Why I published Panorama SUSEP

Personal account of the road to Panorama SUSEP v1.0, written alongside the study itself, motivated by an attempt to explain to a future reader the calculation between publishing and not publishing, and why the first option, despite its evident cost, still seems more defensible.

When one decides to publish research in open environment, instead of keeping it in closed consulting report or in restricted institutional archive, the first reaction expected from outside observers is to classify the decision as naivety or idealism, and that classification is not entirely unfair, given the set of real costs that publication entails and that are rarely compensated by proportional financial return. Nonetheless, in light of the initial discomfort of assuming that cost, there is a series of consequences that make the closed alternative less attractive than it first appears, and it is precisely on this consideration that I intend to dwell in this text.

The Brazilian insurance market was, for years, an environment I knew existed but couldn't see clearly. I knew SUSEP published data, that CNseg compiled annual sectoral reports, that international consultancies sold paid and relatively expensive cuts, but the first time I tried to draw a macro picture of the market by federal state I noticed that nowhere accessible and free was there a structured view that could serve as a starting point for a curious reader, an economic journalist, an actuary at another company, or a regulator wanting to compare their region to the national average. The gap is strange, given that the data is there, public, on the SUSEP portal, in SQLite and CSV, periodically updated, with a license that permits redistribution, and with widely available processing tools. What was missing was someone with hobby time, minimal technical insurance vocabulary, and willingness to build the editorial infrastructure around the numbers, so that they would stop being inert tables and become auditable narrative.

The initial thesis: closed report is the safe path

The dominant practice of the sector, even if not declared explicitly, operates under the premise that serious research justifies itself through closed report. The sixty-page PDF, with executive summary, static charts pasted as images, and generic source citation, is the expected form. Those who produce this research argue, with partial reason, that editorial control is greater, that the client pays for the material and is entitled to exclusivity, that public errata would be unnecessary erosion, and that methodological detail only confuses the non-technical reader. This set of arguments sustains an entire sectoral consulting industry, and it would be intellectually dishonest to dismiss it outright.

The counter-thesis, perhaps less obvious, is that each of these justifications operates against the possibility of the research serving as citable evidence over the long term. Closed reports are not replicable, given that the methodology is rarely sufficiently exposed. Exclusive client is limited audience, given that the publication does not circulate beyond the contract. Public errata would only be erosion if the error were treated as shame, instead of as part of the process. And methodological detail only confuses the non-technical reader if the editorial structure does not know how to separate main narrative from auditable provenance.

The antithesis spelled out: the real cost of publishing

In light of the appeal of the counter-thesis, it is necessary to honestly recognize the cost of publishing. Panorama SUSEP first ran in three afternoons, in a naive pipeline that aggregated whatever SUSEP sent, summed without warning about runoff, and produced a total direct premium number that matched what CNseg reports. I thought it was ready. It wasn't. The following weeks were successive cascading technical audits that revealed six structural problems, each capable, alone, of invalidating an entire analysis:

1. Restatement: SUSEP republishes prior periods when it discovers error, and the pipeline needed to distinguish original from restated to avoid duplication.
2. Runoff: companies that stopped operating continue reporting old claims for years, and adding that to current-year premium creates an artificially inflated loss ratio.
3. CR4 and HHI by economic group: measuring market concentration by CNPJ ignores that three of the four largest belong to the same conglomerate, distorting the competitive read.
4. Float guard: floating-point arithmetic accumulated enough error to break the identity Σ_state = national_total in some lines, requiring integer aggregation and controlled rounding.
5. Outlier line 196 RS 2024: the catastrophic flooding in Rio Grande do Sul produced a loss ratio of 4.27 in a single cut, a true value but one that wrecks any visual scale without explicit treatment.
6. Sanity against CNseg: the total number had to match the official report, not as proof of correctness but as a plausibility ceiling. When it diverges much, there is pipeline error before any analytical conclusion.

Each of these fixes was documented, versioned, and is reproducible via make reproduce in the panorama-susep repository. The reader who doubts any number can download raw data, run the pipeline, and arrive at exactly the same JSONs that feed the site. That is the contract. And the cost of arriving at that contract was indeed much higher than the cost of producing the equivalent closed report.

The pivot question

It is worth noting that incurring this cost, without proportional financial return, in a project that does not pay the bill next month, raises a series of questions that deserve to be asked out loud before the decision is made. Why insist on method when the dominant practice of the sector is loose narrative? Why expose oneself to public error when the closed report offers the comfort of opacity? Why spend three weeks on naive pipeline that will be redone from scratch? Why register errata before the first error? At what point does methodological paranoia turn into paralysis? And, perhaps the hardest of all, why assume there is a reader interested enough that the auditability effort will have any reader at all?

In my view, the reasonable answer to these questions does not lie in either of the two extreme positions, neither in the orthodoxy of the closed report nor in the idealism of pure open notebook. Perhaps it would be worth recognizing that open publication, when well structured, offers three operational advantages that compensate for the cost over the long horizon, even if not in the short.

The third way: open notebook as infrastructure

First, errors are caught early. In open environment, with public pipeline and data under permissive license, any critical reader can identify inconsistency, and that identification happens before the wrong number becomes the foundation of another analysis of mine six months later. In closed environment, that same error lives hidden until it becomes a serious problem. Second, the final version comes out simpler. When you know the pipeline will be exposed, there is real pressure to cut unnecessary shortcuts, qualify where evidence is weak, and avoid conclusion that wouldn't survive the first critical question. Closed reports accumulate appendix and qualification precisely because no one will check. Third, and perhaps most relevant, the work takes on its own life. An open study is citable, indexable, linkable; a URL with a query string that exactly reproduces the visualized cut is worth, in research, more than ten charts in a closed presentation no one can replicate.

In the end, none of this is original. It is the tradition of open notebook science applied to a niche that historically operated in closed environment, in paid consulting, or in academia that didn't reach the interested public. If the study serves as evidence that this mode is viable for Brazilian sectoral research, the parallel goal is met. The main objective, simpler, is that the next reader who wants to know the loss ratio of the auto line in Santa Catarina in 2024 doesn't need to run a pipeline, hire consulting, or give up. Just needs to open the link.

— Natan, Apr/2026