Caldera Re data files

Every worked example on this site runs on the files listed here: one year event loss table (YELT) per cedent of Caldera Re, at two trial counts. The data is fictitious. The Caldera Re appendix describes how it was generated and what each cedent’s book contains, and Demo scale states what 20 trials can and cannot resolve. The figures in the tables are read from the files when the site is built.

The 20-trial demo tier

The worked examples and the figures quoted in the text come from these five files, unless a page names the other tier.

File	Cedent	Trials with rows	Occurrences	Rows	Size
suncoast_20.csv	SunCoast	20	530	1,657	99 KB
pacific_20.csv	Pacific Mutual	20	374	1,257	82 KB
baltica_20.csv	Baltica	20	192	578	34 KB
atlas_20.csv	Atlas Cat Fund	20	391	391	25 KB
trident_20.csv	Trident Re	19	47	94	6 KB
All 5 files				3,977	246 KB

The 1,000-trial tier

The same simulation at a higher trial count, for the resolution contrast that Demo scale describes.

File	Cedent	Trials with rows	Occurrences	Rows	Size
suncoast_1000.csv	SunCoast	1,000	25,754	79,473	4.9 MB
pacific_1000.csv	Pacific Mutual	1,000	17,894	59,451	3.9 MB
baltica_1000.csv	Baltica	1,000	8,567	25,558	1.6 MB
atlas_1000.csv	Atlas Cat Fund	1,000	19,492	19,492	1.3 MB
trident_1000.csv	Trident Re	874	2,106	4,212	266 KB
All 5 files				188,186	11.9 MB

Caldera Re

Neither tier holds a US earthquake tail. SunCoast and Pacific Mutual share one catalog of 300 US earthquake events, and the 1,000-trial tier contains 299 of them. Its costliest US earthquake occurrence costs Pacific Mutual $41.3M and SunCoast $14.5M, against $29.8M and $10.4M in the 20-trial tier. Demo scale states what this means for the contrast between the tiers.

Reading the tables

$N$, the number of trials simulated, is the number every exceedance probability divides by, and the file name records it. Trials with rows can be smaller, because a trial without occurrences has no row. Trident Re’s 20-trial file holds rows for 19 trials, and its $N$ is still 20; Building an EP curve works through the difference. An occurrence is one (trial_id, timestamp, event_id) key, and it can span several rows, one per geography and line of business. Two cedents’ files can hold rows of the same occurrence (SunCoast and Pacific Mutual share the US earthquake events), so occurrence counts do not add across files and the total row leaves them out.

The schema

All ten files share the nine columns of the site’s YELT schema: the four-column core trial_id, timestamp, event_id, loss, and the enrichment peril, geography_id, lob_id, cedent_id, tiv. loss and tiv are in millions of US dollars and strictly positive, under the site’s sign convention.

Loading a file

The runnable examples load a cedent through the site’s caldera_re package, which records $N$ on the frame:

from caldera_re.data import cedents, load_cedent

yelt = cedents.suncoast                  # suncoast_20.csv
yelt = load_cedent("suncoast", 1000)     # suncoast_1000.csv
yelt.attrs["n_trials"]                   # 1000, from the file name

Outside the site, plain pandas reads a downloaded file. State $N$ yourself, because the rows cannot supply it:

import pandas as pd

yelt = pd.read_csv("trident_20.csv", parse_dates=["timestamp"])   # downloaded from the table above
yelt.attrs["n_trials"] = 20              # not yelt["trial_id"].nunique(), which is 19

Checking a download

The files are frozen, and the worked examples load them as published. The program that generated them is maintained privately, so a reader cannot rerun the simulation, only check that a downloaded copy is the file the site uses. On the frame loaded above, pandas recomputes the counts in the tables and the mean trial loss that the dataset summary on the Caldera Re page prints:

yelt["trial_id"].nunique()                                         # 19 trials with rows
len(yelt.drop_duplicates(["trial_id", "timestamp", "event_id"]))   # 47 occurrences
len(yelt)                                                          # 94 rows
yelt["loss"].sum() / yelt.attrs["n_trials"]                        # 41.45, mean trial loss ($M)

The SHA-256 of each file, computed from the published bytes when the site is built, checks the bytes themselves:

File	SHA-256
suncoast_20.csv	995fc69fc2d26c4bee2331a8c1a4102ef8a2f83fe05e395c61ffce781cab4de4
pacific_20.csv	1b96c2efa3903870f15cbf2c8e49d78b50f1ad6c0f2703f07badd67f68f276d6
baltica_20.csv	d2a778687a81dfd8689c31b5b67e6e5fbadcbfa311da45b5d00df76b1401beee
atlas_20.csv	715fe16ff61f7857ea461477c7cc44b97a09b057de467fa77b25eb45f634700d
trident_20.csv	7a970fb92cda918588050d5ec28f9d408f306f59bc235fee530681ff88bd1a84
suncoast_1000.csv	abc2136034251cff1139746b5b450249784f494bc5416a4bd55fe04b34dab3ff
pacific_1000.csv	216e2f576316978783003313e02c669e0b18be5737a71715e9fa00922db1b62e
baltica_1000.csv	e27cfc7e3cd7de21234ce72270cb3f1d44e36382e1167f0e16ed98cb35891088
atlas_1000.csv	d5b4ec11f29d7ccc35b53121cc474be89dce1f3efec9ade5fd02d0479a10499e
trident_1000.csv	a0969af1d10b0f7323779fa597938c459a487efad56be4b208f8ef669747cb3b

sha256sum trident_20.csv on Linux, shasum -a 256 trident_20.csv on macOS and certutil -hashfile trident_20.csv SHA256 on Windows print the digest to compare.