Kinane Habra, Matthias Fengler, Winfried Koeniger, Jonas Bruhin
2026-06-17
Official retail statistics are reliable, but they are available only at a low frequency and often with a delay because they rely on surveys. By the time these data are released, much of the information they contain is already outdated for policymakers and analysts who require timely assessments of current economic conditions. This raises the question of whether high-frequency payment data can improve real-time nowcasting of retail activity. To address this issue, we combine official Swiss retail data with the Consumer Spending Index (CSI), which aggregates millions of anonymized card transactions from Worldline into a timely indicator of consumer spending in Switzerland.
The primary measure of retail trade turnover in Switzerland is the DHU. For this nowcasting experiment, we use monthly growth rates from the seasonally adjusted series for total retail trade (NOGA 47). As a benchmark, the analysis employs a moving average process of order one, in short: MA(1), as suggested by standard model-selection procedures applied to an initial calibration period. Two alternative nowcasting approaches are then considered. The first is a state-of-the-art dynamic factor model that relies on macroeconomic indicators available only at the monthly frequency. The second augments this model with high-frequency transactional payment data from the CSI. The analysis evaluates whether including payment data yields economically meaningful improvements in nowcasting performance.1
We evaluate forecast accuracy using “skill.” Skill compares the forecast error of a model of interest with the forecast error of a benchmark model.2 All results are reported relative to the MA(1) benchmark, allowing for a clear comparison across models: Positive skill values indicate improvements over the benchmark, while negative values indicate inferior performance. We report results for the full sample (March 2019–September 2025), the COVID period (March 2019–June 2021), and the post-COVID period (July 2021–September 2025). This breakdown allows us to assess whether transactional payment data add value primarily during periods of economic stress or also under more normal conditions.
Figure 1 summarizes forecast accuracy across models and subsamples. Models that rely only on macroeconomic indicators deliver modest improvements during the COVID period but underperform relative to the benchmark in both the full sample and the post-COVID period. This suggests that improving upon agnostic time-series benchmarks using additional low-frequency indicators is difficult.
By contrast, augmenting the model with transactional payment data yields substantially larger and more consistent gains across all subsamples. As shown in the figure, incorporating high-frequency payment data improves forecast accuracy by 35% during the COVID period. In the full sample, the augmented model is 28% more accurate than the benchmark. This shows that these gains are largely driven by the COVID period, when spending patterns shifted abruptly and high-frequency payment data provided timely information not captured by monthly aggregates. Importantly, the model incorporating payment data continues to outperform the benchmark in the post-COVID period, with an improvement of 11%. From a practical perspective, this indicates that high-frequency transactional data can improve short-term tracking of retail activity also in a stable economic environment.
Figure 1: Out-of-sample improvement in MAE relative to an MA(1) benchmark, by model and subsample.
(Full sample: March 2019–September 2025; COVID: March 2019–June 2021; Post-COVID: July 2021–September 2025).
To obtain further insights into predictive behavior, we compare in Figure 2 observed retail growth with nowcasts from the MA benchmark (shown in red) and from the two alternative models, with and without transactional payment data from the CSI. The visual evidence corroborates the quantitative results of Figure 1. During periods of economic stress, particularly in 2020 and 2021, the MA benchmark responds slowly and frequently understates the magnitude of movements. The model based solely on macroeconomic indicators performs similarly and does not deliver meaningful improvements over the benchmark. By contrast, the model incorporating transactional payment data captures sharp declines and subsequent rebounds in retail activity more accurately, providing a smoother and more timely signal of the underlying dynamics.
To assess whether these differences are economically meaningful, forecast errors can be evaluated relative to both the level of retail activity and its typical monthly growth rate. Over the full sample, the mean absolute error corresponds to around 0.025% of the average level of DHU for the MA benchmark, compared with 0.018% for the model incorporating transactional payment data. Thus, forecast errors are small when measured relative to the overall level of the index. At the same time, monthly changes in DHU are modest in magnitude, implying that even small forecast errors can be economically relevant when accumulated over time. When scaled by the average monthly growth rate, the MA benchmark’s error corresponds to roughly 16 months of average growth, whereas the payment-data-augmented model reduces this to about 11 months. In economic terms, this represents a reduction in forecast error equivalent to approximately four to five months of trend growth over the sample period.
In practice, the reduction in forecast errors translates into a more reliable real-time signal for monitoring retail activity ahead of official data releases. Transactional payment data do not replace official statistics but provide a timely complement for short-term monitoring, particularly when economic conditions change rapidly. They improve real-time tracking of retail activity across a range of economic conditions.
Figure 2: Retail trade turnover growth (DHU) and nowcasts based on macroeconomic and transactional indicators, with an MA(1) as a benchmark.
Reference
Giannone, D., Reichlin, L., and Small, D. (2008). Nowcasting: The real-time informational content of macroeconomic data. Journal of Monetary Economics, 55(4):665–676.
Habra, K. (2025). Retail trade nowcasting using high-frequency transactional payment data. Master’s thesis, University of St. Gallen, St. Gallen, Switzerland. Unpublished. Supervisors: Matthias Fengler and Lyudmila Grigoryeva