Brazil’s Pix payments fell below forecasts after a betting ban—but the size depends on the model

A payment-system question, not a household balance sheet

According to the paper, Brazil prohibited fixed-odds betting nationwide on September 25, 2026, covering sporting events and online games while preserving other legally authorized lottery categories. The measure included a transition for operators to close and issue refunds.

The authors place the policy against a background of household debt, costly credit and concerns about gambling-related harm. But those concerns are context, not outcomes measured by this study. Its narrower question is whether the number and total value of Pix settlements departed from what earlier patterns would have predicted.

The data cover daily Pix payments settled through Brazil’s Instant Payments System, known as SPI. These aggregate figures do not identify which transfers involved betting platforms or whose finances changed.

Forecasting an alternative path

The researchers fitted their models to 998 days of observations, from January 1, 2024, through September 24, 2026. They then projected payment activity beyond that cutoff and compared the forecasts with observed settlements.

Their models account for recurring patterns associated with weekdays, days of the month, holidays and annual seasonality. They also allow unexplained fluctuations to be related across nearby days. Transaction counts and monetary values are modeled separately; the monetary amounts are nominal, without an inflation adjustment.

The crucial difference concerns the underlying growth path. The first model, S1, uses a linear trend in logged activity. Two alternatives, S2 and S3, add curvature, allowing that trend to accelerate or slow. S3 also changes how short-term dependence is modeled.

The comparisons cover September 26–28 and September 26–November 1. Publication day itself is excluded because its daily total mixes transactions before and after the policy announcement. The longer, 37-day window includes the closure-and-refund transition.

These are comparisons with an estimated alternative future—not percentage changes from the preceding month or another observed period.

The headline result is model-dependent

Under S1, cumulative transaction counts over 37 days were 8.37% below forecast, equivalent to about 789.6 million fewer transactions than projected. Settlement value was 10.83% below forecast, a gap of approximately 460 billion reais.

The approximate 95% intervals excluded zero: from −13.96% to −4.53% for counts, and from −18.28% to −3.83% for value. This is the specification behind the preprint’s headline statistically significant results.

The alternatives produced substantially smaller gaps. Counts were 4.12% below forecast under S2 and 3.75% below under S3. Both estimated settlement values about 3.68% below forecast, corresponding to roughly 145 billion reais. All four of those 37-day uncertainty intervals included zero.

Every model also placed activity below forecast over the three-day window, but every corresponding interval included zero. Agreement on the direction therefore does not establish a precise or statistically robust effect across specifications.

Why the forecasting checks matter

The authors tested forecasts at historical dates before the ban, using only information available at each cutoff. In the recent validation period, S1 overpredicted 37-day monetary totals by an average of 4.60%. Its average absolute percentage error was 4.64%, compared with about 1.2% for S2 and S3.

That matters because a forecast that is already running too high can make later observations look unusually low without a policy causing the entire difference. However, S1 performed better in an earlier validation block, so the curved-trend alternatives were not consistently superior throughout the sample.

Other limitations remain. No 37-day historical validation was performed for transaction counts. The longer validation windows overlap, and the model comparison was retrospective. Tests also found remaining time dependence in some models’ unexplained errors, particularly S1.

The researchers estimated uncertainty by repeatedly resampling blocks of model innovations and refitting the models. That procedure cannot repair an unstable growth trend or account for an omitted simultaneous shock, and the validation did not establish how reliably the intervals capture outcomes.

What the shortfall cannot tell us

A causal interpretation requires the projected path to remain credible, no relevant anticipation of the ban, and no concurrent shocks explaining the departure.

Most importantly, a settlement gap is not a measure of gambling losses avoided, household savings gained or consumption lost. The aggregates cannot separate betting payments, closure-related refunds and other transfers. The preprint documents a model-sensitive departure from forecast; determining its causes and household consequences requires more detailed transaction information.