Evidence Pack: The Reliability of Nowcasting Economic Indicators Using Alternative Data

The fundamental problem of economic policymaking is that it is conducted in the dark. Every major decision—whether a central bank raising interest rates to cool inflation or a government deploying fiscal stimulus to avert a recession—is based on data that describes the past. Official Gross Domestic Product (GDP) figures are typically released 30 to 45 days after a quarter ends, while comprehensive employment and inflation statistics carry their own multi-week delays.

By the time an institution knows what the economy was doing, the macroeconomic reality has often already shifted. This delay, known in econometrics as the "ragged edge" problem, forces policymakers to steer the global economy by looking in the rearview mirror. To close this gap, economists have increasingly turned to a technique borrowed from meteorology: nowcasting.

Where forecasting attempts to predict the future, nowcasting attempts to predict the present. It means estimating the current state of indicators like GDP and inflation before official statistics are published, utilizing high-frequency data that is available immediately. Historically, this relied on slightly faster traditional indicators, such as monthly industrial production reports or purchasing managers' indices (PMIs).

However, the economic shocks of the 2020s exposed the limits of these conventional proxies, which still suffer from publication lags and survey biases. This has triggered a structural shift toward alternative data. Today, the most accurate nowcasting models ingest massive volumes of non-traditional information: satellite imagery of nighttime lights, real-time credit card transactions, global vessel tracking, and internet search trends.

The evidence for this shift is robust, particularly in regions where official statistics are sparse or unreliable. A January 2026 working paper from the International Monetary Fund (IMF) demonstrated that integrating satellite-based nightlight data into predictive models significantly improves the accuracy of quarterly GDP growth estimates in data-constrained economies.[1]

Because nighttime light intensity correlates strongly with human economic activity—manufacturing, transportation, and commercial operations all generate light—satellites provide an unbiased, real-time proxy for output that bypasses the delays of national statistical offices. The IMF study found that introducing satellite data improved the accuracy of their Random Forest models by 13.85 percent.[1]

Beyond physical infrastructure, digital financial footprints offer some of the strongest predictive signals for consumer behavior. Research evaluating the use of electronic payments data—specifically debit and credit card transactions clearing through retail settlement systems—found that these models bear an accurate and timely signal about consumer demand and overall economic activity.[4]

When evaluated against revised GDP numbers, models incorporating high-frequency payments data consistently outperform benchmark macroeconomic models. Crucially, the timeliness of this transaction data improves nowcasting accuracy throughout the quarter, allowing central banks to detect shifts in consumption weeks before official retail sales figures are published.[4]

Text-based alternative data is also proving highly effective for economic monitoring. An exploratory initiative by the Asian Development Bank (ADB) showed that applying machine learning to millions of news articles could generate highly accurate economic nowcasts. By utilizing word co-occurrence networks and sentiment analysis, the ADB was able to reduce the margin of error for GDP estimates to just 0.36 percentage points.[5]

This textual analysis allows economists to quantify the economic mood in real-time, capturing shifts in labor, trade, and investment long before they materialize in quarterly national accounts. The sheer volume and complexity of this alternative data, however, mean it is often incompatible with traditional econometric models like Ordinary Least Squares (OLS) regressions.[5]

High-frequency alternative data is inherently noisy, highly correlated, and non-linear. To extract the signal from the noise, researchers rely on advanced machine learning techniques. Random Forests, Neural Networks, and Dynamic Factor Models (DFMs) are now standard tools, capable of handling thousands of variables without overfitting the model to historical noise.[1]

The European Central Bank (ECB) has successfully applied these techniques to global trade. By augmenting their nowcasting toolkit with real-time satellite data on vessel movements—using the automatic identification system (AIS) that tracks ship positions daily—the ECB significantly improved its ability to capture shifts in global trade dynamics.[2]

The augmented tracker particularly outperformed previous models during periods when financial market variables diverged from physical trade dynamics. For example, in 2022, stock markets fell amid surging inflation and geopolitical shocks, but global trade remained resilient due to the easing of supply bottlenecks. The satellite data correctly identified this decoupling, preventing the model from issuing an erroneous recession signal.[2]

Despite these advances, the evidence highlights critical limitations in the nowcasting framework. The most significant vulnerability is the "structural break"—a sudden, unprecedented shock to the economic system, such as a pandemic lockdown or a geopolitical conflict. Nowcasting models are designed to track the economy along its current trajectory based on historical correlations; they cannot predict black swan events.[6]

When a structural break occurs, the historical relationship between the alternative data proxy and the macroeconomic indicator often breaks down. This modality dependence was starkly illustrated in a March 2026 study evaluating the use of highway traffic data to nowcast GDP.[3]

While physical mobility is generally a reliable proxy for economic activity, the study found that using aggregated highway traffic volumes actually worsened forecast accuracy during the COVID-19 shock, increasing the Root Mean Squared Error (RMSE) compared to a macro-only benchmark. The raw data failed to distinguish between essential freight movement and halted passenger travel.[3]

Only when the traffic data was disaggregated to isolate specific freight vehicle types did the signal regain its predictive value. This underscores a vital lesson for modern econometrics: raw alternative data is not a panacea. It requires careful curation and an understanding of the underlying physical mechanisms it represents.[3]

Ultimately, alternative data and machine learning are not replacing official statistics; they are augmenting them. The most effective nowcasting frameworks operate as a hybrid, anchoring real-time digital signals to the rigorous, albeit delayed, foundation of national accounts.[6]

As central banks and financial institutions continue to integrate these tools, the latency of economic measurement will continue to shrink. The goal is no longer just to measure the economy accurately, but to measure it fast enough to actually do something about it.[6]