What the Insper and Stone study on bets and defaults doesn't answer.

The Folha de S.Paulo newspaper published on Tuesday (September 8th) a groundbreaking study that, according to the report, measures the "size of the damage" caused to bettors by online gambling. Conducted by Sérgio Firpo, a professor at Insper, in partnership with researchers from Stone, the study cross-references banking and Pix data from millions of customers to compare those who started betting after January 2025 — the month of the regulation of betting in Brazil — with those who never bet.

The key figures released: the proportion of bettors in default was, on average, 20% higher in the 12 months following their first bet, compared to the increase observed among non-bettors; the fraction of overdue credit card bills grew by 38,7%; the credit limit made available by banks fell by 16%; and one in ten bettors allocates more than 20% of all the money that leaves their account each month to betting.

This is a more sophisticated design than the CNC study released in April, which BNLData has already subjected to methodological criticism—here, there is individual transaction data, not aggregated sample survey data (PEIC), and there is an attempt to isolate those who bet from those who did not at the individual level. This does not mean, however, that the design solves the classic problems of this type of inference. Points that deserve scrutiny before treating the numbers as a definitive measure of the "loss" of bets:

01. The sample is Stone's customer portfolio, not the general population.

Stone disproportionately serves freelancers, micro-entrepreneurs, and small business owners—a segment with irregular cash flow and a credit risk profile distinct from that of a salaried worker or the average consumer. Extrapolating the results to "Brazilian bettors" en bloc requires a connection that the report does not show has been made. Without knowing how the sample compares to the universe of bettors identified by the Central Bank, it is impossible to assess its representativeness.

02. Reverse causality is not ruled out.

The study compares trajectories before and after the first gamble, but it does not rule out the possibility that people already in financial distress are more likely to start gambling—either seeking a "quick fix" for debt or due to greater exposure to advertising for easy credit at the same stage of life. A pre-treatment trend test (parallelism between the groups before the first gamble) would be the minimum standard to support a causal reading, and there is no indication that it has been reported.

03. The cut in the credit limit may be the result of automated scoring, not a real deterioration in payment capacity.

The president of the Central Bank himself, Gabriel Galípolo, has already stated in testimony to the Senate that the history of gambling is included in the credit score used by banks to price risk. If the "gambling" transaction is directly penalized by the algorithm, the 16% cut in the limit may reflect a regulatory/commercial decision by the bank—not necessarily an objective decrease in the client's ability to pay. The two mechanisms produce the same number, but tell very different stories.

04. Gross stake sent to the betting site is not the same as net loss.

If the metric captures values ​​sent via Pix to betting platforms, it measures deposits, not losses. Without discounting withdrawals and prizes received back, treating "money that left the account for the bet" as equivalent to "money lost" overestimates the effect — a recurring problem in studies that use volume (handle) as a proxy for harm.

January 05, 2025 marks a milestone for bets, but also for other factors.

The same period is marked by rising credit card interest rates, expansion of revolving credit, and increased digital banking—factors that can explain part of the increase in defaults independently of betting. This is the same weakness already pointed out by BNLData in the CNC study. The robustness of the comparison depends on the control group (those who do not bet) having an income and occupation profile equivalent to the group studied—information that the report does not detail.

06. Proprietary and non-replicable database

The microdata belongs to Stone and is not public. Without access to the technical note, the paper, or the analysis code, it is impossible to audit robustness tests, model specifications, or outlier handling. The reader—and the press—are entirely dependent on the presentation that the authors themselves chose to publish.

07. The average may be being pulled down by a tail of heavy bettors.

The study itself highlights that one in ten bettors spends more than 20% of their bank account balance on bets. This suggests that the average effects may primarily reflect this extreme subgroup, and not the behavior of the typical bettor. Reporting the median, in addition to the average, would help clarify whether the "loss" is widespread or concentrated.

Online gambling increases default rates by 20% and reduces credit by 16%.