Julien Hambuckers, Andreas Groll and T. Kneib, "Understanding the Economic Determinants of the Severity of Operational Losses: A Regularized Generalized Pareto Regression Approach", Journal of Applied Econometrics, Vol. 33, No. 6, 2018, pp. 898-935. The losses were provided by UniCredit ORM leader, F. Piacenza (UniCredit S.p.A.). We received 40,871 losses, covering a period ranging from January 2005 to June 2014. The losses were scaled by an unknown factor for anonymity reasons. In addition, the losses were adjusted for inflation, using the Italian consumer price index. We provide the losses adjusted for inflation. Regarding the event types (ET), they were provided by UniCredit and coded into binary variables. In the paper, we use the category DPA (Damage to Physical Assets) as the reference category. The 20 other economic explanatory variables stem from various sources. Firm-specific variables are internal ratios and have been provided by the operational risk department of UniCredit. Macroeconomic and financial variables have been downloaded from Thomson Reuters Datastream for S\&P500, TRSI, MIB, VIX, VFTSE and the unemployment rate, from EUROSTAT website for the GDP growth rates (http://ec.europa.eu/eurostat/data/database), from FRED, the database of the Federal Reserve Bank of St. Louis (https://fred.stlouisfed.org) for the housing price index and the Italian interest rates (notice that FRED also aggregates data from the OECD website - http://stats.oecd.org - where detailed definitions of the time series are provided). Last, the monetary aggregate M1 and the consumer loans rates have been downloaded from ECB database. In the file oploss_unicredit_JAE.xlsx, we provide for each loss (first column), the year and the quarter of registration (second and third column). These values were used to associate one-quarter lagged values of the explanatory variables (e.g. if loss X has been registered in 2005Q2, we associate regressors from period 2005Q1). Detailed descriptions of the covariates and of the headers can be found in the paper, Appendix D. In columns D to J, we provide the ET coded in 7 binary variables (column J must be omitted for the regression). Column K to AD display the economic covariates. An important step in our approach consists in selecting extreme losses. To do so, we compute event type-specific quantiles at level 75%, and we keep only losses above these thresholds to conduct the regression analysis. In the file script_JAE.m, the interested reader will find the Matlab script used to compute these thresholds. The final sample size is 10,217. The file hgk-files.zip contains script_JAE.m and a CSV version of the dataset named oploss_unicredit_JAE.csv. Unix/Linux users should use "unzip -a". Julien Hambuckers jhambuc[at]uni-goettingen.de