Data as argument (vs ornament)

Essay on the difference between the number that supports falsifiable proposition and the number that serves only as ornament in institutional reports, written partly as reaction to corporate presentations that deliberately conflate the two, and partly as consideration of why the confusion is so resistant.

In May 2024, the Guaíba overflowed. The flood that followed was the largest in Rio Grande do Sul in two decades, and lasted weeks. Five months later, when SUSEP consolidated the annual aggregate of line 196 (Named & Operational Risks), the Rio Grande do Sul loss ratio was 4.27. Four reais and twenty-seven cents paid in claims for every real received in premium.

The number sits in a public table: SES SUSEP, line_state, extraction 2025-04-20. Any reader with portal access and twenty minutes of spreadsheet work reaches it exactly. What varies, depending on who cites, is the treatment around it.

In a rushed analysis, this 4.27 becomes "historic losses that demonstrate the resilience of the insurance market". The statement is not false, it is merely elastic, in that "resilience" is a word that accommodates almost any reality. In an auditable analysis, the same 4.27 appears with declared source, shown derivation, explicit runoff caveat, and becomes falsifiable, in the sense that any reader reproduces the number, or refutes it. The difference seems subtle on first reading. In the long run, it defines whether the research ages as citable corpus or as dead archive.

There are two ways to use a number in a public analysis, and the distinction between them is less obvious than it seems to those who grew up reading annual reports of publicly traded companies. One way treats data as evidence, declares source, shows derivation, qualifies the known caveat, and loses value if the proposition falls. The other treats data as ornament, contextualizes little, can rarely be reproduced, and survives intact even when the proposition it supported changes.

The operationalization of the difference is simple. If I say that the loss ratio of the auto line (531) in Santa Catarina, in 2024, was 71%, there is source, slice, formula and caveat behind it, and any reader with access to the same data reaches exactly the same number, or refutes it.

If an institutional presentation says "the sector grew solidly and consistently over the last five years", the statement is not falsifiable, given that "solidly and consistently" is broad enough to accommodate almost any reality. The 4.27 of line 196, by virtue of its magnitude and context, exposes the difference between the two registers with uncommon clarity.

The initial thesis: rigor is expensive, narrative suffices

The dominant practice of Brazilian economic journalism, and of much sectoral consulting, operates under the premise that the average reader has no patience for methodological detail, and that the analyst's function is to translate the number into convincing narrative, not into reproducible proof. That premise is not entirely false. The reader who opens a company's annual report, or a newspaper's economics section, rarely wants to walk through a methodological appendix before forming opinion. Loose narrative meets that demand, and meets it with considerable operational efficiency, given that the marginal cost of producing generic statement is much lower than the cost of producing auditable statement.

The counter-thesis, however, is that loose narrative holds up only as long as no one tests. The decoratively generated number, even if it serves the immediate purpose of illustrating the intended thesis, loses value the instant a critical reader requests the primary reference, or another analyst tries to reproduce the calculation, or time exposes that the alleged "consistency" was a convenient selection of time window. And it loses value cumulatively, given that each ornamental citation diminishes the credibility of the whole a little more.

How to recognize ornamentation

Three signals appear together, with frequency high enough to serve as a heuristic.

The first is ambiguous source citation. "Source: SUSEP" without year, without table, without extraction date, is ornamental citation, given that it doesn't allow reproduction or audit. The effect is to give appearance of rigor without the burden of having it. Compare with "Source: SUSEP/SES, table line_state, extraction 2026-04-20, direct premium net of cancellations", and this second form is real citation, in the sense that a third party reaches exactly the same number starting from it.

The second is the absence of unit or denominator. "Brazil has 75% penetration in product X" is a sentence almost without content until knowing: penetration measured over what population, what age slice, in what year, against what international reference. Without explicit denominator, the number becomes sentiment, and works rhetorically but not technically.

Last, and perhaps most common in corporate presentations, is the visual scale chosen to impress and not to clarify. Bar chart with Y-axis starting at 70 instead of 0, so the difference between 75 and 80 looks like multiplication. Pie chart with nine slices when a table would explain better. Map colored in divergent palette when the metric is monotonic. Each of these choices is deliberate, rarely neutral, and almost always serves the narrative instead of the data.

The pivot question

It is debatable whether the problem is as simple as moral exposure. Whoever today cites "source: SUSEP" without detail is doing science or ornamentation? At what point did rigorous citation become exception instead of norm? Does the average reader know how to distinguish, or has been trained to accept appearance as sufficient? Is there any practical incentive for the producer of the analysis to opt for the auditable form, given the operational economy of the ornamental form? And, perhaps the most uncomfortable of all, to what extent am I myself, in writing this text, making argument or merely decorating a position I already had before starting?

In my view, the reasonable answer recognizes three cumulative pressures that sustain the predominance of the decorative form, and none of them is character flaw of those who produce sectoral reports. The first is time. Treating data as argument costs between five and ten times more per number than treating it as ornament, given that it requires primary source verification, validation against cross-reference, documented derivation, explicit qualification of known caveat, and construction of the reproduction mechanism. Whoever has a one-week deadline to deliver a fifty-page report optimizes for ornament-numbers, which deliver in minutes.

The second is the client. Paid consulting clients rarely want to hear "evidence is ambiguous" or "the standard sector metric aggregates things that shouldn't be aggregated". They want actionable conclusion. There is demand for the appearance of certainty where evidence supports only probable direction, and ornamentation delivers the appearance without the burden of certainty.

The third, perhaps the most determining, is absence of incentive against error. In a closed report, pipeline error isn't discovered by third parties, given that no one replicates. The analyst who misquotes isn't confronted. The decorative number survives because the ecosystem has no mechanism to punish it. In open environment, with public pipeline and data under permissive license, any error is potentially identifiable by critical reader, which deeply changes the risk calculation.

The third way: argument as infrastructure, not morality

In broad terms, perhaps it would be worth shifting the debate from moral dimension to operational dimension. Treating data as argument is not personal virtue of those who prefer "doing it right", it is choice of editorial model that produces measurable consequences. The most immediate consequence is that much analysis becomes infeasible the way it is done today, given that you can't publish sectoral assertion without citing table, date, slice. The less obvious consequence, and in my reading more relevant in the long run, is that analysis treated as argument ages better. The decorative number from 2018 is embarrassment in 2026, given that no one can replicate or understand where it came from. The argumentative number from 2018 is still defensible in 2026, or was corrected by explicit errata, and in either case preserves the continuity of the research.

There is also a political consequence, even if less commented, which is well-made argument being hard to capture. Decorative report can be commissioned to defend prior client position, and frequently is. Argument structured around public data, with open pipeline, declared caveat and pre-committed errata, is in the long run much more resistant to capture, given that the deviation becomes visible. Not immune, I should note, but resistant, and that difference is larger than it appears when comparing credibility profiles over time.

The concrete case

Panorama SUSEP is the attempt to operationalize this difference in a niche historically dominated by ornamentation. Every number visualized in the study carries, in the footer, the primary source with table and extraction date, the derivation formula if any, the pipeline versioned by commit, the license under which the data is redistributed, and the known caveats of the specific slice. The methodology page explains every aggregation decision. The errata page exists since before the first correction, as public commitment.

None of this guarantees absence of error. Guarantees only that error is detectable, correctable, and historically registered. For an area that produces regulatory decision, actuarial allocation and underwriting strategy, that difference is less cosmetic than it appears at first sight. Argument accumulates into usable corpus. Ornamentation accumulates into dead archive, and time treats each accordingly.

— Natan, Apr/2026