Essays on Machine Learning for Risk Analysis in Finance, Insurance and Energy

Abstract

[eng] This thesis provides research catalogued in the area of risk assessment. Specifically, it contributes to the fields of international finance and asset pricing in finance, and risk assessment in energy economics and transportation research. We present in this thesis a generalization of the spillover indexes to analyze interconnectedness at firm level, and define the aggregate influence from a sector and a country on a firm. We also discuss which factors are relevant for predicting conditional quantiles across the distribution of returns and present a method for selecting factors based on the investor interests. We study the performance of quantile regression against quantile time-series models. Finally, we present a regression framework which estimates VaR and CTE ensuring noncrossing conditions for various quantile levels, and discuss results on energy and telematics data. Within the financial contagion literature, we aim to provide a better understanding of international spillovers and a method for visualize which country and sector are its main drivers. We show that not all companies are driven by their own country or sector, which should be considered by investors and risk managers when assessing company risk and managing investments. In this paper we show that a large percentage of firms’ stocks are driven by their country. But contrary to the belief where country is the main driver of a company’s return movements, a part depends mainly on its sector. We note that 1) the financial services and energy companies are positioned at the center of the network, and 2) northern and western Europe are highly interconnected, while eastern and southern Europe present lower spillovers. 3) For the British energy firms British Petroleum (BP) and Royal Dutch Shell, we evidence greater spillovers from France than from Great Britain itself. 4) We identify which non-Russian firms are most influenced by Russia, simulating a risk management analysis in the event of of turmoil distresses such as the recent Ukrainian conflict. 5) We show the improvement on spillover information when using individual firm connectedness and aggregating spillovers afterwards against calculating spillovers directly from indexes. 6) We finally show that eastern Europe has increased interconnectedness with the rest of the continent after the Covid-19 pandemic. Regarding the asset pricing literature, we aim to understand the key elements that predict extreme quantile levels of a stock return. We study which factors for a 7-factor asset pricing specification are more relevant for each part of the distributions’ tail. The 7-factor specification is constituted by the factors size, book-to-market, operating profitability, investment, momentum, market beta and liquidity. We present a method to add more factors depending on the investors’ interests. We use quantile regression models for predicting quantile levels 0.05, 0.25, 0.5, 0.75 and 0.95 of the stock returns using cross-sectional characteristics as covariates from the Open Source Cross-Sectional Asset Pricing Dataset (Chen and Zimmermann, 2021). We observe that the factor size changes from positive to negative sign when predicting lower quantiles to higher quantiles. We show that extreme quantile level estimations perform better than the median in terms of pseudo-R2. Regarding factor significance, the variable investment has lower predictive power than other factors in terms of t-statistics for all tested quantile levels. Liquidity gains significance if quantile levels increase. For book-to-market, profitability, momentum and market beta, median predictions of returns are more significant than extreme quantile level estimations. The opposite happens for size, which presents higher relevance for predicting extreme quantile levels of the returns’ distribution. We observe that during crisis periods, some factors lose significance. This is the case of profitability and momentum for quantile levels 0.05 and 0.5, and size, book-to-market and market beta for quantile level 0.5. We add additional factors individually and compare the weighted average pseudo-R2 obtained across all 5 quantile levels. The weighting depends on the strategy that the investor follows. For all strategies tested, the most relevant factors to add to the 7-factor specification are momentum seasonality and net operating assets. Following, for strategies more interested in predicting losers’ tails (left part of the distribution), adding asset growth is recommended, but if the investor is interested in the winners’ tail (right part of the distribution), the recommended factor to add is enterprise multiple. Within the asset pricing literature, we encourage the use of cross-sectional information against time-series factors to predict extreme quantile levels of the right-hand side of the response distribution during periods of high volatility. By using this methodology, we do not restrict the information on panel-like datasets, which allows us to study more companies, and provide estimates for newly added firms. We use quantile regression specification with cross-sectional characteristics obtained from the Open Source Cross-Sectional Asset Pricing Dataset (Chen and Zimmermann, 2021) and compare results against a CAViaR (Engle and Manganelli, 2004) specification. Fama and French (2020) evidence that the average returns are better explained by using cross-sectional factors than by using time-series factors. We show that this only applies on extreme quantile levels during high volatility periods. We show that individual firm Hits (exceedances above VaR) calculated using time-series models tend to accumulate, while using cross-sectional data we avoid concentrations. We show that cross-sectional information improves the prediction of Value-at-Risk (VaR) and Conditional Tail Expectations (CTE). We finally discuss changes on capital requirements for a firm. In general, by using cross-sectional information, capital requirements should be increased from when time-series information is used. During turmoil periods the opposite happens: capital requirements should decrease compared to when using the CAViaR specification. Inside the area of non-crossing quantiles, we define the non-crossing property for VaR and CTE for several quantile levels. We define a regression framework based on neural networks that creates an environment for predicting VaR and CTE for several quantile levels while asserting non-crossing conditions. The proposed neural network predicts VaR and CTE as positive excesses of the previous VaR and CTE. We prove that this definition satisfies the non-crossing property and show its improvement against the Monotone Composite Quantile Regression Neural Network (Cannon, 2018) and a quantile regression and CTE linear approach on an energy consumption and telematic datasets. We show the estimation improvements on extreme quantile levels of the right part of the distribution against the other tested models by using Murphy diagrams (Ehm et al., 2016). We present examples with crossing predictions to demonstrate the infeasibility of such results in a business context, which we overcome using the proposed model.