In statistics, the Gumbel distribution is a key tool for modeling extreme values across diverse fields such as finance, hydrology, and engineering. However, existing tests for verifying the goodness-of-fit of the Gumbel distribution, such as the Kolmo...
In statistics, the Gumbel distribution is a key tool for modeling extreme values across diverse fields such as finance, hydrology, and engineering. However, existing tests for verifying the goodness-of-fit of the Gumbel distribution, such as the Kolmogorov-Smirnov (K-S) or Anderson-Darling (A-D) tests, are omnibus tests. Omnibus tests measure overall misfit across the entire distribution, resulting in low sensitivity in the sparse tail regions of the data. This leads to a significant problem where extreme risk can be underestimated. Therefore, this study proposes two new tail-sensitive goodness-of-fit test statistics—Expected Shortfall Discrepancy (ESD) and Tail Concentration Ratio (TCR)—based on financial risk management theory where the Gumbel distribution is actively applied, aiming to overcome these limitations. ESD measures the standardized difference between the empirical expected loss (Empirical ES) of the data and the theoretical expected loss (Theoretical ES) of the Gumbel model to verify the average tail size. TCR splits the tails into near and far tails and verifies the tail shape and concentration using a chi-squared test that compares the observed frequency of each interval to the theoretical expected frequency. Monte Carlo simulation results show that the proposed ESD and TCR tests exhibit significantly higher power than existing K-S and A-D tests for alternative distributions with heavier tails than the Gumbel distribution. Furthermore, in an analysis using actual S&P 500 monthly maximum loss data, the K-S and A-D tests failed to reject the Gumbel distribution at the 0.05 significance level. However, the proposed ESD and TCR tests revealed that the Gumbel model significantly underestimates tail risk and rejected the null hypothesis at the 0.05 significance level. This study demonstrates that specialized verification for the tail region is essential to ensure the reliability of extreme value modeling and presents ESD and TCR as practical alternatives that address the shortcomings of existing testing methods.