Measuring Consumption with Credit Card Data: Benchmarking and Beyond

Researchers have created a new way to track how much people spend in every U.S. county every month using credit card data. This method is much faster and more detailed than traditional government statistics. It helps us see how different groups of people react to economic changes like interest rate hikes.

In macroeconomics, understanding consumption—the total spending by households—is essential for tracking the health of an economy. Traditionally, policymakers rely on official statistics like Personal Consumption Expenditures (PCE). These are comprehensive but suffer from significant delays and coarse geographic detail. While researchers use private-sector transaction data to fill these gaps, these datasets often lack the representative scale required for national accounts. This paper addresses that gap by leveraging regulatory data to build a high-frequency, granular map of American spending.

The resolution gap in national accounts

The current gold standard for measuring consumption, the National Income and Product Accounts (NIPA), provides a complete picture but suffers from two main flaws. First, it has low frequency, meaning data arrives with a significant time lag. Second, it has low geographic granularity, meaning it lacks the detail to show what is happening in specific local areas.

Because official statistics are released with delays, they struggle to capture rapid economic shocks. They also struggle to show how those shocks ripple through specific local communities. Furthermore, traditional measures often include "sticky" components. These are expenditures like housing costs or insurance-paid healthcare that do not behave like everyday discretionary spending. This creates a mismatch when researchers use high-frequency data to predict slower, aggregate economic indicators.

Building a granular consumption proxy

To solve this, the authors utilize the Federal Reserve’s Y-14M reports. This is a massive regulatory dataset containing monthly account-level information for over 350 million credit cards. Unlike many private datasets that focus on where a transaction occurs, the authors use the cardholder's billing zip code. This attributes spending to where the person lives. This effectively transforms the data into a household-based measure.

The construction process follows a specific logic to ensure comparability:

1. Filtering for active users: The authors focus only on cards with nonzero purchase volume. This avoids diluting averages with dormant accounts.
2. Creating an "Adjusted PCE": The authors construct a custom benchmark. They take the official PCE and subtract categories unlikely to be charged to a card. These include motor vehicles, housing, and healthcare paid by insurance.
3. Scaling by averages: The authors measure the growth in average spending per card. This prevents the data from being skewed by the sheer number of new cards entering the market.

This architecture allows the dataset to function as a high-fidelity proxy. It targets the parts of the economy that move most quickly in response to financial changes.

Validating the 92% correlation

The authors report that this credit card-based measure is remarkably accurate at the national level. In a regression analysis, the growth in average credit card spending explained 92% of the variation in monthly adjusted PCE growth [Table 2]. This means the credit card data tracks the core movement of national consumption very closely. Even during the COVID-19 pandemic, the credit card series tracked the sharp declines and recoveries seen in official retail sales .

The utility of the data extends far beyond national trends. The study shows the dataset captures meaningful geographic variation. It aligns closely with the U.S. Economic Census at the county level .

Most importantly, the authors use the data to study the effects of monetary policy. They look at the impulse response, which is the predicted change in spending following a specific economic event.

They focused on contractionary monetary policy (a policy where the central bank raises interest rates to slow the economy). They found that when rates rise, spending in low-income counties drops by approximately 9% at its trough. In contrast, high-income counties see a smaller decline of 7% . This level of detail was previously impossible with traditional public data.

Limits of the transaction lens

Despite its strength, the authors define the boundaries of what this data can do. The Y-14M reports provide monthly account totals rather than individual line items. Therefore, the data cannot distinguish between spending categories. You can see how much a county spent, but not if they bought groceries or electronics. The authors warn that the data is best for studying broad consumption dynamics rather than shifts between specific goods and services .

There are also structural biases to consider. Credit card spending disproportionately captures discretionary, interest-sensitive expenditures. This makes the data more volatile than total PCE. It acts like a magnifying glass for economic shifts. This is useful for research but requires careful calibration. Additionally, the relationship between card spending and consumption varies by category. Researchers must choose their benchmarks with precision to avoid misinterpretation.

The verdict: A powerful tool for local economics

The Y-14M dataset is ready for high-stakes research. It is not a direct replacement for the PCE. Instead, it is a high-resolution microscope for the discretionary economy. If your goal is to study national aggregates, stick to the official accounts. However, if you need to understand how economic shocks propagate through specific neighborhoods, this dataset is a significant upgrade.

For practitioners, the takeaway is clear. When modeling consumption with card data, always use an "adjusted" benchmark. This benchmark should exclude non-card-payable categories like housing and insurance. If you ignore this, you will likely misinterpret the heightened volatility of the card data. You might mistake natural sensitivity to interest rates for a measurement error.

Figures from the paper