model explanation helps an investor to understand and decide which stocks to buy or sell based on the models’ output. We argue that different performance and interpretation between long and short models and the possibility of ensembling are key advantages of modeling these positions separately.