Do Pix Transactions Improve Monthly Nowcasts of Brazilian Economic Activity? Evidence That Seasonal Control Absorbs the Signal

Abstract

This paper investigates whether transaction data from Pix, Brazil’s instant payment system, improve the monthly nowcasting of Brazilian economic activity, measured by the Central Bank Economic Activity Index (IBC-Br). Pix data are released within days of each month’s end, well before the IBC-Br, which gives them a potential timing advantage. Using monthly Central Bank data from January 2021 to May 2026, we estimate first-difference predictive regressions relating the IBC-Br to its own lagged change, the Selic policy rate, and Pix transaction growth, under four seasonal treatments: no seasonal controls, quarter-of-year dummies, month-of-year dummies, and the officially seasonally adjusted IBC-Br. Out-of-sample accuracy is evaluated with expanding windows of three minimum lengths, Diebold–Mariano and Clark–West tests, and bootstrap critical values that reproduce the small-sample estimation noise of the early windows. Pix growth is strongly associated with the IBC-Br, and improves forecasts, only in specifications that leave month-of-year seasonality uncontrolled. Once monthly seasonality is controlled, Pix adds no forecasting gain and its coefficient is small and statistically insignificant, while the seasonally controlled models forecast the IBC-Br considerably more accurately than any specification without them. With the officially adjusted series, the association has the opposite sign to an activity signal. A recursive model-selection rule never selects a specification with Pix, and the model-averaged effect is economically small. These conclusions are robust to longer estimation windows, to controls external to the IBC-Br, and to simulated data revisions. Aggregate Pix volume is therefore not, at the current sample length, a useful monthly nowcasting input once its calendar component is removed. The paper also documents an aggregation pitfall in the Pix open-data interface and provides replication code.

1. Introduction

The availability of official economic data in Brazil follows, as in most countries, a quarterly cadence (Gross Domestic Product, IBGE) or a monthly cadence with a lag of several weeks (Central Bank Economic Activity Index, IBC-Br). This lag creates an information gap that is relevant not only to monetary policymakers, risk managers, and market analysts, but also to a broader community of business decision-makers who must act on incomplete information about the current state of the economy, whether on the Selic rate trajectory, on portfolio allocation, on credit policy, on inventory planning, or on budget cycles. The nowcasting literature developed precisely to fill this gap, combining indicators available before the target variable’s own release to generate real-time estimates of variables that will only be officially disclosed weeks or months later (Giannone et al., 2008).

In the Brazilian case, this information gap has additional relevance because of the inflation-targeting regime operated by the Central Bank, under which decisions about the Selic rate are taken at periodic meetings of the Monetary Policy Committee (Copom) based, among other inputs, on assessments of the output gap and the intensity of current economic activity. Indicators that anticipate the IBC-Br, the official monthly GDP proxy still released with a lag of about forty days, have direct value for the monetary policy decision-making process, in addition to their value for market analysts and risk managers in the private sector.

Payment systems are among the promising sources of such indicators, since they record in real time the flow of economic transactions underlying real activity. Recent studies document forecasting-accuracy gains from incorporating payment data into nowcasting models in Norway (Aastveit et al., 2024), Switzerland (Felber & Beyeler, 2023), and Italy (Aprigliano et al., 2019), among other countries, typically using debit-card data or traditional bank-clearing system data.

Brazil offers a relevant case because of Pix, the instant payment system launched by the Central Bank of Brazil in November 2020, which within a few years came to be used more widely than debit and credit cards combined, measured by transaction volume. Unlike the traditional payment systems explored in the international literature reviewed above, Pix has a distinctive feature relevant to nowcasting: because it is a public system operated directly by the Central Bank, transaction data are released systematically and free of charge through an open-data API, without the commercial or institutional access barriers that typically limit the use of private payment data for academic research purposes in other countries.

Also relevant, the BCB itself began, in October 2025, to release monthly Payments-by-Economic-Activity Statistics (Estatísticas de Pagamentos por Atividade Econômica, EPAE), a series that disaggregates Pix payment flows across 23 sectors of the economy (Central Bank of Brazil, 2025). In the technical box introducing this statistic, the BCB argues that variation in payment flows across economic sectors, particularly via Pix, constitutes a potential and complementary indicator of economic activity, disclosed with greater speed than the quarterly GDP (Central Bank of Brazil, 2025, p. 48). This recent institutional recognition directly motivates this paper’s research question: to what extent does this institutional intuition hold up under a formal statistical nowcasting exercise?

Before stating the paper’s objective, two terminological points raised by the empirical setting deserve clarification, since they bear directly on how the contribution below should be read. First, this paper’s use of the term nowcasting follows Giannone et al. (2008) in its general sense, forecasting the present or the very recent past of a variable before its own official release, using whatever information is available before that release date, rather than in the narrower sense of a classical mixed-frequency bridge equation, in which a higher-frequency indicator is temporally aggregated to forecast a lower-frequency target (for example, monthly indicators bridging to quarterly GDP, as in Baffigi et al. (2004)). Both the IBC-Br and the Pix series used in this paper are observed at the monthly frequency, so the exercise below is better described as a monthly predictive regression in the tradition of the bridge-equation literature, rather than as a bridge equation in the classical, mixed-frequency sense; Section 3.2 develops this distinction in detail. Second, and consequently, Pix growth is not a high-frequency indicator relative to the IBC-Br in the sampling sense, since both are monthly series; its informational advantage, if any, is a timing advantage rather than a frequency advantage. The BCB discloses aggregate Pix transaction data through its open-data API within days of each month’s end, whereas the IBC-Br for that same month is only released about forty days later. A predictive regression that uses the current month’s Pix growth to help forecast the current month’s IBC-Br, before the latter is disclosed, therefore exploits a publication-lag advantage specific to Pix, not a difference in observation frequency between the two series.

Given this context, the objective of this paper is to test systematically whether the volume of Pix transactions improves the monthly nowcast of the IBC-Br, through a predictive regression that incorporates Pix growth as a regressor available ahead of the IBC-Br’s own release, in addition to the IBC-Br’s own lagged change and the Selic rate, and to submit that result to alternative seasonal treatments, alternative forecast-accuracy tests, and out-of-sample validation with estimation windows of different lengths, rather than stopping at a single in-sample significance test. This paper contributes to the literature by testing a substantive hypothesis about the informational content of Pix, not yet evaluated in the Brazilian literature, in direct dialog with the BCB’s own recent institutional recognition of the potential of the new EPAE statistic as a complementary indicator of economic activity, and by documenting a negative result that is informative beyond this application: the positive association between Pix growth and activity that appears, in-sample and out-of-sample, in specifications without full seasonal control disappears once month-of-year effects are controlled for, either with monthly indicator variables or with the officially seasonally adjusted series, and the specifications that control for those effects forecast the IBC-Br considerably more accurately than those that do not. Its contribution to practice follows from this result: it cautions business and finance professionals against reading aggregate Pix growth as a monthly activity signal without first removing its calendar component, and it sets out the evaluation steps such a signal should pass before being used in the applications discussed in Section 4.7. The remainder of the paper is organized as follows: Section 2 reviews the literature on economic nowcasting and on payment data as a predictor of economic activity; Section 3 describes the data, the empirical strategy, and the information set assumed at each forecast origin; Section 4 reports the results, the robustness analysis, and their academic and managerial implications; and Section 5 concludes, discussing limitations and suggestions for future research.

2. Literature Review

2.1. Economic Nowcasting: Foundations

The term nowcasting was coined in meteorology and adapted to economics to describe the forecast of the present, the very recent past, or the very near future of an economic variable before its official release (Giannone et al., 2008). The central motivation is publication lag: indicators such as Gross Domestic Product are released weeks or months after the reference period, creating an information gap that monetary policymakers, risk managers, and market analysts must fill with partial and asynchronous information.

Dynamic factor models and bridge equations have become the institutional standard for nowcasting at central banks: the Federal Reserve Bank of New York uses a dynamic factor model that synthesizes dozens of high-frequency indicators into a single common factor, recently updated to better accommodate pandemic-era data volatility (Almuzara et al., 2023), the Federal Reserve Bank of Atlanta uses a bridge equation in its GDPNow indicator, aggregating statistical forecasts of thirteen subcomponents of U.S. GDP (Higgins, 2014), and the Bank of England uses a combination of MIDAS, dynamic factors, and a Bayesian mixed-frequency VAR (Anesti et al., 2022). A common thread across these institutional approaches is a preference for models with interpretable structure, even at the cost of being statistically simpler than machine-learning alternatives.

The bridge equation, specifically, originates in the methodological proposal of Baffigi et al. (2004), who formalized the use of simple linear regressions to “bridge” high-frequency information available before the end of the quarter to the low-frequency variable of interest, in the original case euro area GDP. The basic formulation of the method consists of regressing the low-frequency variable (for example, quarterly GDP) on a small set of high-frequency indicators previously temporally aggregated to the same frequency as the dependent variable, and then using short-term forecasts of those indicators, typically obtained from simple autoregressive models, to generate the nowcast of the low-frequency variable before its official release. This two-stage approach is well suited to handling the “ragged edge” problem, in which different high-frequency indicators are available at different reference dates within the same forecast period, a recurring issue in the mixed-frequency nowcasting literature (Giannone et al., 2008). As discussed in Section 1, this temporal-aggregation stage is the defining feature that separates a classical bridge equation from the same-frequency predictive regression estimated in this paper.

Since its original formulation, bridge equations have been adopted by a wide range of central banks and macroeconomic research institutes around the world, because of their estimation simplicity (typically by ordinary least squares) and their transparency for communication with policymakers, features that Golinelli and Parigi (2007) document in subsequent applications of the method in the euro area, and that continue to motivate its adoption at monetary policy institutions even decades after its original proposal.

A relevant application for this paper is that of Barlas et al. (2021), who use aggregated bank transaction data (person-to-business and business-to-business) from the Turkish bank Garanti BBVA to estimate bridge equations for nowcasting Turkish consumption, investment, and GDP, an emerging economy with a GDP release lag of ten to thirteen weeks, longer than that typically observed in developed economies. The authors document that the informational value of bank transaction data is especially high at the start of the nowcasting process, when traditional (“hard data”) indicator information is still scarce, a pattern structurally analogous to the Brazilian IBC-Br case, whose release lag of approximately 40 days, although shorter than the Turkish one, still leaves a relevant window in which transaction data, such as Pix, may carry incremental information. The main methodological difference between Barlas et al. (2021) and this paper is institutional scale: whereas the Turkish authors use transaction data from a single private commercial bank (Garanti BBVA), representing a fraction of the country’s banking activity, this paper uses data from the instant payment system operated directly by the Central Bank of Brazil, which covers the entirety of Pix transactions carried out in the country, reducing the sample-representativeness concern that accompanies the use of data from a single commercial bank.

This broader modeling tradition, of which the classical bridge equation is the mixed-frequency variant and the monthly predictive regression estimated in this paper is a same-frequency variant, has the advantage of establishing a direct and interpretable statistical relationship between the variable of interest (in this paper’s case, the IBC-Br) and a small number of indicators available before the official release. Dynamic factor models, in contrast, require estimating a latent state space from a potentially wide panel of series and typically demand longer samples for stable estimation of factor loading parameters. This feature makes the approach well suited to this paper’s empirical context: the series of interest (Pix transactions) has only been available for a relatively short period (since November 2020), which limits the use of methodologies that require longer samples for adequate statistical convergence.

2.2. International Applications of Bridge Equations in Emerging Economies

The application of bridge equations to emerging economies, although less extensive than in developed economies, has relevant precedents that serve as a direct methodological reference for this paper. Gálvez-Soriano (2020), of the Bank of Mexico, compares a dynamic factor model, two bridge equation models, and two principal-component-based models for nowcasting Mexican quarterly GDP, finding that the average of the bridge equation forecasts statistically outperforms the other models according to the Diebold and Mariano accuracy test, and is also more accurate than the median forecast of market analysts surveyed by Bloomberg and by the Bank of Mexico’s own expectations survey. The author attributes this superiority to the fact that, when the available set of indicators has high and heterogeneous variance, extracting a reduced number of common factors, as dynamic factor models do, discards idiosyncratic information that the bridge equation, by treating each indicator individually, is able to preserve.

Assunção and Fernandes (2022), in an application to Portuguese GDP, combine bridge equations with the Hodrick-Prescott filter to forecast the as-yet-unreleased values of high-frequency indicators within the current quarter, showing that this simple multivariate approach performs comparably to, or better than, targeted diffusion index combination methods and the univariate Theta method. The authors also revisit the historical origin of the method, attributed to Klein and Sojo (1989), who first proposed constructing separate bridge equations for the nominal expenditure-side components of GDP, with real values subsequently obtained by dividing each nominal component by its respective implicit deflator, a methodological alternative to the direct-aggregation approach adopted by Baffigi et al. (2004) and followed in this paper.

Taken together, these international applications document three findings that guide this paper’s empirical strategy. First, bridge equations tend to outperform dynamic factor models when the number of available indicators is small and their variance is high, a situation similar to this paper’s, in which only two regressors (Selic and Pix growth) are available. Second, the predictive performance of bridge equations in emerging economies has been validated with formal statistical tests of comparative accuracy, and not merely with descriptive error metrics. Third, the method’s simplicity allows its replication and adaptation to different institutional contexts, from quarterly Mexican and Portuguese GDP to monthly Brazilian IBC-Br, without substantive changes to the model’s core specification.

An additional point that emerges from comparing these studies is the diversity of variables already tested in bridge equations around the world: economic sentiment indicators and cement sales in Portugal (Assunção & Fernandes, 2022), expectations surveys and analyst estimates in Mexico (Gálvez-Soriano, 2020), industrial production and survey-based indicators in the euro area (Golinelli & Parigi, 2007), and, closer to this paper’s scope, payment system data in Italy, Switzerland, and Turkey (Aprigliano et al., 2019; Barlas et al., 2021; Felber & Beyeler, 2023). This diversity suggests that the bridge equation functions less as a method tied to a specific type of regressor, and more as a flexible framework for incorporating any indicator whose relationship with the target variable is theoretically justifiable, which reinforces the suitability of this modeling tradition for testing this paper’s hypothesis about Pix as an indicator of economic activity.

Outside the bridge-equation and dynamic-factor traditions, a separate strand of the forecasting literature handles persistence and seasonality through parsimonious model structure rather than through large sets of indicator variables. Orlando and Bufalo (2021) adapt a single-factor CIR-type model (CIR#) to forecast interest rates and extend the framework to series with pronounced seasonality, such as tourism demand (Orlando & Bufalo, 2023), and Ceci et al. (2024) apply a related specification (CIR3) to the industrial production of electric and gas utilities, a series with marked seasonal patterns. These studies compare competing forecasts with the Diebold–Mariano test and its small-sample correction, as this paper does. They are not substitutes for the payment-data nowcasting literature, but they offer an instructive contrast for the present paper: they embed seasonality in a small number of model parameters, whereas the predictive regressions estimated here control for it with quarter-of-year or month-of-year indicator variables, at a cost in degrees of freedom that is not trivial in a sample of about five years. Section 4.3 returns to this contrast.

2.3. Payment Data as a High-Frequency Indicator

Aprigliano et al. (2019) use Italian payment system data to forecast quarterly economic activity, finding relevant predictive gains during periods of greater macroeconomic volatility, when traditional survey-based indicators tend to lose precision. Felber and Beyeler (2023), of the Swiss National Bank, obtain a similar result using payment transaction data to forecast Swiss private consumption, documenting that the informational value of this data is especially high in the first weeks of each month, before the release of competing traditional indicators. Aastveit et al. (2024), in the Journal of Applied Econometrics, use debit card data to forecast Norwegian household consumption, documenting pronounced forecasting gains during the COVID-19 pandemic, a period in which consumption patterns shifted abruptly toward digital channels and traditional indicators, based on periodic sample surveys, became lagged.

Golinelli and Parigi (2007) document subsequent applications of bridge equations in the euro area using survey-based indicators and industrial production data, confirming the robustness of the method to different compositions of high-frequency regressors. More recently, Cascaldi-Garcia et al. (2024), in the International Journal of Forecasting, build a multi-country now-casting model for the euro area and its three largest member countries, finding that bridge-equation-style pooling of area-wide and country-specific indicators remains informative even after two decades of continuous institutional refinement of euro area nowcasting practice, while Antolín-Díaz et al. (2024), in the Journal of Econometrics, show that explicitly modeling heterogeneous dynamics and fat-tailed shocks materially improves nowcasting accuracy during periods of high volatility, a consideration directly relevant to the COVID-era portion of this paper’s sample.

A recurring result in this literature is that the informational value of payment data tends to be especially high during periods of greater uncertainty or structural disruption, when traditional data-collection channels (sample surveys, questionnaires) face longer lags or response bias. A related line of research explores inter-industry payment data in the United Kingdom: Hötte (2025) empirically validates a database of sectoral Bacs system payments as an input for macroeconomic models, finding a strong contemporaneous correlation between sectoral flows and gross value added, a development structurally analogous to the Brazilian Payments-by-Economic-Activity (EPAE) statistic discussed in Section 3, and directly related to this paper’s institutional motivation.

Beyond payment data narrowly defined, the broader family of non-traditional high-frequency indicators explored in recent nowcasting research includes textual and sentiment-based signals: Ashwin et al. (2024), in the Journal of Applied Econometrics, show that sentiment extracted from newspaper articles materially improves real-time nowcasts of euro area GDP growth, particularly early in the quarter, before other hard-data indicators become available, a timing pattern consistent with the value of Pix data highlighted in this paper, which is likewise available within days of month-end rather than with the multi-week lag typical of survey-based indicators.

An aspect that this literature has explored little, and that this paper seeks to address in part in Section 3.1, is the validation of the integrity of payment data extracted from public programmatic-access sources (APIs). Unlike traditional sample-survey datasets, whose collection and processing methodology is documented in published technical manuals, central bank open-data APIs frequently lack detailed documentation of the exact behavior of their query parameters, which can introduce subtle aggregation errors if the researcher does not validate the extracted data against independent references, a methodological concern revisited with a concrete example in Section 3.1 of this paper.

2.4. Pix and the Literature on Financial Inclusion and Instant Payments

The academic and institutional literature on Pix has, to date, concentrated mainly on the determinants of its adoption, its effects on financial inclusion, and the regulatory implications of payment digitalization, with few exceptions devoted to its application as a statistical input for macroeconomic models. The International Monetary Fund (2023) documents the design features of Pix that contributed to its rapid adoption (free of charge for individuals, 24 h availability, mandatory interoperability among participating institutions), and highlights that the granularity of transaction data, segmented by users’ demographic and geographic characteristics, constitutes a relevant informational byproduct beyond the original goal of financial inclusion, even though this was not the declared policy objective at the time the system was conceived.

Da Cunha Duarte (2025), in the Eurasian Economic Review, analyzes the potential of Pix to promote women’s digital financial inclusion in Brazil, documenting usage asymmetries by age group and highlighting the absence of a gender-sensitive regulatory framework on the part of the Central Bank. Although that study’s focus differs from this paper’s, it reinforces a relevant methodological point: the demographic and geographic granularity of Pix transaction data (the same granularity exploited technically in this paper for nowcasting purposes) is being progressively recognized in the institutional and academic literature as a multi-use informational asset, beyond the monitoring of financial inclusion itself.

This multiplicity of potential uses for the same dataset (financial inclusion, gender-sensitive regulation, and, in this paper, macroeconomic nowcasting) is, in itself, an argument in favor of more Brazilian academic research exploring the Pix database, given that the costs of initial data extraction and processing (documented in Section 3.1 of this paper) can be amortized across multiple subsequent lines of investigation. The use of Pix specifically as an input for short-term macroeconomic forecasting models remains, however, an academically underexplored research gap in Brazil, which this paper seeks to fill.

2.5. Summary and Positioning of the Study

Table 1 positions the present study relative to the international literature reviewed, highlighting the gap that this paper seeks to address: the absence of Brazilian studies that use Pix as a nowcasting regressor within a monthly predictive-regression framework, in the bridge-equation tradition, with explicit statistical inference and data extracted directly from the primary source. The comparison with Barlas et al. (2021) is instructive: both studies apply this modeling tradition to transaction data in emerging economies, but the present study benefits from a universal-coverage data source (the entire Pix system, operated by the Central Bank), rather than a sample from a single commercial bank, which reduces representativeness concerns that necessarily characterize studies based on data from individual private financial institutions.

Table 1. Positioning of the reviewed literature.

Study	Country	Payment Data	Method	Brazil/Pix
Klein and Sojo (1989)	USA	—	Bridge equation (origin)	No
Giannone et al. (2008)	General/Methodological	—	Dynamic factors	No
Baffigi et al. (2004)	Euro area	—	Bridge equation	No
Golinelli and Parigi (2007)	G7	Various indicators	Bridge equation	No
Higgins (2014)	USA	—	Bridge equation (GDPNow)	No
Aprigliano et al. (2019)	Italy	Payment system	Bridge equation	No
Gálvez-Soriano (2020)	Mexico	—	Bridge equation, factors, PCA	No
Barlas et al. (2021)	Turkey	Bank transactions	Bridge equation, factors, BVAR	No
Anesti et al. (2022)	United Kingdom	—	MIDAS, factors, Bayesian VAR	No
Assunção and Fernandes (2022)	Portugal	—	Bridge equation	No
Felber and Beyeler (2023)	Switzerland	Payment transactions	Bridge equation	No
Aastveit et al. (2024)	Norway	Debit card	Predictive regression	No
Hötte (2025)	United Kingdom	Inter-industry payments	Correlation/regression	No
Central Bank of Brazil (2025)	Brazil	Pix (EPAE)	Descriptive	Yes
Present study	Brazil	Pix (transactions, complete)	Monthly predictive regression (bridge-equation tradition)	Yes

Note. Own elaboration based on the literature review in Section 2.

3. Data and Methodology

3.1. Data Sources

The Central Bank Economic Activity Index (IBC-Br) is the official monthly proxy for Brazilian GDP, released with a lag of approximately 40 days. This paper uses two versions of it, both extracted from the BCB’s Time Series Management System (SGS): the seasonally unadjusted index (SGS code 24363), which is the main dependent variable, and the officially seasonally adjusted index (SGS code 24364), used as an alternative dependent variable. The monthly accumulated Selic rate (SGS code 4390) is used as a control variable, since the monetary policy cycle may simultaneously affect economic activity and the volume of Pix transactions. Two further series serve as additional controls that are external to the IBC-Br and available shortly after each month’s end: the monthly average BRL/USD exchange rate (SGS code 3698) and the BCB Commodity Index, IC-Br (SGS code 27574), both used in monthly log changes. Monthly IPCA inflation (SGS code 433) and the IBC-Br’s industrial component (SGS code 29603) are used in the sectoral decomposition of Section 4.5. All SGS series were downloaded in October 2026, so the IBC-Br vintage is the one available at that date (Section 3.3). These series are widely used in applied Brazilian macroeconomics and present no known quality or availability problems, unlike the Pix series discussed below.

The choice of the Selic rate as the control variable, rather than other monetary policy indicators, follows directly from the tradition of Aprigliano et al. (2019) and is justified by the monetary policy transmission channel to economic activity: increases in the benchmark interest rate raise the cost of credit and tend to cool consumption and investment with a lag of a few months, while potentially affecting the volume of Pix transactions contemporaneously, for example through changes in Brazilian households’ saving and spending behavior. Including the Selic rate as a control regressor helps net out the predictive association between Pix and the IBC-Br that could otherwise be attributable to the monetary policy cycle itself, rather than isolating a causal effect; the specification establishes a conditional predictive association, not a causal relationship, and the contemporaneous Selic rate is itself related to macroeconomic conditions and monetary-policy responses that could also correlate with Pix growth.

The BCB makes Pix data available through its Open Data API (Olinda platform, OData protocol, EstatisticasTransacoesPix resource), with monthly granularity disaggregated by transaction type (P2P, P2B, B2P, B2B, P2G, B2G, G2x), initiation method, region, and age bracket of participants. The complete panel extraction (November 2020 to May 2026, 714,576 unique dimension combinations) was carried out and consolidated into a monthly total-volume series (sum across all transaction types), reported in Table 2. The Python 3.12.6 code used for data extraction, correction, and estimation is provided in the Supplementary Materials (Code S1).

Table 2. Verified data points on Pix transactions.

Period	Volume (R$ Billion)	Transactions (Million)	Source
Nov/2020, all transaction types	25.9	29.5	Own extraction, BCB EstatisticasTransacoesPix API
Nov/2020, P2P only	12.1	25.7	Own extraction; consistent with the value cited in the BCB EPAE technical box for month 1
Dec/2020	107.2	124.9	Own extraction, BCB EstatisticasTransacoesPix API
Sep/2025	2651.6	6133.3	Own extraction, BCB EstatisticasTransacoesPix API
May/2026 (last available month)	2957.5	7121.9	Own extraction, BCB EstatisticasTransacoesPix API

Note. Values under “all transaction types” correspond to the sum of P2P, P2B, B2P, B2B, P2G, B2G, and transactions involving government (G2x). The “P2P only” row isolates the person-to-person segment for comparison with press statistics, which typically report this subset. Own elaboration based on the BCB EstatisticasTransacoesPix API.

One methodological point of rigor deserves explicit mention. In the initial extraction from the EstatisticasTransacoesPix API, it was found that the reference-period filter parameter (@Database) does not behave as an exact match: each monthly query cumulatively returns all records from the requested month up to the latest available month, rather than only the specific month requested. This behavior generates record duplication when multiple monthly queries are naively concatenated. In this paper’s case, direct concatenation of 68 monthly queries produced 28,487,932 raw rows, versus 714,576 unique records after deduplication on the full combination of dimensions (true reference period, transaction type, initiation method, region and age bracket of payer and payee, value, and quantity). Deduplication was validated by comparing the P2P (person-to-person) subset for Pix’s first month of operation against the value disclosed in the BCB’s technical box on EPAE: the value obtained (R$ 12.1 billion) matches the officially cited value (R$ 12.0 billion) with a difference of 0.8%, confirming the series’ integrity after the correction procedure. This validation process is reported here for methodological transparency and because it illustrates a replicable risk for other researchers who may use the same data source. In particular, researchers seeking to reproduce or extend this paper’s results should be aware of this specific API behavior before concatenating multiple monthly queries, on pain of artificially inflating aggregate totals in a way that is not evident from simple inspection of the row count returned by each individual query.

Pix began operating for the general public on 16 November 2020, so the November 2020 observation covers only about half a month of activity, and the December 2020 growth rate (about 144 log points, Table 2) mechanically reflects the comparison of a partial launch month with a full month rather than any change in economic activity. Because this single observation lies far outside the range of every later month, it has very high leverage in any regression that includes Pix growth: in the first-difference specification of Section 3.2, excluding it raises the estimated Pix coefficient from about 6 to about 25. The primary estimation sample therefore begins in January 2021, the first month whose growth rate compares two full months of operation, and ends in May 2026, giving 65 monthly observations. Results that include December 2020 are reported as a robustness check in Section 4.4.

3.2. Empirical Strategy

This paper follows the standard nowcasting structure (Giannone et al., 2008): forecasting the current month t’s IBC-Br using information available before its own official release. Because the IBC-Br in levels is non-stationary (Section 3.5), and a regression of its level on its own lag leaves non-stationary residuals (Section 4.1), all predictive regressions are specified in first differences:

Δy_t = α + β₁·Δy_{t − 1} + β₂·Selic_t + β₃·ΔlogPix_t + γ′s_t + ε_t,

where y_t is the IBC-Br, ΔlogPix_t is the monthly logarithmic growth rate of the total quantity of Pix transactions (sum across all transaction types: P2P, P2B, B2P, B2B, P2G, B2G, G2x), and s_t is a vector of seasonal controls described below. The restricted model, without Pix, sets β₃ = 0. The forecast of the level is ŷ_t = y_{t − 1} + Δŷ_t, so forecast errors for the level and for the change coincide. Logarithmic growth, rather than the level, is used for Pix because it is the conventional transformation for rendering transaction-volume series approximately stationary (Aastveit et al., 2024). For comparison only, the same regressors are also estimated with the level y_t as the dependent variable and y_{t − 1} in place of Δy_{t − 1}; these level regressions are reported as reference rows, since their residuals are not stationary.

Seasonality is the central modeling issue in this application. Pix transaction volume displays pronounced calendar patterns (around Black Friday in November, the mandatory year-end bonus in December, and the start of the year), and the unadjusted IBC-Br is subject to related calendar effects (business-day counts, holiday timing). Part of the co-movement between the two series could therefore reflect shared calendar effects rather than information about activity. Each specification is estimated under four seasonal treatments: (i) no seasonal controls; (ii) three quarter-of-year dummies; (iii) eleven month-of-year dummies; and (iv) no dummies, with the officially seasonally adjusted IBC-Br (SGS 24364) as the dependent variable. Treatments (ii) and (iii) differ sharply in their cost in degrees of freedom, a point raised in the parsimonious seasonal-modeling literature discussed in Section 2.2 and examined directly in Section 4.3.

The core specification controls only for the Selic rate, following Aprigliano et al. (2019), because the Selic rate is known in real time throughout the month being forecast, so including it does not compromise the timing of the exercise described in Section 3.3, unlike most other standard drivers of the IBC-Br, whose own publication lags are comparable to or longer than the IBC-Br’s. Because this leaves room for omitted-variable bias, Section 4.5 adds two controls that are external to the IBC-Br and also available shortly after each month’s end, the exchange rate and the IC-Br commodity index, and reports separately a decomposition with IPCA inflation and the IBC-Br’s own industrial component.

The specification adopted in this paper differs from Baffigi et al.’s (2004) classic bridge equation in one central respect: while the original formulation treats the low-frequency variable (quarterly GDP, in the original case) and the high-frequency indicators as distinct frequencies to be reconciled through temporal aggregation, this paper works entirely at the monthly frequency, since both the IBC-Br and Pix are released on that cadence. This eliminates the need for a temporal-aggregation stage or an auxiliary model to forecast the within-period values of a higher-frequency regressor, a technical issue central to the classic application of the bridge equation to quarterly data, but one that does not arise for the pair of monthly series used in this paper. For this reason, and as flagged in Section 1, the remainder of this paper refers to the specifications above as monthly predictive regressions, reserving the term bridge equation for the classical mixed-frequency method reviewed in Section 2.1 and for the cited studies that implement it; where the two terms could otherwise be read as interchangeable, this paper’s own model is the one intended by “predictive regression.”

The evaluation follows the pseudo-out-of-sample procedure with an expanding window: at each forecast origin, all models are re-estimated by ordinary least squares using only the observations available up to that origin, including the coefficients on the seasonal dummies, and a one-step-ahead forecast is produced. Because the seasonal specifications are parameter-intensive relative to the sample, results are reported for three minimum estimation windows of 18, 29, and 39 observations, which yield 47, 36, and 26 one-step-ahead forecasts beginning in July 2022, June 2023, and April 2024, respectively, all ending in May 2026. With eleven monthly dummies, the first window leaves 3, 14, and 24 residual degrees of freedom in the three cases, and the minimum window can therefore be chosen so that every model is estimated with adequate degrees of freedom (see Section 4.2). Forecasts are also compared with a naive (random-walk) benchmark and with a seasonal-naive benchmark, the value of the same calendar month one year earlier.

Forecast accuracy is measured by the root mean squared error (RMSE) and the mean absolute error. Because the model with Pix nests the model without it, we report both the Diebold–Mariano test (Diebold & Mariano, 1995) with the Harvey et al. (1997) small-sample correction and the Clark–West test (Clark & West, 2007), which corrects the downward bias of the former in nested comparisons. The Clark–West statistic relies on asymptotic approximations that may be unreliable when the larger model is estimated with few degrees of freedom in the early windows. We therefore also compute bootstrap p-values in the spirit of Clark and McCracken (2012): the restricted model is estimated on the full sample, the dependent variable is regenerated recursively from resampled residuals under the null of no Pix effect with the other regressors held fixed, and the complete expanding-window exercise, including re-estimation of both models at every origin, is repeated 999 times. The bootstrap thus reproduces the small-sample estimation noise of the early windows. In-sample inference uses Newey and West (1987) standard errors with three lags.

3.3. Information Set and Real-Time Considerations

Because the central premise of this paper is that Pix growth is available to a forecaster before the IBC-Br value it seeks to anticipate, the information set assumed at each forecast origin is stated here precisely, rather than left implicit in the model equations above. For a forecast of the IBC-Br value referring to month t, produced at a forecast origin shortly after month t’s end (once that month’s Pix, Selic, exchange-rate, and commodity-index data become available, but before the IBC-Br release for month t), the predictive regressions use the following information set:

IBC-Br: the values up to month t − 1, the latest released observation, which determine the lagged change Δy_{t − 1} and the level y_{t − 1} to which the forecast change is added; the IBC-Br for month t itself is the quantity being forecast and is not yet available at the assumed origin, given its publication lag of about forty days.

Selic: the accumulated monthly rate for month t, which the BCB discloses on a rolling daily basis and which is therefore fully known by month t’s end, well before the IBC-Br release for that month.

Pix: the transaction growth rate for month t, ΔlogPix_t, computed from data the BCB discloses through its open API within days of month t’s end (Section 3.1), and therefore also available before the IBC-Br release for month t.

Exchange rate and IC-Br: the monthly averages for month t, which the BCB publishes in the first days of month t + 1, and therefore also available before the IBC-Br release for month t. IPCA inflation and the IBC-Br’s industrial component do not share this timing and are used only in the full-sample decomposition of Section 4.5.

This timeline is what gives Pix its putative informational advantage in this application: not a higher sampling frequency than the IBC-Br, as discussed in Section 1, but an earlier publication date for the same monthly reference period. A conventional monthly indicator with a publication lag comparable to, or longer than, the IBC-Br’s own forty days would not offer this advantage, whatever its in-sample informational value; the case for Pix as a nowcasting input rests specifically on its short publication lag.

One limitation of the empirical implementation concerns data vintages. The IBC-Br is revised after its first release, and a fully real-time evaluation would compare each forecast with the first-release value and use the vintage of the lagged regressors available exactly at each origin. The BCB’s SGS platform does not provide an archive of past IBC-Br vintages, so the exercise uses the series as available in October 2026 for both the lagged regressors and the evaluation target, and is better described as a pseudo-real-time evaluation of predictive content than as a genuine real-time vintage exercise. Because a first-release dataset cannot be assembled from public sources, we bound the potential revision bias in two ways.

First, we compare two vintages of the IBC-Br, downloaded in mid-2026 and in October 2026. Over November 2020 to May 2026, the sample mean of the index changes by 0.02 points. For a specification that includes December 2020 and uses the IBC-Br in levels, the Pix coefficient is 5.68 with the earlier vintage and 5.62 with the later one, and the Clark–West p-values for the comparison with month-of-year dummies are 0.011 and 0.016, respectively; no qualitative conclusion changes. Second, we add independent noise with standard deviations of 0.25, 0.5, and 1.0 index points to every IBC-Br observation (200 draws each), well above the revision observed between the two vintages, rebuild all derived variables, and repeat both the full-sample estimation and the out-of-sample exercise with a 29-observation minimum window. Table 3 reports the results.

Table 3. Sensitivity of the Pix results to simulated IBC-Br revisions.

Revision Noise (s.d., Index Points)	Specification	Median Change in RMSE from Adding Pix	Draws in Which Pix Lowers RMSE	Draws with Clark–West p < 0.05	Draws with in-Sample Pix p < 0.05
0.25	First difference	−5.0%	100.0%	100.0%	100.0%
0.25	First difference + quarterly dummies	−13.2%	100.0%	100.0%	100.0%
0.25	First difference + monthly dummies	+0.8%	0.0%	0.0%	0.0%
0.50	First difference	−5.0%	100.0%	99.0%	100.0%
0.50	First difference + quarterly dummies	−13.0%	100.0%	100.0%	100.0%
0.50	First difference + monthly dummies	+0.8%	0.0%	0.0%	0.0%
1.00	First difference	−4.3%	98.0%	78.5%	100.0%
1.00	First difference + quarterly dummies	−11.1%	100.0%	100.0%	100.0%
1.00	First difference + monthly dummies	+0.8%	3.5%	0.0%	0.0%

Note. Independent normal noise added to every IBC-Br observation, 200 draws per noise level; primary sample (January 2021 to May 2026); expanding window with a 29-observation minimum. Negative values in the third column indicate that the model with Pix has the lower RMSE. Own elaboration.

Across all three noise levels, the specifications without month-of-year controls continue to show a lower RMSE with Pix in nearly every draw, and the specification with month-of-year dummies continues to show no gain from Pix and no significant Pix coefficient in any draw. Revisions of a plausible size therefore do not alter the pattern documented in Section 4. Assembling a first-release dataset from the BCB’s internal revision history remains a direction for future work (Section 5).

3.4. Descriptive Statistics

Table 4 presents descriptive statistics for the series used in the primary sample (January 2021 to May 2026).

Table 4. Descriptive statistics of the series used (January 2021 to May 2026).

Variable	Mean	Std. Dev.	Min	Max	N
IBC-Br, unadjusted (index)	103.67	5.76	91.68	118.07	65
IBC-Br, seasonally adjusted (index)	103.70	4.65	95.33	110.93	65
Monthly change in IBC-Br, unadjusted	0.19	4.07	−7.64	11.97	65
Selic, monthly accumulated (%)	0.89	0.29	0.13	1.28	65
Pix growth (log-diff.)	0.062	0.096	−0.121	0.359	65
Exchange rate, BRL/USD (log-diff.)	0.000	0.027	−0.063	0.061	65
IC-Br commodity index (log-diff.)	0.008	0.034	−0.052	0.107	65
IBC-Br Indústria (log-diff.)	0.002	0.046	−0.097	0.107	65

Note. IBC-Br in index form (2003 average = 100); Selic monthly accumulated rate, in percentage points; log-differences are monthly. Own elaboration.

The correlation between the level of the IBC-Br and the Selic rate is positive and moderate (0.62), reflecting the period of rising benchmark interest rates that partially accompanied the post-pandemic recovery. The correlation between the monthly change in the unadjusted IBC-Br and contemporaneous Pix growth is positive (0.33), in the direction a nowcasting role would require. Two features of the data, however, anticipate the results of Section 4. First, Pix growth has a marked calendar pattern: its average in March (0.17) and December (0.13) is two to three times its overall mean, which is also the pattern of the unadjusted IBC-Br around those months. Second, the correlation between Pix growth and the monthly change in the officially seasonally adjusted IBC-Br is slightly negative (−0.11). Once the calendar component is removed from the IBC-Br, the positive simple association therefore does not survive. The predictive regressions in Section 4 test whether the conditional relationship behaves in the same way.

3.5. Unit Root and Stationarity Tests

Before estimating the predictive regressions, we formally test the stationarity properties of each series using the Augmented Dickey–Fuller (ADF) test (Dickey & Fuller, 1979) and the Kwiatkowski–Phillips–Schmidt–Shin (KPSS) test (Kwiatkowski et al., 1992), which have complementary null hypotheses (respectively, a unit root and stationarity) and are therefore reported jointly to avoid over-reliance on either test in isolation. Table 5 reports the results for each series.

Table 5. Unit root and stationarity tests (January 2021 to May 2026).

Series	ADF (p-Value)	KPSS (p-Value)	Conclusion
IBC-Br, unadjusted (level)	0.842	0.010	Non-stationary
IBC-Br, unadjusted (first difference)	<0.001	0.100	Stationary
IBC-Br, seasonally adjusted (level)	0.845	0.010	Non-stationary
IBC-Br, seasonally adjusted (first difference)	<0.001	0.100	Stationary
Selic (level)	0.035	0.016	Borderline/conflicting
Pix growth (monthly log-diff.)	<0.001	0.010	Conflicting
Pix growth (Zivot-Andrews, structural break)	0.003	n.a.	Stationary

Note. ADF null hypothesis: unit root; KPSS null hypothesis: stationarity. KPSS p-values are bounded by the tabulated range (0.010 to 0.100). n.a.: not computed. The final row reports the p-value of the Zivot and Andrews (1992) test, which allows for a single endogenous break in the intercept (one lag; statistic = −5.61; estimated break in October 2021); with two lags the p-value rises to 0.093. Own elaboration.

The IBC-Br, adjusted or not, is non-stationary in levels and stationary in first differences by both tests, which is why the predictive regressions are specified in first differences. For Pix growth, the ADF test now rejects a unit root clearly once the launch month is excluded from the sample, while the KPSS test still rejects stationarity, a conflict that is common in series of this length, where both tests have limited power to separate a persistent stationary process from a non-stationary one. The dominant feature of Pix growth in this period is a steady deceleration as the system matures, a pattern more consistent with a shifting mean than with a stochastic trend. A Zivot–Andrews test allowing for one break in the mean rejects a unit root (p = 0.003, with a break in late 2021), although the result is sensitive to the lag length. We therefore treat Pix growth as stationary around a slowly declining mean, and return to the consequences of that declining mean in Section 4.3. The Selic rate shows the same borderline pattern; as a policy rate bounded by the monetary policy regime, it is treated as stationary in the regressions.

4. Results and Implications

This section proceeds in seven steps. Section 4.1 reports full-sample estimates and residual diagnostics for each specification, and Section 4.2 reports the out-of-sample comparison. Section 4.3 explains why seasonal control is decisive for the results, Section 4.4 examines the reliability of the nested-model tests and the influence of the launch month, and Section 4.5 adds controls external to the IBC-Br and a sectoral decomposition. Section 4.6 derives a single economic magnitude using a principled selection rule, and Section 4.7 discusses implications for financial-market practice.

4.1. Full-Sample Estimates and Residual Diagnostics

Table 6 reports the Pix coefficient for each specification estimated on the primary sample, together with residual diagnostics.

Table 6. Full-sample estimates of the Pix coefficient and residual diagnostics (January 2021 to May 2026).

Specification	Pix Coef.	NW s.e.	p-Value	Adj. R2	df	BG p	ADF Resid. p
First difference, no seasonal controls	24.832	5.012	<0.001	0.196	61	<0.001	0.717
First difference + quarterly dummies	27.655	5.830	<0.001	0.418	58	<0.001	0.665
First difference + monthly dummies	1.762	3.493	0.616	0.857	50	0.003	<0.001
Seasonally adjusted IBC-Br, first difference	−1.204	1.989	0.547	0.007	61	0.008	<0.001
Level, no seasonal controls (reference)	21.193	4.701	<0.001	0.677	61	<0.001	0.854
Level + quarterly dummies (reference)	25.145	6.071	<0.001	0.731	58	<0.001	0.711
Level + monthly dummies (reference)	0.956	3.529	0.787	0.926	50	0.001	0.040

Note. NW: Newey–West; df: residual degrees of freedom; BG: Breusch–Godfrey. All specifications include the Selic rate and the lagged dependent variable (the lagged change in first-difference specifications, the lagged level in level specifications). Newey–West standard errors with three lags. Breusch–Godfrey test with three lags (null: no serial correlation). ADF on residuals: null of a unit root. Own elaboration.

Three results stand out. First, the level specifications leave non-stationary residuals unless month-of-year effects are controlled for (ADF p of 0.85 and 0.71), which is why they are retained only as reference rows. Second, in first differences without monthly controls, the Pix coefficient is large and significant (24.8 without seasonal controls and 27.7 with quarterly dummies), but the residuals still fail the ADF test (p about 0.7), because the unadjusted monthly change in the IBC-Br retains a strong calendar pattern that neither specification removes. Third, once that pattern is removed, either with month-of-year dummies or by using the officially adjusted series, the residuals become stationary (ADF p < 0.001), and the Pix coefficient falls to 1.76 (p = 0.62) with monthly dummies and turns negative (−1.20, p = 0.55) with the adjusted series. The specifications whose residuals behave as the model requires are therefore exactly those in which Pix carries no significant information. Monthly dummies also raise the adjusted R² of the first-difference regression from 0.20 to 0.86. Serial correlation remains in all specifications (Breusch–Godfrey p ≤ 0.008), which is why inference relies on Newey–West standard errors.

Multicollinearity does not drive these results. In the first-difference specification without seasonal controls, all variance inflation factors lie below 1.6. With monthly dummies, the variance inflation factor of Pix growth rises to 4.27, still below conventional thresholds of concern, but this rise is itself informative: a substantial part of the variation in Pix growth is explained by the calendar month.

4.2. Out-of-Sample Forecast Accuracy

Table 7 reports the out-of-sample comparison between each specification with and without Pix, for the three minimum estimation windows.

Table 7. Out-of-sample accuracy with and without Pix growth, by specification and minimum estimation window (January 2021 to May 2026).

Window	Specification	k	df	RMSE no Pix	RMSE Pix	DM p	CW p	Boot. CW p
18	Level	4	14	3.946	3.625	0.207	0.009	0.004
18	Level + monthly dummies	15	3	1.936	1.970	0.362	0.395	0.322
18	Level + quarterly dummies	7	11	3.963	3.476	0.139	0.006	0.004
18	First difference	4	14	4.205	3.830	0.248	0.009	0.004
18	First difference + quarterly dummies	7	11	3.857	3.302	0.155	0.004	0.004
18	First difference + monthly dummies	15	3	1.897	1.927	0.142	0.776	0.637
18	Official SA IBC-Br (SGS 24364), first difference	4	14	0.780	0.785	0.616	0.612	n.a.
29	Level	4	25	3.904	3.747	0.522	0.043	0.034
29	Level + monthly dummies	15	14	2.009	2.041	0.126	0.925	0.880
29	Level + quarterly dummies	7	22	3.843	3.435	0.226	0.020	0.020
29	First difference	4	25	3.866	3.668	0.501	0.023	0.019
29	First difference + quarterly dummies	7	22	3.442	2.981	0.208	0.009	0.011
29	First difference + monthly dummies	15	14	1.877	1.891	0.136	0.931	0.892
29	Official SA IBC-Br (SGS 24364), first difference	4	25	0.584	0.575	0.190	0.054	0.038
39	Level	4	35	4.084	3.952	0.665	0.113	0.113
39	Level + monthly dummies	15	24	2.187	2.200	0.083	0.964	0.949
39	Level + quarterly dummies	7	32	3.933	3.526	0.329	0.051	0.061
39	First difference	4	35	4.077	3.864	0.558	0.058	0.055
39	First difference + quarterly dummies	7	32	3.573	3.159	0.360	0.035	0.042
39	First difference + monthly dummies	15	24	2.009	2.018	0.408	0.790	0.651
39	Official SA IBC-Br (SGS 24364), first difference	4	35	0.581	0.568	0.121	0.029	0.021

Note. One-step-ahead forecasts of the IBC-Br level (of the seasonally adjusted IBC-Br in the last row of each block); expanding window; all parameters re-estimated at each origin. Minimum windows of 18, 29, and 39 observations give 47, 36, and 26 forecasts, starting in July 2022, June 2023, and April 2024. Window: minimum estimation window; k: number of parameters in the model with Pix; df: residual degrees of freedom of that model in the first window. DM: Diebold–Mariano test with the Harvey–Leybourne–Newbold correction, two-sided, squared-error loss. CW: Clark–West test, one-sided. Bootstrap p-values from 999 replications under the null of no Pix effect (Section 3.2); n.a.: not computed. RMSE of the naive forecast: 4.04, 3.76, and 3.97; of the seasonal-naive forecast: 3.38, 3.25, and 3.52; of the random walk on the seasonally adjusted series: 0.78, 0.57, and 0.58, for the three windows. Own elaboration.

The out-of-sample results mirror the full-sample ones. In the specifications without month-of-year controls, adding Pix lowers the RMSE at every window, by about 5 to 9 percent in the first-difference specification and by 12 to 14 percent with quarterly dummies. The Diebold–Mariano test does not reject equal accuracy in any of these comparisons, while the Clark–West test rejects at the 18- and 29-observation windows and, in the first-difference specification with quarterly dummies, also at 39 observations, with bootstrap p-values close to the asymptotic ones. In the specifications with month-of-year dummies, adding Pix slightly raises the RMSE at every window, and the Clark–West p-values lie between 0.40 and 0.96. This remains true at the 39-observation window, where the monthly dummy model has 24 residual degrees of freedom in its first estimation, so the absence of a Pix gain is not an artifact of saturated early windows.

The comparison across specifications matters as much as the comparison within them. The models with month-of-year dummies, with or without Pix, forecast the unadjusted IBC-Br with an RMSE of about 1.9 to 2.2 at every window, against 3.0 to 4.2 for every specification without monthly controls, including those with Pix, and against 3.2 to 3.5 for the seasonal-naive benchmark. Lengthening the minimum window removes the Pix gain without removing the accuracy advantage of the monthly seasonal model, which indicates that the eleven-dummy specification is not simply fitting noise. A forecaster choosing among these models on accuracy would choose a monthly seasonal model, and in that model Pix does not help.

4.3. Why Seasonal Control Is Decisive

The quarterly dummy specification deserves particular attention, because Pix appears more useful in it than in any other specification and because it costs only three degrees of freedom. Its Pix gain does not, however, indicate a seasonality-independent signal. Quarter-of-year dummies leave the within-quarter monthly pattern of the unadjusted IBC-Br in the residual, and Pix growth is itself strongly seasonal at the monthly frequency, with average growth in March and December well above that of the other months (Section 3.4). The economical reading is that Pix growth absorbs part of the monthly seasonality that quarterly dummies leave behind. Two facts support this reading: the quarterly model with Pix remains far less accurate than the monthly model without Pix (Table 7), and the gain disappears as soon as month-of-year effects are controlled directly (Table 6 and Table 7).

The officially seasonally adjusted series offers a test of the same question that does not depend on estimating any seasonal parameters. With this series as the target, the full-sample Pix coefficient is negative and insignificant (Table 6). Out-of-sample, Pix does not help at the 18-observation window and lowers the RMSE by about 1.5 and 2 percent at the 29- and 39-observation windows, with Clark–West bootstrap p-values of 0.038 and 0.021. This gain does not reflect an activity signal. The recursively estimated Pix coefficient is negative in most windows, and in nearly all of them from the 29-observation minimum onward, so the gain comes from a negative association between the adjusted IBC-Br change and Pix growth, whose dominant feature in this period is the steady deceleration documented in Section 3.5. Pix growth then acts as a slowly moving intercept term rather than as a procyclical indicator of the kind the paper set out to test. The same sign is obtained when Pix growth is itself deseasonalized recursively within each training window.

These results also bear on the contrast with the parsimonious seasonal-modeling literature discussed in Section 2.2 (Ceci et al., 2024; Orlando & Bufalo, 2021, 2023). The eleven monthly dummies are an expensive way to represent seasonality in a sample of about five years, and the concern that they overfit early estimation windows is legitimate. In this application, however, the expense appears to be worth paying: at every window length the monthly specification forecasts more accurately than the quarterly one, and the officially adjusted series, which embeds the seasonal pattern without consuming any degrees of freedom in the regression, leads to the same conclusion about Pix. A parsimonious structural representation of seasonality of the kind developed in that literature would be a natural refinement once the sample is longer, but the evidence available here gives no indication that it would restore a positive Pix effect.

4.4. Reliability of the Nested-Model Tests and the Launch Month

Because the Clark–West test can be unreliable when the larger model is estimated with few degrees of freedom, Table 7 and Table 8 report bootstrap p-values alongside the asymptotic ones. In this design, the bootstrap does not indicate that the asymptotic test over-rejects: the bootstrap 95 percent critical values lie between 1.31 and 1.76, close to the asymptotic value of 1.645, and the bootstrap p-values are close to the asymptotic ones. The concern about the early windows nonetheless matters, for a different reason, as Table 8 shows for the sample that includes December 2020.

Table 8. Seasonally controlled comparisons in the sample including December 2020, by minimum estimation window.

Specification	Window	df	RMSE no Pix	RMSE Pix	DM p	CW p	Boot. CW p
Level + monthly dummies	19	4	1.957	1.935	0.124	0.016	0.005
Level + monthly dummies	30	15	2.026	2.019	0.679	0.179	0.133
Level + monthly dummies	40	25	2.207	2.192	0.393	0.148	0.111
First difference + monthly dummies	19	4	1.891	1.892	0.930	0.460	0.358
First difference + monthly dummies	30	15	1.886	1.876	0.347	0.142	0.111
First difference + monthly dummies	40	25	2.022	2.009	0.324	0.129	0.115
First difference + quarterly dummies	19	12	3.858	3.626	0.050	0.009	0.009
First difference + quarterly dummies	30	23	3.448	3.222	0.084	0.018	0.022
First difference + quarterly dummies	40	33	3.580	3.348	0.174	0.051	0.058

Note. Window and df as in Table 7. Sample December 2020 to May 2026; minimum windows of 19, 30, and 40 observations give 47, 36, and 26 forecasts. Bootstrap as described in Section 3.2. Own elaboration.

When the launch month is included, the comparison between the level specifications with and without Pix under monthly dummies produces a significant Clark–West statistic at the shortest window (p = 0.016), where the model with Pix has only four residual degrees of freedom in its first estimation, but the significance disappears at 30 and 40 observations (p = 0.179 and 0.148). In the first-difference specification with monthly dummies, the Clark–West test does not reject at any window. A significant nested-test result that holds only when the larger model is nearly saturated, and only in the sample that contains the launch-month outlier, should not be read as evidence of predictive content. The excluded observation matters in-sample as well: in first differences without seasonal controls, the Pix coefficient is about 6 with December 2020 and about 25 without it (Section 3.1), because a single growth rate of about 144 log points dominates the variation in the regressor.

4.5. Controls External to the IBC-Br and a Sectoral Decomposition

Controlling only for the Selic rate leaves the specifications open to omitted-variable bias. Table 9 adds, first, two controls that are external to the IBC-Br and available at a short lag, the monthly log changes in the exchange rate and of the IC-Br commodity index, and, second, IPCA inflation and the monthly log change in the IBC-Br’s industrial component.

Table 9. Pix coefficient with additional controls, by seasonal treatment (January 2021 to May 2026).

Controls Added	First Difference	First Difference + Quarterly Dummies	First Difference + Monthly Dummies
None	24.832 (<0.001)	27.655 (<0.001)	1.762 (0.616)
Exchange rate (Δlog)	24.565 (<0.001)	27.233 (<0.001)	1.763 (0.618)
IC-Br commodity index (Δlog)	24.863 (<0.001)	27.925 (<0.001)	1.832 (0.584)
Exchange rate and IC-Br	24.535 (<0.001)	27.225 (<0.001)	1.867 (0.579)
IPCA inflation	25.166 (<0.001)	27.779 (<0.001)	3.903 (0.229)
IBC-Br Indústria (Δlog), sectoral decomposition	16.876 (<0.001)	18.495 (0.003)	−1.637 (0.368)

Note. Pix coefficient, with the Newey–West p-value in parentheses. All specifications include the lagged change in the IBC-Br and the Selic rate. The exchange rate (SGS 3698) and the IC-Br (SGS 27574) are published shortly after each month’s end and are consistent with the information set of Section 3.3. IPCA and IBC-Br Indústria (SGS 29603) are not available at the same short lag; IBC-Br Indústria is also a component of the dependent variable, so its row is a sectoral decomposition rather than a test of predictive content. Own elaboration.

The controls that are external to the IBC-Br leave the Pix coefficient essentially unchanged in every seasonal treatment: it remains large and significant without monthly controls and small and insignificant with them. The omitted-variable concern therefore does not explain the pattern in Table 6 and Table 7; seasonality does. Adding the industrial component of the IBC-Br lowers the coefficient by about one third in the specifications without monthly controls. Because the industrial component is part of the dependent variable, this row does not test whether Pix carries aggregate predictive content; it indicates how much of the co-movement between Pix growth and the IBC-Br runs through the industrial sector. The correlation between Pix growth and the monthly change in industrial production is 0.20 in the primary sample, consistent with the business-to-business share of Pix transactions.

4.6. Economic Magnitude Under a Principled Selection Rule

Reporting a different magnitude for each specification would leave the reader without a basis for assessing economic significance. We therefore derive a single magnitude in two complementary ways, both fixed before the results were examined.

The first is a recursive selection rule. All twelve candidate models (the level and first-difference specifications without seasonal controls, with quarterly dummies, and with monthly dummies, each with and without Pix) produce out-of-sample forecasts from the 18-observation minimum window. After a burn-in of 12 forecasts, at each origin the rule selects the model whose mean squared error over all previous out-of-sample forecasts is smaller than that of every other candidate. It selects the first-difference model with monthly dummies and without Pix at 23 of 35 origins, and the level model with monthly dummies and without Pix at the remaining 12; a model with Pix is never selected. With a 29-observation minimum window and a burn-in of six forecasts, the rule selects a model without Pix at 27 of 30 origins. When the candidates are restricted to models with Pix, the rule again selects the two monthly dummy specifications.

The second is a model-averaged estimate. Table 10 weights the six specifications with Pix by their Bayesian information criterion (BIC) on the full primary sample and computes a model-averaged coefficient and standard error following Buckland et al. (1997).

Table 10. BIC weights and model-averaged Pix coefficient (January 2021 to May 2026).

Specification	Pix Coefficient	Std. Error	BIC	BIC Weight
Level, no seasonal controls	21.193	5.445	351.17	0.000
Level + quarterly dummies	25.145	5.086	348.65	0.000
Level + monthly dummies	0.956	4.304	288.77	0.191
First difference, no seasonal controls	24.832	5.976	365.25	0.000
First difference + quarterly dummies	27.655	5.152	353.56	0.000
First difference + monthly dummies	1.762	4.138	285.88	0.809
Model-averaged	1.608	4.181

Note. Conventional OLS standard errors. Model-averaged standard error computed as Σ w_k [se_k² + (β_k − $\overline{\mathsf{\beta }}$)²]^{1/2}. The level and first-difference specifications are models of the same quantity, y_t conditional on its own past, estimated on the same observations, so their BIC values are comparable. Own elaboration.

The weights concentrate on the two monthly dummy specifications, and the averaged coefficient is 1.61 with a standard error of 4.18. With a standard deviation of Pix growth of 0.096 in the primary sample, a one-standard-deviation change in Pix growth is associated with a change of about 0.15 IBC-Br index points, with an approximate 95 percent interval from −0.63 to 0.94 points, against a standard deviation of the monthly change in the IBC-Br of 4.07 points. Both procedures therefore lead to the same answer: under a principled selection rule, the economic magnitude of the Pix effect is small and not distinguishable from zero. The larger coefficients of the specifications without monthly controls (about 25, or about 2.4 index points per standard deviation) receive negligible weight because those specifications fit and forecast considerably worse.

4.7. Managerial and Financial-Market Implications

Although the Selic rate transmission channel discussed in Section 3.1 situates this paper’s motivation within a monetary policy framing, this journal’s readership includes a substantial share of financial-market practitioners for whom the practical relevance of these findings deserves brief, explicit comment.

The results do not support using aggregate Pix growth as a monthly nowcasting input for the IBC-Br at the current sample length. A corporate treasurer, credit risk manager, or equity analyst who regresses activity on raw Pix growth will find a strong and apparently significant relationship, and will even find modest out-of-sample gains, but only in specifications that leave the calendar component of both series uncontrolled; a simple monthly seasonal model without Pix forecasts more accurately than any of those specifications. The practical lesson is procedural and extends to other payment-based indicators: before such an indicator is used to inform decisions, its calendar component should be removed, its contribution should be evaluated against a seasonal benchmark rather than a naive one, and nested-model test results should be checked against estimation windows long enough for the larger model to be estimated reliably.

One practical advantage of the framework carries over regardless of the negative result: because the predictive regressions are fully disclosed and re-estimable from public data with ordinary least squares, the evaluation can be repeated at low cost as the Pix sample lengthens, by disaggregated transaction type, or for other emerging economies with instant payment systems at an earlier stage of implementation (International Monetary Fund, 2023).

5. Conclusions

This paper tested whether the volume of Pix transactions improves the monthly nowcasting of the IBC-Br, with the degree of rigor a financial-market practitioner should demand before deploying such a signal. The answer, based on data available through May 2026, is that it does not, once the calendar component shared by Pix and the unadjusted IBC-Br is controlled.

The evidence is asymmetric rather than balanced. Pix growth is strongly and significantly associated with the IBC-Br, in-sample and out-of-sample, only in specifications that leave month-of-year seasonality uncontrolled, including those with quarter-of-year dummies. Those specifications also leave non-stationary residuals and forecast the IBC-Br considerably less accurately than a monthly seasonal model without Pix. Once monthly seasonality is controlled, either with month-of-year dummies or with the officially seasonally adjusted series, the Pix coefficient becomes small and statistically insignificant, or takes the opposite sign to an activity signal, and adding Pix does not improve out-of-sample accuracy at any estimation window, including windows long enough for every model to be estimated with adequate degrees of freedom. A significant nested-model test result in favor of Pix appears only in nearly saturated early windows of the sample that includes the launch month. A recursive model-selection rule never selects a specification with Pix, and the model-averaged effect is about 0.15 index points per standard deviation of Pix growth, with an interval that includes zero. These conclusions are unchanged by controls external to the IBC-Br and by simulated data revisions well above those observed between two vintages of the series.

This negative result carries three implications. Academically, it shows that, for a payment indicator with a strong calendar pattern, a significant in-sample coefficient and a significant nested-model test can both arise from shared seasonality, and that a seasonal benchmark, rather than a naive one, is the relevant standard of comparison for payment-data nowcasts. Institutionally, it qualifies, for the aggregate monthly series, the expectation expressed by the Central Bank of Brazil (2025) in the technical box introducing the EPAE statistic that Pix payment flows may carry complementary information about economic activity: any such information is not visible in aggregate Pix volume once calendar effects are removed, although it may be present in disaggregated or sectoral payment data. Managerially, as discussed in Section 4.7, it cautions against using raw Pix growth as an activity signal in credit, treasury, or investment decisions.

The main limitation of the study is its sample size (65 monthly observations in the primary sample), inherent to the recency of the system. It limits the power of every test reported in Section 4 and the precision with which monthly seasonal effects can be estimated, so the conclusion should be read as the absence of detectable predictive content at the current sample length rather than as proof that none exists. The exercise is also pseudo-real-time, since public first-release vintages of the IBC-Br are not available, although the bounds reported in Section 3.3 indicate that revisions of a plausible size do not change the conclusions.

As a priority for future work, the following extensions would test whether a signal exists that the aggregate series does not reveal:

(i)

Assembling a first-release vintage dataset of the IBC-Br from the BCB’s revision history, to replace the pseudo-real-time evaluation with a genuine real-time one.

(ii)

Decomposing Pix by transaction type (P2P, P2B, B2B) as separate regressors, exploiting the disaggregated database already available, to test whether a specific, less calendar-sensitive transaction segment concentrates any genuine predictive signal.

(iii)

Running formal cointegration tests between the IBC-Br and Pix once the sample grows sufficiently for adequate test power.

(iv)

Incorporating the BCB’s Payments-by-Economic-Activity (EPAE) series, available from October 2025 onward, as an additional source of sectoral disaggregation, in line with the approach of Hötte (2025) for inter-sectoral payment data in the United Kingdom.

(v)

Extending the evaluation to forecast horizons longer than one month, following the practice of Aprigliano et al. (2019) and Gálvez-Soriano (2020), once the sample supports it.

(vi)

Representing seasonality through a small number of structural parameters, in the spirit of the CIR-type models discussed in Section 2.2 (Ceci et al., 2024; Orlando & Bufalo, 2021, 2023), as a more parsimonious alternative to the monthly indicator variables used in this paper.

(vii)

Exploring a genuinely higher-frequency design using weekly or daily Pix data, which the API in principle supports, as a more direct test of timeliness than the monthly design used throughout this paper.

These extensions would not reverse the finding for aggregate monthly Pix volume documented here, but they would establish whether the absence of predictive content is a property of the aggregate series or of payment data more generally. Until they do, we do not recommend using aggregate Pix growth as a monthly nowcasting input for Brazilian economic activity.