Research question and scope
Orbital collisions can generate fragments that increase the probability of subsequent collisions. This study examines how that feedback propagates into losses for a hypothetical insured portfolio.
The comparator is a constant annual charge equal to 0.5% of insured hull value. This is an explicit modelling assumption, not a survey of insurance quotations. The analysis does not test whether commercial insurers already account for debris risk.
Collisional cascading refers to feedback in which collisions produce debris that contributes to further collisions. Net growth depends on production and removal rates. In the baseline model, the linear production term exceeds decay, while the debris–debris term grows approximately as the square of the debris count. This is a property of the assumed model, not evidence that an observed orbital shell has crossed a measured critical threshold.
The physical model supplies debris trajectories to a loss calculation. Risk metrics are computed across the maximum annual loss from each simulated path, then converted to a premium using assumed risk, capital and expense loadings. The resulting 16.5× ratio is conditional on this scenario and pricing rule.
The model provides a framework for examining changes in exposure and pricing assumptions. It does not estimate an optimal premium, establish insurer solvency, or demonstrate how operators would respond to different insurance prices.
Collision and debris model
The model uses a uniform orbital shell spanning 700–900 km, centred at 800 km. It assigns 36,500 debris objects larger than 10 cm and 8,000 active satellites to that shell. These are scenario inputs, not an observed census of that altitude band.
Growth is generated by collision events and assumed fragment yields rather than a separately fitted cascade growth rate. Initial populations, cross-sections, velocities, yield distributions and removal rates remain model parameters and require validation.
Uniform flux approximation
The expected annual collision rate between debris and satellites:
| Parameter | Description | Baseline |
|---|---|---|
| Nd | Debris population (>10 cm) | 36,500 |
| Ns | Active satellites in shell | 8,000 |
| σ | Satellite collision cross-section | 10 m² |
| v | Mean relative velocity | 10 km/s |
| T | Seconds per year | 3.16 × 107 s |
| V | Shell volume | 1.29 × 1011 km³ |
With the baseline inputs, the approximation gives ~7 debris–satellite and ~16 debris–debris collisions per year. Cross-section is converted to km² when used with velocity in km/s and volume in km³.
Under uniform mixing, the debris–debris collision rate is proportional to Nd(Nd − 1)/2 and is approximately quadratic for large populations. Debris–satellite rates are linear in Nd when satellite count is fixed. Separate log-normal yield assumptions for the two channels determine fragment production.
A single stochastic path from a simplified browser model, using a fixed random seed (42). This illustration is not the 1,000-path research ensemble and does not reproduce its risk or premium estimates. The numerical loop stops at 50 years or a debris count above 20 million.
Ready · seed 42
Historical events and yield assumptions
The three fragmentation events below provide context for the scale of debris production. Their reported counts are approximate catalogue totals, not observations directly fitted to the operative yield distributions. They do not independently validate the cascade model.
Reported event characteristics
| Event | Date | Fragments (>10cm) | Mass (kg) |
|---|---|---|---|
| Fengyun-1C (ASAT) | 2007-01-11 | ~3,500 | 880 |
| Iridium 33 × Cosmos 2251 | 2009-02-10 | ~2,300 | 1,510 |
| Kosmos 1408 (ASAT) | 2021-11-15 | ~1,500 | 2,200 |
The assumed log-normal distributions have medians of approximately 245 fragments for debris–satellite events (μ=5.5, σ=0.8) and 3 for debris–debris events (μ=1.1, σ=0.5), before integer rounding. These are priors informed by event scales, not a fitted calibration. The sampler does not explicitly model impact mass or energy.
Loss mapping and premium calculation
Debris trajectories are mapped to annual collision exposure, conditional expected losses and a premium for a hypothetical portfolio of 500 satellites valued at $50M each. This calculation is separate from the illustrative browser simulation.
From collision exposure to loss
Step 1: The annual collision intensity per satellite is Λ(t) = Nd(t) σvT/V. The implementation uses the capped linear approximation phit = min(Λ, 1). For a Poisson process the exact probability is 1 − exp(−Λ). The two are close only at small intensities; the cap does not make the approximation exact at high exposure. Baseline exposure is approximately 0.09% per satellite per year.
Step 2: Multiply collision probability by insured value and aggregate over 500 policies to obtain conditional expected portfolio loss. This loss bridge uses expected claim counts rather than sampling individual claim events. Dependence between claims can affect portfolio tails without changing the sum of marginal means.
Step 3: Take the maximum annual loss along each of the 1,000 paths. Compute the mean, value at risk (VaR) and conditional value at risk (CVaR, or expected shortfall) across this sample. The headline premium uses these statistics; it is not calculated from a one-year-ahead distribution.
Stylised premium decomposition
This stylised pricing model splits the premium into four parts — it is not the calculation prescribed by the Solvency II Directive:
| Component | Description | Amount |
|---|---|---|
| Expected Loss (EL) | Mean of maximum annual portfolio losses | $921.4M |
| Risk Loading | 1.5× (CVaR₉₅ − EL) tail surcharge | $966.0M |
| Capital Charge | 4% × max(VaR₉₉ − EL, 0) | $32.0M |
| Expense Loading | 15% × EL; assumed expense/profit loading | $138.2M |
| Total Premium | $2,058M / yr |
The risk loading ($966.0M) exceeds the expected loss component ($921.4M) under the selected loading multiplier of 1.5. The reported CVaR at 95% is $1,565M. These values describe the simulated sample and selected pricing rule, not a fitted market tariff. Component amounts are rounded.
Reported risk metrics
| Metric | Definition | Value |
|---|---|---|
| VaR₉₅ | 95th percentile of worst-year loss | $1,365M |
| CVaR₉₅ (ES) | Expected shortfall beyond VaR₉₅ | $1,565M |
| VaR₉₉ | 99th percentile of worst-year loss | $1,721M |
| CVaR₉₉ (ES) | Expected shortfall beyond VaR₉₉ | $1,854M |
With 1,000 paths, approximately 50 observations lie in the upper 5% tail and 10 in the upper 1% tail. Both quantiles and tail means are subject to sampling uncertainty. Sample size alone does not determine a standard error; confidence intervals require a stated estimation procedure and additional validation.
Results and interpretation
fixed benchmark
36.5K → 1.51M
per policy per year
fixed benchmark
The $1,933M difference compares a premium derived from maximum annual losses over a 50-year scenario with a constant $125M annual benchmark. It is neither a current funding shortfall nor a measured market pricing error. Expected loss at year 0 is 0.18× the benchmark, but excludes premium loadings. These quantities should not be treated as equivalent measures.
A validated model could support comparisons of orbit selection, mitigation and insurance design. Whether pricing changes would encourage mitigation depends on contracts, costs and operator behaviour, none of which is estimated here.
Methodology and limitations
Computational procedure
Step 1 — Collision-driven growth. Debris evolution is governed by Poisson-distributed collision events at each time step (Δt = 0.1 yr). Fragment yields follow assumed log-normal distributions informed by historical event scales. There is no separately tuned cascade-rate parameter; the initial population, yields, benchmark and portfolio size remain modelling choices and are reported as such.
Step 2 — Monte Carlo ensemble. 1,000 independent 50-year paths provide an empirical sample of debris trajectories and conditional expected losses.
Step 3 — Risk quantification. The mean, VaR and CVaR are computed from the maximum annual loss along each path. This transformation is distinct from estimating the distribution of losses at a fixed calendar year.
Step 4 — Premium calculation. The stress premium is the mean loss plus 1.5 × max(CVaR95 − EL, 0), 0.04 × max(VaR99 − EL, 0), and 0.15 × EL. These coefficients are pricing assumptions, not regulatory requirements.
Limitations
Single-shell approximation. The model represents only the 800 km shell. In reality, debris migrates across altitudes due to atmospheric drag and collision-induced velocity changes. A multi-shell model would capture cross-altitude contamination effects.
Population assumptions. The satellite count is fixed at 8,000; debris count evolves. The initial 36,500 objects represent a catalogue-scale count assigned to one shell rather than a measured shell-specific population. Launch schedules, retirement and cross-shell movement are not resolved.
No active debris removal. The model does not include ADR missions, which would slow the debris growth.
Simplified fragment physics. The fragment yield is an event-anchored log-normal prior, not a formal calibration. Collision mass and energy do not enter the sampler.
Catalogue-scale count with a generous debris–debris term. The model tracks objects larger than 10 cm but adds the ~3 fragments from each debris–debris event to that same count, and both channels share the 10 m² satellite cross-section. Excluding those fragments lowers N(50) from 1.51M to 758,000 and the expected loss from $921.4M to $462.6M; a 1 m² debris–debris section gives 793,000 and $484.1M. These sensitivity results show that the headline ratio depends strongly on the debris–debris assumptions; they do not establish a universal upper bound on risk.
Additional interpretation limits
Claim-count uncertainty. The headline loss mapping uses expected claim counts. It does not sample a full realised claims distribution. A separate dependence model is not an input to the headline premium.
Horizon sensitivity. Quadratic feedback can produce rapid growth or finite-time divergence in the mean-field approximation. Late-horizon results require sensitivity and numerical convergence checks, not extrapolation as stable rates.
Benchmark and coverage. The 0.5% benchmark is assumed. Changing it changes the reported ratio. A benchmark for general in-orbit failures and a model restricted to debris collisions also differ in coverage.
Validation status. The analysis has not been peer reviewed. Independent replication, empirical calibration and uncertainty quantification remain necessary before operational or underwriting use.
Further work
Multi-shell model with cross-altitude migration; NORAD TLE data integration for empirical debris tracking; dynamic constellation deployment modeling; and reinsurance layer pricing (excess-of-loss, stop-loss structures).