| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 1.54 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
A interpretabilidade de um modelo pode ser definida como o grau de compreensão sobre o
funcionamento interno de um modelo, permitindo perceber as causas sobre o qual recai o resultado. Na
modelação estatística, pode verificar-se um trade-off entre a interpretabilidade e a precisão da previsão:
modelos de interpretabilidade natural revelam uma menor precisão relativamente a modelos mais
complexos. A previsão do Produto Interno Bruto é um aspeto fundamental para a aplicação de políticas
económicas, com particular atenção para políticas monetárias e políticas orçamentais. Nesse sentido e
tendo em conta que os modelos de machine learning preveem melhor a volatilidade económica, tem se
vindo a assistir a uma revolução do método de previsão macroeconómico utilizado pelas principais
organizações políticas nacionais e internacionais.
O objetivo deste trabalho prende-se em investigar se um black-box, Extreme Gradient Boosting
(XGBoost) pode superar os métodos de interpretabilidade natural selecionados, Regressão linear e
Árvore de decisão na previsão do Produto Interno Bruto (PIB) per capita para dados de painel e
identificar medidas de importância das variáveis para melhorar a transparência dos modelos de machine
learning. Se as organizações políticas e financeiras forem capazes de prever e interpretar corretamente
os fatores, podem implementar políticas mais eficazes.
Para a análise de interpretabilidade de XGBoost, utilizam-se os seguintes métodos, Shapley Additive
ExPlanations (SHAP), Permutation Feature Importance e Partial Dependence Plot. Através destes
métodos pretendemos mostrar que os resultados obtidos por XGBoost podem ser interpretados sem um
grande esforço computacional e garantir maior vantagem competitiva. Contudo, como será aprofundado
durante este trabalho, verifica-se que o modelo XGBoostserá o modelo com melhor precisão de previsão
para os dados de painel.
The interpretability of a model can be defined as the degree of understanding of the inner workings of a model, allowing one to understand the causes upon which the result is based. In statistical modelling, there can be a trade-off between interpretability and forecast accuracy: models with natural interpretability show lower accuracy than more complex models. Gross Domestic Product forecasting is a fundamental aspect for the implementation of economic policies, with particular attention to monetary policies and fiscal policies. In this sense and considering that machine learning models better predict economic volatility, there has been a revolution in the macroeconomic forecasting method used by the main national and international political organizations. The objective of this work is to investigate whether a black-box, Extreme Gradient Boosting (XGBoost) can outperform selected natural interpretability, Linear regression and Decision tree methods in forecasting Gross Domestic Product (GDP) per capita for panel data and to identify measures of variable importance to improve the transparency of machine learning models. If political and financial organizations can correctly predict and interpret factors, they can implement more effective policies. For the XGBoost interpretability analysis, the following methods are used, Shapley Additive ExPlanations (SHAP), Permutation Feature Importance and Partial Dependence Plot. Through these methods we intend to show that the results obtained by XGBoost can be interpreted without a large computational effort and ensure greater competitive advantage. However, as will be deepened during this work, it is verified that the fixed effects model will be the model with the best prediction accuracy for panel data.
The interpretability of a model can be defined as the degree of understanding of the inner workings of a model, allowing one to understand the causes upon which the result is based. In statistical modelling, there can be a trade-off between interpretability and forecast accuracy: models with natural interpretability show lower accuracy than more complex models. Gross Domestic Product forecasting is a fundamental aspect for the implementation of economic policies, with particular attention to monetary policies and fiscal policies. In this sense and considering that machine learning models better predict economic volatility, there has been a revolution in the macroeconomic forecasting method used by the main national and international political organizations. The objective of this work is to investigate whether a black-box, Extreme Gradient Boosting (XGBoost) can outperform selected natural interpretability, Linear regression and Decision tree methods in forecasting Gross Domestic Product (GDP) per capita for panel data and to identify measures of variable importance to improve the transparency of machine learning models. If political and financial organizations can correctly predict and interpret factors, they can implement more effective policies. For the XGBoost interpretability analysis, the following methods are used, Shapley Additive ExPlanations (SHAP), Permutation Feature Importance and Partial Dependence Plot. Through these methods we intend to show that the results obtained by XGBoost can be interpreted without a large computational effort and ensure greater competitive advantage. However, as will be deepened during this work, it is verified that the fixed effects model will be the model with the best prediction accuracy for panel data.
Description
Mestrado Bolonha em Métodos Quantitativos para a Decisão Económica e Empresarial
Keywords
Produto Interno Bruto per capita Dados de painel Time-fixed effects XGBoost Interpretabilidade de modelos Gross Domestic Product per capita Panel data Model interpretability
Pedagogical Context
Citation
Rebelo, Maria Adriana Carmo (2022). “Interpretabilidade de modelos Machine Learning para a previsão macroeconómica”. Dissertação de Mestrado. Universidade de Lisboa. Instituto Superior de Economia e Gestão
Publisher
Instituto Superior de Economia e Gestão