A captive is an insurance company owned by the group it insures, and its existence answers a simple idea: beyond a certain volume, a group pays the market more in premium than it receives in claims, and it can keep that differential by carrying frequency itself. The reasoning is sound on numerous, repetitive risks, motor fleets, employers' liability, ordinary damage, because the law of large numbers operates inside the group as it would at an insurer. On political risk it does not operate, and from that comes the whole difficulty of this module.
One must see precisely why it does not, because the error comes not from ignorance of statistics but from a badly applied analogy. A group holding assets in eight countries believes it holds eight risks, and it holds one, declined eight times. Hostile measures spread by political imitation, by financial contagion, by membership of a single regional bloc, and a wave of nationalisations in a sector rarely touches one state alone. Frequency is low, severity is high, and above all the events are correlated at the very moment they occur. A captive carrying those risks does not pool, it concentrates, and its result follows the same distribution as the group's.
That correlation has a name in reinsurance, and the mechanism it designates deserves being understood rather than cited. What correlates in the tail of a distribution is not what correlates in ordinary conditions: two countries whose economies have nothing in common in normal times become joined when one political movement crosses the region, or when one commodity price collapses and puts their budgets under strain in the same quarter. A captive that calibrated its retention on correlations measured in ordinary conditions discovers at the first serious event that its eight lines trigger together, which its model judged all but impossible.
One must then say what a captive actually serves on this subject, because it is not useless, it is useful otherwise. It serves to fund a retention the group has decided to keep, smoothing its cost over time rather than suffering it in the year it falls. It serves to access the reinsurance market, often broader and cheaper than the direct market, which is frequently its true reason for being. It serves to unify cover terms across subsidiaries that each bought their own, and to surface at group level an exposure nobody saw consolidated. None of those uses requires carrying severity; all of them require instead knowing where severity stops.
A second rule meets the first here and must be stated, that of capital. A captive is a real insurance company, subject to a solvency requirement in its territory of establishment, and that requirement is computed on the distribution of losses it carries. A captive whose commitments are strongly correlated must immobilise more capital than a captive of the same volume carrying independent risks, which is precisely the prudential sanction for concentration. The group that thought it was saving an insurer's margin discovers it is funding regulatory capital, and the apparent saving shrinks accordingly.
There remains the question that decides in practice, that of the level of retention, and it is not resolved by comparing a premium with an expected loss. On a low frequency, high severity distribution, expected loss is a poor guide: it describes an average the group will never experience, since it will suffer neither an eighth of a loss each year nor the mean of a population it does not constitute. The useful test lies elsewhere and it is one of liquidity: what simultaneous loss across several lines can the group absorb without breaching a banking covenant, without abandoning an investment under way, without downgrading its rating? Retention is set below that threshold, and what exceeds it is transferred, whatever its apparent price.
One last error deserves naming because it is frequent and invisible. A group has its captive carry a modest retention on each country, say five million, judging each line reasonable taken alone. Eight lines at five million make forty million of simultaneous exposure, and that number appears nowhere if the analysis was done line by line, which is the usual way of doing it. Consolidating retentions is a three minute calculation few groups keep current, and it is the only one that says what is actually carried.
What remains to be done fits in four lines. Consolidate retentions before judging them, adding the lines rather than appraising them separately. Calibrate on a simultaneity scenario rather than on an expected loss. Remember that what a captive brings here is access to reinsurance, smoothing and consolidated visibility, not pooling. And accept that a low frequency, high severity risk is precisely the one worth buying outside, which is the reverse of the intuition that created the captive.
A European mining group holds assets in three countries of one southern African region: a copper mine valued at 210 million euros, a cobalt mine at 160 million, and a processing plant at 90 million. Its captive, established in Dublin with 55 million of own funds, carries a 12 million retention per country per event, the balance being reinsured. The risk committee approved each line separately: 12 million against 210, against 160 and against 90 were judged modest. In 2025 the same party wins elections in two of the three countries on a platform of renegotiating mining concessions, and the copper price has fallen 31 per cent in eighteen months. The finance director asks whether the structure holds. What should he be told?
The answer begins with a calculation nobody made, and that is already the diagnosis. The committee approved three 12 million lines taken separately; the real exposure is 36 million if all three trigger together, against 55 million of own funds, that is 65 per cent of the captive's capital on one regional event. Line by line analysis is what made that figure invisible, and it is the method rather than the level of retention that is at fault. Then comes the question of simultaneity, and the facts of 2025 settle it the wrong way. Three countries in one region, one political movement winning two elections on an identical platform, and a copper price collapse putting public budgets under strain in the same quarter: these are exactly the channels through which exposures independent in ordinary conditions become joined in the tail. The model that approved 12 million per country very probably used correlations measured over calm periods, which is the usual and wrong way to measure them. Three things must therefore be said to the finance director, in this order. The structure does not hold as it stands, not because the retentions are too high in themselves, but because they are not consolidated and the relevant threshold is what the group can absorb on one event, not per country. Retention must then be reset by a liquidity test rather than by an expected loss: what simultaneous loss does the group absorb without breaching a banking covenant, knowing that the copper price has already reduced its capacity, which worsens both terms at once. Finally, the captive keeps its full usefulness here, but not the one attributed to it: it gives access to reinsurance, it smooths, and it makes consolidated exposure visible. What it cannot do is pool three declensions of one risk, and it is precisely correlated severity that is worth transferring outside.
- 01A group present in eight countries does not hold eight risks but one declined eight times: the law of large numbers does not operate and the captive concentrates instead of pooling.
- 02What correlates in the tail is not what correlates in ordinary conditions: political imitation, financial contagion, regional membership, one commodity price.
- 03A strongly correlated captive immobilises more regulatory capital at equal volume: the saved insurer's margin is paid for in capital.
- 04Expected loss is a poor guide at low frequency: the test is one of liquidity, the simultaneous loss absorbable without breaching a covenant.
- 05Consolidating retentions is a three minute calculation nobody keeps current, and it is the only one that says what is actually carried.