The loss record came first. Every time.

Every asset class that ended up with real risk infrastructure got there in the same order, and the order is the lesson. The loss record came first. The models came second, and they became infrastructure only because they had been tested against the record. The instruments came third, and they exist because an independent party publishes a number that a contract can be written against. Carbon is trying to do this backwards, and it is not the first market to try.

Andrew, and what the industry discovered about itself

Hurricane Andrew made landfall in Florida on 24 August 1992. It produced roughly 15.5 billion dollars of insured losses in 1992 money, around 25 billion in later terms, and it took a number of insurers with it. The Insurance Information Institute records at least eleven insurer insolvencies attributed to the storm, with industry counts running higher across 1992 and 1993.

The uncomfortable finding was not that the storm was large. It was that the industry did not know what it was exposed to. Carriers had priced Florida property on historical loss experience, actuarial projection and underwriting judgement. Those are respectable tools and they were not wrong so much as blind: they described a past that did not contain the event, and nobody had assembled a view of what the portfolio would do if the event arrived.

What actually got built, and in what order

The response is usually remembered as "the industry adopted catastrophe models". That skips the part that mattered.

The first thing built was loss data. Claims experience organised into actuarial tables. Loss development triangles, which record how an accident year's losses mature as claims are reported and settled, and which are the reason an actuary can reserve for claims that have happened but have not yet arrived. Industry loss aggregation services that publish an estimate of insured loss from a named event, creating an objective reference point that two parties who disagree about everything else can both point at.

Only then did the models become credible. Vendors built stochastic event sets, hazard modules, vulnerability functions and financial modules; a carrier feeds a portfolio in and receives an exceedance probability curve out. By the mid 1990s this was standard practice. Florida created a catastrophe fund whose structure was set by modelled results, and in 1995 established a commission specifically to review hurricane loss models against standards.

The sequence is the point. The record forced the collection. The collection made the models testable. Testability made them a shared language. And the shared language is what the market actually bought.

The same story, three more times

Credit. Rating agencies did not become infrastructure by issuing letters. They became infrastructure by publishing default studies: how often did a bond of this rating actually default, over what horizon, with what recovery. The study is what made the letter mean something, and the entity that published the study defined the vocabulary everyone else had to speak. Ratings are now embedded in regulation, investment mandates, collateral eligibility and capital requirements. That embedding, not the opinion, is the moat.

Structured credit. Loan level disclosure, agency pool data, cash flow modelling and performance databases spanning multiple cycles. The 2008 crisis was not primarily a failure of data availability. It was a failure of labels covering the relevant regime, plus a correlation assumption that had never been tested against a national downturn because the sample did not contain one. Two lessons travel: aggregated data hides correlated risk, and a model calibrated on a benign period will not survive the regime it was never shown.

Commodities. Assay and inspection by independent parties. Certification against defined specifications. Warehouse warrants issued by approved warehouses, which are the document of title that makes a metal contract deliverable. Every physical market converged on buyer paid or exchange paid inspection, for one reason: a buyer will not pay for a cargo on the seller's word, and seller paid inspection is not credible at size.

Where carbon actually is

Carbon has the third layer and not the first. It has ratings, which are opinions issued before the fact. It has monitoring, increasingly good, which answers what is happening now. It has insurance, written since 2022, priced by underwriters working from engineering first principles because there is nothing else to work from. What it does not have is any record of what happened afterwards.

The consequence is precise and it is not rhetorical. No carbon rating product has ever been publicly backtested against realised outcomes, because the outcomes have never been collected. Nobody can state the loss rate for credits of a given type and vintage over a given horizon and defend it. A carbon underwriter today is in the position of a property underwriter in 1991: real exposure, real premium, and no event set.

The bar is not ours to set, and someone else has already set it. The United Nations Development Programme's guidance on designing insurance for nature states that a parametric product requires ten to fifteen years of data from independent third party sources. Nothing in carbon credit performance clears that bar, because the series has never been started.

Four lessons that transfer without modification

The record precedes the models. Andrew forced the data collection that made the models credible, not the other way round. A market that builds models first ends up with a disagreement it cannot settle.

The models became infrastructure because they were validated. Nobody argues catastrophe models are right. They argue about which one to use, and they all pay. That argument is only possible because there is something to test against.

The independent number is what enables the instrument. Catastrophe bonds and industry loss warranties exist because a third party publishes an objective figure that a contract can trigger on. The trigger is the whole point, and it cannot be published by a party to the trade.

The winners assembled the data and never stopped. Not the best science initially. The ones who built the series and maintained it, then let everyone else calibrate against it.

What this implies for anyone carrying carbon today

If the pattern holds, the sequence in carbon is already legible. The record gets built and published. Models get calibrated against it and become arguable, which is an improvement on unarguable. Then the instruments arrive: cover priced on evidence, collateral valued independently, and eventually contracts that pay on a monitored, independently determined outcome.

The uncomfortable implication for a holder is that the first stage has not happened yet, which means every position currently on a balance sheet was priced without it. That is not an accusation. It is the same position every property underwriter was in before 1992, and the same position every bond investor was in before default studies. The difference is that this time the events are already public, already dated and already documented. They are simply scattered across registries, gazettes and court records in different jurisdictions and formats, and nobody has assembled them.

That is the work. It is unglamorous, it is mostly reconciliation, and it is the thing every market in this list eventually discovered it could not price without.