Quantitative risk dashboard

Sample · claims aggregate risk

What is a 1-in-200 claims month, and what should a 200k xs 50k layer cost?
Source: claims.csv

events: 2,000periods: 36 × Msimulations: 50,000Generated: 2026-11-15 00:29
Expected events per M
55.6
Poisson · dispersion 1.07
Mean severity
USD 27.2k
Lognormal mixture (2)
Expected loss per M
USD 1.51M
σ USD 663k
1-in-200 loss (99.5% VaR)
USD 4.01M
TVaR USD 4.64M
Capital buffer at 99.5%
USD 2.5M
VaR minus expected loss

What if…

Stress the frequency and severity, move the reinsurance layer or the confidence level, then press Recalculate. Expected values and layer prices are exact; VaR and TVaR are re-simulated in your browser with a 95% interval.

Interpretation & hypotheses

Take it away

The PDF includes your edits, notes, working hypothesis and current what-if scenario. The Excel file holds the cleaned data, every table and every chart's numbers.

Interpretation

A 1-in-200 claims month costs about USD 4.01M (95% interval USD 3.94M to 4.09M); the exact FFT calculation gives USD 3.99M, inside that interval. An average month costs USD 1.51M, so covering a 1-in-200 month needs about USD 2.5M of capital above the expected loss.

Claims split into two groups: about 86% are routine claims and 14% are large ones, and the two-component model reproduces the observed average claim (USD 27.2k against 27.3k in the data). Counts show no clustering and no trend over the 36 months, so the Poisson frequency is a fair description.

A layer of USD 200k excess of USD 50k per claim pays about USD 10.4k per claim on average, a pure premium of USD 578k a month; the simulation agrees (USD 578k).

With that capital and a 25% premium loading, the chance of running out of money within 12 months is about 3.1% (± 0.2%). The main caveat is the short history: 36 months cannot show how often a truly extreme claim arrives, so treat the far tail with care.

Your working hypothesis

What do you expect, or want to test, before acting on these numbers? Saved in this browser and printed in the PDF.

Hypotheses tested

Every formal test behind this dashboard. H₀ is the null hypothesis; it is rejected when p < 0.05.

Null hypothesis (H₀)TestStatisticp-valueDecision at 5%What it means
There is no trend in events per M over timeSpearman rank correlation with timeρ = 0.0320.852✓not rejectedThe rate looks stable over the period.
Event counts are Poisson (no clustering)Likelihood ratio vs negative binomial (boundary-corrected)LR = 0.020.441✓not rejectedNo evidence of clustering beyond chance.
The loss amounts follow the fitted Lognormal mixture (2)Kolmogorov-Smirnov, parametric bootstrap (199 refits, n = 2,000)D = 0.00900.94✓not rejectedThe model is consistent with the data.
The loss amounts follow the fitted Lognormal mixture (2) (tail-sensitive)Anderson-Darling, parametric bootstrapA² = 0.1700.785✓not rejectedNo evidence the tails deviate from the model.
The model's mean loss equals the data'sz-test against the sample meanz = -0.060.953✓not rejectedThe model reproduces the observed average.

Aggregate loss per period

Collective risk model S = X₁ + … + X_N, simulated 50,000 times and computed independently with the FFT of the probability generating function; the two must agree.

Distribution of total loss per period

Bars: simulation. Line: exact discretised compound distribution.

02e-074e-076e-078e-07USD 1MUSD 2MUSD 3MUSD 4Mtotal loss per perioddensity
Data table
bin frombin todensity
USD 179kUSD 286k3.735e-09
USD 286kUSD 394k2.559e-08
USD 394kUSD 501k8.721e-08
USD 501kUSD 608k1.952e-07
USD 608kUSD 715k3.141e-07
USD 715kUSD 822k4.523e-07
USD 822kUSD 929k5.786e-07
USD 929kUSD 1.04M6.613e-07
USD 1.04MUSD 1.14M6.777e-07
USD 1.14MUSD 1.25M7.072e-07
USD 1.25MUSD 1.36M6.981e-07
USD 1.36MUSD 1.46M6.742e-07
USD 1.46MUSD 1.57M6.217e-07
USD 1.57MUSD 1.68M5.605e-07
USD 1.68MUSD 1.79M4.945e-07
USD 1.79MUSD 1.89M4.241e-07
USD 1.89MUSD 2M3.649e-07
USD 2MUSD 2.11M3.126e-07
USD 2.11MUSD 2.21M2.672e-07
USD 2.21MUSD 2.32M2.123e-07
USD 2.32MUSD 2.43M1.899e-07
USD 2.43MUSD 2.54M1.419e-07
USD 2.54MUSD 2.64M1.285e-07
USD 2.64MUSD 2.75M9.842e-08
USD 2.75MUSD 2.86M7.526e-08
USD 2.86MUSD 2.96M6.761e-08
USD 2.96MUSD 3.07M5.752e-08
USD 3.07MUSD 3.18M4.501e-08
USD 3.18MUSD 3.29M3.81e-08
USD 3.29MUSD 3.39M3.007e-08
USD 3.39MUSD 3.5M2.222e-08
USD 3.5MUSD 3.61M1.662e-08
USD 3.61MUSD 3.71M1.961e-08
USD 3.71MUSD 3.82M1.065e-08
USD 3.82MUSD 3.93M9.525e-09
USD 3.93MUSD 4.03M7.657e-09
USD 4.03MUSD 4.14M9.151e-09
USD 4.14MUSD 4.25M5.976e-09
USD 4.25MUSD 4.36M5.603e-09
USD 4.36MUSD 4.46M4.482e-09
USD 4.46MUSD 4.57M4.109e-09
USD 4.57MUSD 4.68M2.428e-09
USD 4.68MUSD 4.78M1.494e-09
USD 4.78MUSD 4.89M1.494e-09
USD 4.89MUSD 5M1.868e-09

Return-period loss curve

The total loss exceeded on average once every N periods.

USD 1MUSD 2MUSD 3MUSD 4MUSD 5M1-in-21-in-51-in-101-in-201-in-501-in-1001-in-2001-in-5001-in-1000return periodloss
Data table
xloss
1-in-2USD 1.4M
1-in-5USD 1.98M
1-in-10USD 2.36M
1-in-20USD 2.73M
1-in-50USD 3.22M
1-in-100USD 3.6M
1-in-200USD 4.01M
1-in-500USD 4.5M
1-in-1000USD 5M

Risk measures

LevelReturn periodVaR · simulation95% CIVaR · latticeTVaR · simulationTVaR · lattice
90%1-in-10USD 2.36MUSD 2.35M – USD 2.37MUSD 2.37MUSD 2.9MUSD 2.91M
95%1-in-20USD 2.73MUSD 2.7M – USD 2.74MUSD 2.73MUSD 3.28MUSD 3.28M
99%1-in-100USD 3.6MUSD 3.55M – USD 3.66MUSD 3.6MUSD 4.2MUSD 4.21M
99.5%1-in-200USD 4.01MUSD 3.94M – USD 4.09MUSD 3.99MUSD 4.64MUSD 4.65M

Cross-checks (analytic vs simulated vs exact)

Differences beyond the Monte Carlo error mean something is wrong.

QuantityAnalytic / exactSimulatedDifference
Mean E[S] = E[N]·E[X]USD 1.51MUSD 1.51M ± USD 5.78k-0.07%
Std dev √(E[N]Var[X] + Var[N]E[X]²)USD 663kUSD 659k-0.62%
99% VaR · lattice vs simulationUSD 3.6MUSD 3.6M+0.13%
Probability mass captured by the FFT of the probability generating function100%100%

Frequency

Events per M

Complete periods only; empty periods count as zero.

40506070802022-012022-072023-012023-072024-012024-07periodeventsfitted mean
Data table
xevents
2022-0153
2022-0252
2022-0375
2022-0455
2022-0565
2022-0659
2022-0756
2022-0852
2022-0952
2022-1059
2022-1145
2022-1263
2023-0149
2023-0264
2023-0364
2023-0457
2023-0546
2023-0644
2023-0744
2023-0844
2023-0960
2023-1061
2023-1150
2023-1248
2024-0148
2024-0251
2024-0362
2024-0451
2024-0546
2024-0657
2024-0766
2024-0865
2024-0953
2024-1064
2024-1159
2024-1261

Trend test

Spearman ρ (counts vs time)p-valuefirst 12 meanlast 12 meanratio
0.0328.52e-0157.256.91.00×

Frequency model

Dispersion index variance/mean = 1.07. Likelihood-ratio test of Poisson against negative binomial: LR = 0.02, p = 0.4412 (boundary-corrected ½χ²₁). No evidence of over-dispersion; Poisson is adequate.

Observed vs fitted count probabilities

Share of periods with exactly k events.

observedPoissonNegative binomial
0%2%4%6%8%10%0204060events per period (k)probability
Data table
xobservedPoissonNegative binomial
00%0%0%
10%0%0%
20%0%0%
30%0%0%
50%0%0%
60%0%0%
70%0%0%
80%0%0%
100%0%0%
110%0%0%
120%0%0%
130%0%0%
150%0%0%
160%0%0%
170%0%0%
180%0%0%
200%0%0%
210%0%0%
220%0%0%
230%0%0%
250%0%0%
260%0%0.001%
270%0.001%0.001%
280%0.002%0.002%
300%0.006%0.007%
310%0.011%0.013%
320%0.019%0.023%
330%0.032%0.037%
350%0.084%0.094%
360%0.13%0.14%
370%0.19%0.21%
380%0.28%0.31%
400%0.56%0.6%
410%0.76%0.8%
420%1%1.1%
430%1.3%1.3%
452.8%2%2.1%
465.6%2.5%2.5%
470%2.9%2.9%
485.6%3.4%3.4%
502.8%4.2%4.2%
515.6%4.6%4.6%
528.3%4.9%4.9%
535.6%5.2%5.1%
552.8%5.4%5.3%
562.8%5.3%5.2%
575.6%5.2%5.1%
580%5%4.9%
602.8%4.3%4.3%
615.6%3.9%3.9%
622.8%3.5%3.5%
632.8%3.1%3.1%
655.6%2.3%2.3%
662.8%1.9%2%
670%1.6%1.6%
680%1.3%1.4%
700%0.84%0.88%
710%0.66%0.69%
720%0.51%0.54%
730%0.39%0.41%
740%0.29%0.31%

Count models

ModelMLE parameterslog-likAICΔAICχ²dfχ² p
Poissonλ = 55.56-123.8249.70.03.7040.4475
Negative binomialr (size) = 1584, p = 0.9661-123.8251.72.00.8930.8273
Geometric (from 0)p = 0.01768-180.9363.9114.2134.6540.0000

Severity

Lognormal mixture (2) is the best fit by AIC. Goodness of fit is not rejected (Kolmogorov-Smirnov p = 0.940, Anderson-Darling p = 0.785; parametric bootstrap with 199 refits on n = 2,000).

Severity distribution

Bars: individual losses. Lines: fitted models.

Lognormal mixture (2)Burr XIIInverse Gaussian
00.250.50.7511.25USD 1kUSD 10kUSD 100kUSD 1Mvalue (log scale)density of log10(value)
Data table
bin frombin todensity
USD 466USD 5740.02774
USD 574USD 7060.03884
USD 706USD 8690.04439
USD 869USD 1.07k0.1054
USD 1.07kUSD 1.32k0.172
USD 1.32kUSD 1.62k0.3551
USD 1.62kUSD 1.99k0.4883
USD 1.99kUSD 2.45k0.5271
USD 2.45kUSD 3.02k0.6825
USD 3.02kUSD 3.71k0.9044
USD 3.71kUSD 4.57k1.043
USD 4.57kUSD 5.62k0.9932
USD 5.62kUSD 6.92k0.91
USD 6.92kUSD 8.52k0.8434
USD 8.52kUSD 10.5k0.7768
USD 10.5kUSD 12.9k0.5992
USD 12.9kUSD 15.9k0.4494
USD 15.9kUSD 19.5k0.2996
USD 19.5kUSD 24k0.1997
USD 24kUSD 29.6k0.1054
USD 29.6kUSD 36.4k0.1165
USD 36.4kUSD 44.8k0.09433
USD 44.8kUSD 55.1k0.2108
USD 55.1kUSD 67.8k0.1165
USD 67.8kUSD 83.5k0.1221
USD 83.5kUSD 103k0.1387
USD 103kUSD 126k0.09433
USD 126kUSD 156k0.1165
USD 156kUSD 191k0.1609
USD 191kUSD 236k0.111
USD 236kUSD 290k0.07213
USD 290kUSD 357k0.05549
USD 357kUSD 439k0.03329
USD 439kUSD 540k0.03329
USD 540kUSD 665k0.005549
USD 665kUSD 818k0.02774
USD 818kUSD 1.01M0.01665
USD 1.01MUSD 1.24M0
USD 1.24MUSD 1.52M0
USD 1.52MUSD 1.88M0.005549

Q-Q plot · Lognormal mixture (2)

Points on the dashed line mean the fitted model reproduces the observed quantiles. Bending away at the right end = the tail is mis-modelled.

USD 100USD 1kUSD 10kUSD 100kUSD 1MUSD 10MUSD 100USD 1kUSD 10kUSD 100kUSD 1MUSD 10Mmodel quantileobserved quantile
Data table
model quantileobserved quantile
344466
940931
1.24k1.25k
1.47k1.49k
1.62k1.59k
1.8k1.76k
1.97k1.91k
2.09k2.05k
2.25k2.22k
2.41k2.37k
2.52k2.5k
2.67k2.69k
2.82k2.87k
2.94k2.96k
3.09k3.11k
3.25k3.28k
3.36k3.38k
3.52k3.54k
3.68k3.71k
3.81k3.81k
3.98k3.97k
4.15k4.16k
4.29k4.32k
4.47k4.43k
4.66k4.62k
4.8k4.76k
5.01k4.98k
5.22k5.16k
5.39k5.41k
5.61k5.62k
5.86k5.91k
6.04k6.08k
6.31k6.38k
6.59k6.63k
6.81k6.89k
7.13k7.16k
7.47k7.52k
7.75k7.8k
8.14k8.28k
8.58k8.75k
8.92k9.1k
9.45k9.58k
10k10.1k
10.5k10.6k
11.3k11.4k
12.1k12.3k
12.9k12.9k
14.1k13.9k
15.8k15.5k
17.4k17.1k
20.4k19.8k
25.4k24.5k
31.8k33.3k
45k47.1k
63.4k60.2k
79.9k74.7k
108k105k
148k158k
196k192k
323k322k
1.51M1.88M

Severity models

ModelMLE parameterslog-likAICΔAICKS DAkaike weight
Lognormal mixture (2)w1 = 0.8569, μ1 = 8.507, σ1 = 0.775, w2 = 0.1431, μ2 = 11.47, σ2 = 0.9444-20,936.941,883.80.00.0090100%
Burr XIIc = 2.88, d = 0.2936, scale = 2665-20,999.842,005.7121.90.03230%
Inverse Gaussianmean μ = 2.73e+04, shape λ = 5016-21,111.842,227.6343.80.13860%
Log-logisticshape β = 1.444, scale α = 6478-21,171.042,345.9462.10.06170%
Lognormalμ (log-mean) = 8.931, σ (log-sd) = 1.312-21,243.242,490.4606.60.11710%
Pareto II (Lomax)tail index α = 1.325, scale θ = 9726-21,311.042,626.0742.20.12330%
Weibullshape k = 0.6211, scale λ = 1.554e+04-21,709.443,422.71,538.90.17880%
Gammashape α = 0.4957, scale θ = 5.506e+04-21,983.843,971.52,087.70.25910%
Exponentialmean θ = 2.73e+04-22,429.044,859.92,976.10.39640%
Lognormal mixture (3)w1 = 0.3477, μ1 = 7.988, σ1 = 0.6399, w2 = 0.5002, μ2 = 8.839, σ2 = 0.6277, w3 = 0.152, μ3 = 11.39, σ3 = 0.9825-20,935.941,887.83.90.0100not selected: BIC prefers Lognormal mixture (2)
Lognormal mixture (4)w1 = 0.346, μ1 = 7.915, σ1 = 0.6052, w2 = 0.2259, μ2 = 8.455, σ2 = 0.3775, w3 = 0.2656, μ3 = 9.212, σ3 = 0.4316, w4 = 0.1625, μ4 = 11.3, σ4 = 1.027-20,934.241,890.36.50.0065not selected: BIC prefers Lognormal mixture (2)

Model checks · Lognormal mixture (2)

A good fit reproduces the observed mean, spread and centre. Large gaps mean conclusions depend on the model's tail more than on the data.

QuantityModelDataVerdict
MeanUSD 27.2kUSD 27.3k ± USD 3.76k✓ within the 95% interval
Standard deviationUSD 84.7kUSD 85.9k
MedianUSD 5.83kUSD 5.87k
Expected count above the largest observation n·S(max)0.24≈ 1 or less is consistent✓

Excess-of-loss layer

Per-event layer USD 200k xs USD 50k: the reinsurer pays min((X − R)⁺, L) on every event.

Layer pricing

QuantityValue
Expected ceded per event ∫ S(x) dx over the layerUSD 10.4k
Pure premium per period = E[N] × expected ceded per eventUSD 578k
Simulated ceded per period (check)USD 578k ± USD 2.55k
Layer share of expected loss38.3%
Probability the layer is hit in a period99.8%
99.5% VaR of ceded lossUSD 1.47M

Solvency (Cramér-Lundberg ruin model)

Surplus U(t) = u + c·t − S(t) with Poisson arrivals at the fitted rate. Ruin = surplus below zero at any claim.

Severity has no finite moment generating function, so no Lundberg coefficient exists. The subexponential asymptotic ψ(u) ≈ (1/(θ E[X])) ∫_u^∞ S(x) dx describes ruin by one huge claim and is only accurate when u is large relative to single claims; with many claims per period the diffusion approximation is the relevant one.

Ruin analysis

QuantityValue
Initial capital uUSD 2.5M
Premium loading θ25%
Premium income per M c = (1+θ)λE[X]USD 1.89M
Ruin probability within 12 M (simulated, 20,000 paths)3.1% ± 0.24%
Diffusion approximation, within 12 M (reliable only for many, light-tailed claims; the simulation is the reference)1.2%
Diffusion approximation, infinite horizon exp(−2θλE[X]u / λE[X²])1.4%
Single-large-claim asymptotic, infinite horizon (valid only for very large u)0%

Method

Collective risk model. S = Σᵢ₌₁ᴺ Xᵢ, N independent of the i.i.d. Xᵢ. E[S] = E[N]E[X]; Var[S] = E[N]Var[X] + Var[N]E[X]².

Exact compound distribution. Severity discretised on a lattice of span h by the rounding method fₖ = F((k+½)h) − F((k−½)h). Panjer: f_S(x) = [1/(1 − a f₀)] Σ_{y=1}^{x} (a + by/x) f_y f_S(x−y), with (a,b) = (0, λ) for Poisson and (1−p, (r−1)(1−p)) for NB. Large E[N]: f_S = IFFT(P_N(FFT f)) with zero padding.

Simulation. N drawn from the fitted count model, severities from the fitted severity model; quantile CIs from order statistics.

Layer pricing. E[min((X − R)⁺, L)] = ∫_R^{R+L} S_X(x) dx; pure premium = E[N] × that.

Ruin. Cramér-Lundberg U(t) = u + ct − S(t), c = (1+θ)λE[X]. Lundberg: λ + cR = λM_X(R), ψ(u) ≤ e^(−Ru); exponential claims: ψ(u) = e^(−Ru)/(1+θ). Subexponential claims: ψ(u) ~ (1/(θE[X])) ∫_u^∞ S_X(x) dx.

Assumptions and limitations

  • Frequency and severity are assumed independent, and events independent of each other. Contagion or common shocks make the real tail heavier than shown.
  • Severity parameters are fitted on all history; inflation or a change in the mix of events is not trended. Trend the amounts first if the period is long.
  • The ruin model uses Poisson arrivals at the mean rate; with over-dispersed counts (dispersion 1.07) the true ruin probability is higher.
  • Monte Carlo figures carry sampling error; intervals are shown where it matters.