From the archive · field 2020

The algorithm found three people we had already met

We ran fifteen in-depth interviews and two hundred questionnaires on the Brazilian crypto market. The interviews found a cautious newcomer. Weeks later, clustering on the survey data returned a 37% group that was exactly him. The convergence was the finding. So was the shared blind spot.

The client wanted to enter the Brazilian crypto market and needed to know who they would be talking to. The question was commercial and blunt: who is this investor, what do they want from a platform, and what would make them leave the one they already use.

The design we proposed had three legs that rarely travel together. Fifteen in-depth interviews — ten investors and five digital influencers. A survey of two hundred investors. And a collection of public data: 5,029 videos on the subject across 280 YouTube channels, and the 330,794 comments left on them up to 11 September 2020.

The person who moderated the interviews and the person who ran the notebooks were on the same project. That is what made the main finding possible — and it was not about crypto.

Who was buying, and the surprise that they were not outsiders

The quantitative portrait was sharp: 88% men, 70% aged 35 or under, education above the Brazilian average, and a little over half earning up to R$5,724 a month. A fifth worked in IT, development or software.

The reasonable expectation in 2020 was that this audience sat outside the traditional financial system — that was the category's own narrative. The data said otherwise. Among crypto investors, 84% also held a savings account, against 31% of the traditional Brazilian investor according to 2019 market data. In equities, 22% against 3%. Only 2% held no traditional investment at all. And only 4% had no digital bank account.

These were not people who had left the system. They were people who had added a layer on top of it.

What the interviews found

The qualitative material described a market split between veterans and a recent wave. When we asked people to describe the typical crypto investor, the answer changed depending on who was answering — and the variation was the finding.

“More libertarian. People who want to get rid of the government, get rid of all of it, and protect themselves.”

Investor, 33

“Aggressive. They want to take the risk, see if they make a gain, they come in for a high return.”

Investor, 28

“This is someone with a bold investment profile. Not the kind who puts money in a savings account.”

Influencer, 48

“People interested in the technology, or in the political and social ideals behind cryptocurrencies.”

Influencer, 22

Three figures took shape as we read: the libertarian who came in on conviction, the aggressive one who came in for the return, and a third character who appeared mostly in what people said about others — the insecure newcomer, arrived within the last two years, researching on Google, choosing a platform on what friends said.

That third one was the most commercially interesting and the weakest as a finding. He came from indirect description, not self-description. He is exactly the kind of thing a qualitative report asserts with confidence and nobody can check.

And what the algorithm found

The survey was analysed with k-means clustering on five variables: semantic identification (cautious–aggressive, analyst–explorer, strategist–conqueror), familiarity with market concepts, age, education and individual income.

Three groups came out, and the split was close to even: 37%, 32% and 31%.

The 37% group gathered the investors with the least time in the market — two and a half years on average — the lowest income and the lowest education of the three, sitting at the cautious pole of the semantic identification. They wanted deposits from minimum amounts, zero fees on small volumes, support in Portuguese, and explanatory videos on the platform. They had chosen their exchange by searching Google and asking people they knew.

It was the insecure newcomer from the interviews. Not an approximation: the same character, with measured income, measured time in market, and measured size.

That is what two methods are for. The qualitative work produced the hypothesis and the language — cautious, searches Google, asks a friend. The quantitative work said how many they were and what they earned. Neither alone would have delivered both. And if the clustering had returned groups the interviews did not recognise, that would have been information too: it would have meant one of the two readings was wrong, and it would have been worth finding out which one before delivery.

The third finding came from the public data and could not have come from anywhere else: of the 280 channels discussing the subject, 7% accounted for 91% of everything published. The conversation looked broad and was concentrated.

What the 2020 photograph was already showing

A study is a photograph of a moment. The final report was delivered on 4 November 2020. Pix, Brazil's instant payment system, went live twelve days later. The photograph caught Brazilian financial behaviour in the last weeks before it changed register.

Two things in that photograph aged well.

The first is that 33% of investors named “an exchange regulated within Brazilian legal requirements” as a desired attribute of an ideal platform. An unregulated market asking to be regulated. In December 2022 Brazil passed Law 14.478, the legal framework for virtual assets; a 2023 decree gave the Central Bank authority to regulate and license providers; in November 2025 the resolutions creating the SPSAV regime followed.

The second is more prosaic and was the most useful to the client at the time: 72% wanted to be able to deposit minimum amounts, and 55% wanted support in Portuguese. Read alongside the median income of the sample, those two lines are not preferences — they are the specification of a small-ticket market that has to be served in the local language. Anyone entering here designing for the large investor would be designing for the minority.

And where it was wrong

Hindsight is generous, so it is worth naming the error, which was ours and belonged to both methods at once.

We asked the main reason for investing in crypto. Long-term gain, 38%. Less bureaucratic than the banking system, 22%. Short-term gain, 19%. To use as a currency for exchange and payment, 12%. Recommendation from friends, 8%.

We treated the 12% as what it looked like: the fourth reason out of six, a minority. The public data agreed and reinforced it. Across 58,695 records of a cryptocurrency being mentioned in comments, Bitcoin was 86%. Stablecoin appeared in 0.5%. Tether, in 0.1%.

In February 2025, the president of Brazil's Central Bank stated publicly that around 90% of crypto volume moved in Brazil was in stablecoins. Federal tax authority data for the first half of 2025 pointed to R$227 billion in transactions, with USDT accounting for roughly two thirds.

The 12% minority was the market. The 0.5% was the vector.

There is a second error, smaller and funnier: the second most mentioned coin in the comments, at 5%, was a project that features in no market discussion today. Ethereum appeared at 2%. Social media attention is not prediction — it is attention. We measured it well and read it as if the two were the same thing.

What we learned and have used since: in a fast-growing market, the small line in the table deserves an explicit question, not a rounding. Not because every minority becomes a majority — most do not — but because the cost of asking is one more question on the questionnaire, and the cost of not asking is this piece.

Why this matters if you commission research in Brazil

The claim most suppliers make is that they do qualitative and quantitative. Almost always that means two teams, two reports, and a conclusions chapter stitching them together at the end.

What happened on this project differed in one detail that changes the outcome: the qualitative hypothesis was tested against the data before delivery, because the same two people were on both sides. The cautious newcomer stopped being a moderator's impression and became 37% of the sample, with an income band and a time in market. Had the numbers disagreed, we would have known in time to change the reading instead of defending it.

The point is not that the algorithm validated the anthropologist. It is that both were at the same table, which is what made it possible to ask one about the other.