Impact of natural disasters on U.S property liability insurers' and indices' stock price volatility

Abstract

Purpose – The purpose of this thesis is to assess if natural disasters impact the volatility of 19 property-liability insurers in the United States of America (USA) and 3 stock indices over a 10-year period using GARCH (1,1), IGARCH (1,1), EGARCH (1,1) and GJR-GARCH (1,1). Additionally, we implement the Value at Risk (VaR) and Extreme Value Theory (EVT) method to generate the worst loss over a target horizon that will not be exceeded with a given level of confidence. In this regard, this thesis will be a pioneer in examining the performance of capital markets in a context of unusually high uncertainty and build upon existing mixed-viewed literature with regards to capital market behavior in such conditions. Design/methodology/approach – The daily closing prices for each property-liability insurer and stock index are collected over a sample period from January 1st 2010 till December 31st 2020. The sampled period is segregated into two sub-sample periods: the in-sample period extending from January 1st 2010 till December 31st 2017, and the out-of-sample period extending from January 1st 2018 till December 31st 2020. Accordingly, in-sample returns are then calculated from daily closing prices and utilized to estimate the parameters of the selected volatility models, based on the constraints and assumptions of each model. Subsequently, the calculated in-sample parameters are implemented to forecast the volatilities for both periods (in-sample and out-of-sample). Next, the three chosen error metrics RMSE, MAE and MAPE are used to identify the optimal model for each stock during both the in-sample and out-of-sample periods. Next, a dummy variable was employed to measure the impact of natural disasters on the chosen stocks and indices. Afterwards, through the use of the Volatility Update Historical Simulation method, future return scenarios are generated for the Dow Jones U.S Property and Casualty Insurance Index (DJUSIP) on daily basis over the period December 6th 2018 till December 31st 2020. The Value at Risk (VaR) is then calculated for 250 days at four confidence levels (90%, 95%, 97.5% and 99% confidence levels). Eventually, in order to determine the accuracy of the underlying VaR model the Kupeic test is performed. Lastly, we incorporate Extreme Value Theory (EVT) into our calculations as it assumes a separate distribution for extreme losses in order to estimate the probability of extreme values. Findings – Results showed that the IGARCH (1,1) has proven to be the optimal model for the majority of the chosen insurance companies during the in-sample period. On the other hand, the EGARCH (1,1) model performed best for a substantial number of insurance companies, particularly, AFG, UFCS.O, GBLI.O, HALL.O and the chosen stock indices (SPX, IXIC and DJI). As for the remaining stocks, PGR, MSADY.PK, CINF.OQ, WRB, WTM, HMN and HCI the GARCH (1,1) and GJR-GARCH (1,1) proved to outperform other models. The same calculations applied for the in-sample period are applied to the out-sample period. The results reflect homogeneity among the indices, SPX, IXIC, DJI and two insurance companies, AFG and UFCS.O. In addition, the EGARCH model was also the most accurate model for RLI, SIGI.O, ARGO.K, UVE, DGICA.Oand FNHC.O but only for the out-of-sample period. Among all of the chosen stocks, IGARCH out-performed other models for SAFT.O for both in-sample and outX sample periods. Alternatively, IGARCH performed best for the out-sample period of HMN and GBLI.O whereby. With regards to the remaining stocks, the GARCH (1,1) model proved to be the best performing model for both in-sample and out-sample period for PGR, WRB, WTM and HCI. Specifically, for the out-sample period, the GARCH model out-performed other models for CB, MSADY.PK and CINF.OQ while the IGARCH (1,1) and GJR-GARCH were chosen for the in-sample period, respectively. Lastly, the GJR-GARCH is the optimal model for HALL.O for the out-sample period. After determining the optimal model for the in-sample and out-sample periods, we set the pre-disaster period to 0 and to 1 for both the- one month and three-month post disaster periods. The outcome highlighted that that during the in-sample period, volatility is more likely to be negatively impacted by natural disasters and during the out-sample period, the majority of stocks’ volatility are positively impacted by natural disasters. Furthermore, using the Rolling Window procedure and by incorporating the optimal model into the Volatility-Weighted Historical Simulation method, the Value at Risk (VaR) was estimated for 250 days between 06/06/2018 till 03/06/2019 at 90%, 95%, 97.5% and 99% confidence levels for Dow Jones Property & Casualty Insurance Index (DJUSIP:DJI). The computed VaR results were then compared to actual returns in order to determine the number of days/exceptions in which actual returns exceeded VaR estimates across the 250 days period. Lastly, the Kupiec Test was performed and the outcome reflected that VaR provides a very accurate measure in determining the level of downside risk at all confidence intervals. Lastly, we incorporate Extreme Value Theory (EVT) into our calculations at 95% and 99% confidence interval, the VaR was estimated to be 2.33% and 7.79%. Based on the VaR, the Expected Shortfall (ES) was estimated to be 6.79% and 7.80%, respectively at 95% and 99% confidence level. When comparing the VaR obtained through EVT and the volatility adjusted model, we note that the volatility adjusted model yielded a higher VaR thus, we can conclude that the model is overestimating the loss. Research limitations/implications – This thesis has potential limitations. A particular limitation is that with the majority of research in this area the analysis on the impact of natural disasters has been made in segregation from other effects, such as macroeconomic, political and calendar announcements. While this simplifies research, it is tricky as natural disasters may be vulnerable to contamination caused by macroeconomic announcements independent of the disaster or catastrophe itself. For example, Shelor et al. (1992) analysis of the 1989 Loma Prieta earthquake compromised the outcome of the research as it failed to take into consideration the lowering of official US interest rates two days later. Moreover, multiple property-liability insurance companies were excluded from the dataset as there was no sufficient data for the chosen timeframe (01/01/2010 till 31/12/2020) andthere are many property-liability insurance companies that are private and thus, these could not be included. Therefore, the dataset used could have been wider and more inclusive. Originality/value – The findings of this thesis investigated the behavior of the 19 U.S insurance stocks, 3 U.S indices, a property-liability composite index for Value at Risk (VaR) and Extreme Value Theory (EVT) and 252 natural disasters, over the period extending from 01/01/2010 till 31/12/2020, which makes its dataset comprehensive, exhaustive and novel to those of preceding research.