The Data Space

Input data shared by every by-policy projection.

The six input CSVs are read here, once per model, and referenced from Projection as data. Projection is parameterized by point_id, so each Projection[N] is a separate ItemSpace with its own cells cache; if the readers lived there, every model point would re-read every file. Holding them in an unparameterized Space reads each file once no matter how many policies are projected.

Inputs are external files: plain CSVs in the model folder’s parent directory, products/risikolebensversicherung/, rather than data stored inside the model. The model folder therefore holds nothing but formulas — no _data/, no IOSpec, no embedded values — so a diff of the model shows logic changes only. This follows annuallife.TradLife_A; contrast basiclife.BasicTerm_S, which keeps its inputs inside the model through modelx’s IOSpec machinery.

The consequence worth knowing: the model is not portable on its own. Copying the RLV_DE_S folder without its parent’s CSVs produces a model that reads and then fails on first evaluation.

input_dir() resolves the directory from _model.path.parent at run time, so the model works wherever the repository is checked out. Each table has a filename Reference and a reader Cells:

Reference	Cells	File
model_point_file	model_point_table()	model_point_table.csv
mort_table_file	mort_table()	mort_table.csv
benefit_schedule_file	benefit_schedule()	benefit_schedule.csv
nvg_schedule_file	nvg_schedule()	nvg_schedule.csv
lapse_file	lapse_table()	lapse_table.csv
freq_loading_file	freq_loading_table()	freq_loading_table.csv

Every file but the model point table carries a per-row ``provenance`` column. That is this library’s second ruling and it is machine-checked in tests/test_model_conventions_de.py: a number in a shipped input file says where it came from, in the same [S#] / [R#] / [REG-R#] / [std] vocabulary the documents use. model_point_table.csv is the single exemption, because a model point is a configuration — one policy’s own terms — rather than an assumption.

The mortality table is a proxy, and here is its anchor

mort_table.csv is a [std] Gompertz-form proxy, not a fitted or supervised table:

mort_rate(sex, smoker, x) = base(sex) x smoker_mult(smoker) x 1.095^(x - 30)

base(M) = 0.00040      base(F) = 0.00020
smoker_mult(N) = 1.00  smoker_mult(R) = 2.20        ages 18 to 80

The German first-order basis for a term product is DAV 2008 T, with its DAV 2008 T NR and DAV 2008 T R smoker variants — derived by the DAV Arbeitsgruppe Biometrische Rechnungsgrundlagen from German insurers’ own policy data over 2006 to 2008, adopted 4 December 2008 and restated as a Fachgrundsatz dated 29 November 2022, and expressly suitable for premium calculation but not for policies written without a Gesundheitsprüfung. Those tables are the property of the Deutsche Aktuarvereinigung, are not public, and are not redistributed here. They are cited by name; the shipped proxy stands in for them.

What a replacement must preserve, so that the notes’ worked example still closes:

1. The 50/50 unisex non-smoker blend is 0.00030 x 1.095^(x - 30). That is the [std] best-estimate scale the research file constructed and froze, and it is what the tariff — which may not rate on sex — is actually built on. A replacement table whose male and female non-smoker rates average to this at every age reproduces every premium in the model unchanged.

2. The female-to-male ratio is 0.50 at every age, the order of magnitude reported for insured lives at the ages this product is sold [unverified]. It is what decides the size of the unisex cross-subsidy between model points 1 and 2, and it moves no premium.

3. The smoker multiplier is 2.20, the mid-point of the two-to-three range reported for insured-lives smoker mortality at working ages [unverified]. It reproduces a premium ratio near 2.0 between model points 1 and 3 once the sum-related and per-policy expense elements, which do not scale with mortality, are added back.

The 9,5 % per year of age is the slope of the research file’s own construction; it is a fitted-in-spirit gradient with no German source, and on model point 14’s forty-year run it is the single most exposed number in the model. A population table is the wrong starting point for a replacement: a Destatis series without a selection adjustment overstates claims by a wide margin at the issue ages 25–45 this product is sold at. Dropping a licensed or company table in place of this one changes the basis with no formula change.

The other four assumption files

The three schedule files are keyed on policy_year, the contractual 1-based label, and their values are left as they are: Projection reads them at t + 1, t being the model’s 0-based period index.

benefit_schedule.csv carries the three German Versicherungssumme shapes as factors on the initial sum: konstant (1.0 at every year, the majority form, shipped for forty years), linear_fallend ((21 - policy_year)/20, i.e. f(t) = (20 - t)/20, shipped for the twenty-year term model point 4 is written on) and annuitaet_fallend_3pct (the outstanding balance of a thirty-year annuity loan at 3,00 % nominal, shipped for model point 5’s thirty-year term). The falling schedules are term-specific by construction: an amortisation shape is agreed at issue for a stated term, so a schedule id is written for the term it belongs to rather than truncated from a longer one. A model that hard-codes a constant sum insured cannot represent two of the three shapes the German market sells, which is why the schedule is a first-class input.

nvg_schedule.csv carries the cumulative Nachversicherungsgarantie multiplier sum_uplift: keine is 1.0 throughout and is the base run, and nvg_zwei_erhoehungen steps to 1.2 at policy year 6 (t = 5) and 1.4 at policy year 12 (t = 11). No event list, cap, exercise window or age limit was established from any document, so the take-up is exogenous — a schedule, not a modelled decision. What the model does with it is not exogenous: each increment carries its own three-year § 161 window, so suicide_factor is a weighted average across tranches in the years after an increase.

lapse_table.csv is 6 % in policy year 1, 4 % in years 2 and 3 and 3 % thereafter, all [std], argued from three structural features rather than from data: nothing is forfeited by lapsing, exit is frictionless in time because the Versicherungsperiode follows the Zahlweise, and the need that motivated the purchase amortises. The GDV whole-market Stornoquote is a book average dominated by long-dated savings contracts and is deliberately not used. The table’s own row for the final policy year still reads 3 %; the model zeroes it, because a lapse and an expiry at the end of the last period t = n - 1 are the same event paying the same nothing.

freq_loading_table.csv carries the Ratenzahlungszuschlag — 1.000 annual, 1.02 half-yearly, 1.03 quarterly, 1.05 monthly — a German market convention with no carrier attribution, and the instalment count that goes with it. Whether carriers strike the loading on the Bruttobeitrag or the Zahlbeitrag was not established; this model loads the billed amount, so the Brutto = Zahl + Verrechnung identity holds at every frequency.

Cells Descriptions

input_dir()

The directory holding the input CSVs: the model folder’s parent.

Inputs are external files, not data stored inside the model, so the model folder is pure formulas. The path is resolved at run time from where the model was read, following annuallife.TradLife_A.

model_point_table()

The model point table, read from model_point_table.csv.

The one input file with no provenance column, and the only one exempt from the rule: a model point is a configuration — one policy’s issue age, sum insured, Zahlweise and schedule ids — rather than an assumption about the world.

mort_table()

The second-order annual death rates, from mort_table.csv.

Indexed by table_id, sex, smoker and attained age. A [std] Gompertz proxy standing in for DAV 2008 T, which is proprietary and is cited by name rather than shipped; see the Space docstring for the three anchors a replacement must preserve. The table is second order — a best estimate for medically selected lives. The first-order tariff rate is built from it in Projection.mort_rate_tar by taking the unisex blend and applying the Sicherheitszuschlag, so there is exactly one unsourced mortality level in the model rather than two stacked on each other.

benefit_schedule()

The Versicherungssumme factors by schedule id and 1-based policy year.

Read from benefit_schedule.csv. konstant, linear_fallend and annuitaet_fallend_3pct are the three German shapes; the two falling ones are written for the term of the model point that uses them, an amortisation schedule being agreed at issue for a stated term. policy_year is the contractual label and runs from 1; Projection.benefit_factor reads it at t + 1.

nvg_schedule()

The cumulative Nachversicherungsgarantie multiplier by schedule id and 1-based policy year.

Read from nvg_schedule.csv. keine is 1.0 throughout — the base run — and nvg_zwei_erhoehungen steps twice. Take-up is exogenous: no event list, cap, window or age limit was established from any document, so an increase is supplied as a schedule rather than modelled as a decision. policy_year is the contractual label and runs from 1; Projection.sum_uplift reads it at t + 1.

lapse_table()

The annual lapse rates by 1-based policy year, read from lapse_table.csv.

Entirely [std]: no Risikoversicherung-specific German rate exists in the source corpus, and the whole-market Stornoquote is deliberately not used. The final policy year’s row is a live 3 % and is overridden to zero by Projection.lapse_rate, which is where the convention belongs — the zero is a property of the last policy year, not of the table.

freq_loading_table()

The Ratenzahlungszuschlag and instalment count by Zahlweise.

Read from freq_loading_table.csv: the multiplier prem_freq_load applied to the billed amount, and instalments, the number of payments a year, which is carried for reporting and enters no cash flow on an annual grid.