High-Frequency Economic Index (HFEI) Overview

• High-Frequency Economic Index (HFEI) is a real-time statistical measure combining multi-frequency economic data to track underlying business conditions.
• It employs mixed-frequency dynamic factor models, Bayesian techniques, and state-space approaches to reconcile data with varying publication lags.
• HFEI enables timely nowcasting and informed policy-making by providing interim economic signals for regional, national, and sectoral analysis.

A High-Frequency Economic Index (HFEI) is a statistical indicator designed to track underlying economic activity in real time by combining variables observed at different frequencies and with different publication lags. In the recent literature, the term covers several related constructions: quarterly regional indices inferred from annual regional gross value added (GVA) and quarterly national indicators, pseudo-weekly national indicators estimated from weekly, monthly, and quarterly official statistics, daily latent-factor measures of business conditions, and granular prediction systems trained on low-frequency labels and high-frequency covariates (Barbaglia et al., 2024, García-Albán et al., 10 Jul 2025). Across these formulations, the common objective is to extract timely signals on output, business conditions, or recession risk when official aggregates are sparse, delayed, or asynchronously released.

1. Scope and analytical role

The central motivation for HFEI construction is the mismatch between policy needs and official statistical release schedules. European regional accounts for output are published at an annual frequency and with a two-year delay, while official aggregates such as GDP are typically available only quarterly and with long publication lags (Barbaglia et al., 2024, García-Albán et al., 10 Jul 2025). HFEIs therefore provide interim measures of economic conditions for planning, implementing and evaluating locally targeted economic policies, for real-time macroeconomic monitoring, and for business-cycle dating.

Representative implementations differ in geography, target frequency, and data architecture, but they share the aim of synthesizing dispersed information into a single coherent indicator.

Framework	Frequency and scope	Core mechanism
Regional HFEI (Barbaglia et al., 2024)	Quarterly estimates for 162 regions in 12 European countries	Mixed-frequency dynamic factor model with annual regional GVA and quarterly national indicators
Ecuador HFEI (García-Albán et al., 10 Jul 2025)	Pseudo-weekly national indicator	Bayesian Dynamic Factor Model with weekly, monthly, and quarterly official indicators
ADS Index (Diebold, 2020)	Daily U.S. real activity	State-space dynamic factor model with six mixed-frequency indicators
CFNAI-style imputation (Ng et al., 2023)	Weekly analog of a monthly diffusion index	Imputation of missing high-frequency values using factors and dynamic state-space smoothing
MF-AGL (Toda et al., 2021)	Real-time prediction for smaller areas	Aggregate learning from high-frequency granular covariates and low-frequency aggregate labels

The empirical targets also vary. In the European regional framework, the latent quarterly series is a high-frequency estimate of regional GVA (Barbaglia et al., 2024). In the Ecuador application, the HFEI is additionally converted into pseudo-weekly recession probabilities via a time-varying mean regime-switching model (García-Albán et al., 10 Jul 2025). In the ADS framework, the latent factor measures “good growth,” with positive values above average growth and negative values below average (Diebold, 2020). This suggests that “HFEI” is best understood as a methodological class rather than a single canonical index.

2. Statistical formulations

A dominant formulation is the linear Gaussian state-space model. In the regional mixed-frequency dynamic factor model of Barbaglia et al., all regional GVA observations are stacked in an $N \times 1$ vector $y_t^*$ at quarterly frequency, and the system is written in canonical form as

$$y_t^* = \Lambda_t f_t + \epsilon_t, \qquad f_{t+1} = \Phi f_t + \eta_t,$$

where $f_t = (\tilde f_t', z_t')'$ contains unobserved “regional” factors $\tilde f_t$ and observed national or common indicators $z_t$ such as quarterly national GVA growth, interest rates, and unemployment (Barbaglia et al., 2024). The model also imposes a cross-sectional restriction linking observed national GVA growth to the output-share-weighted sum of regional factor loadings, so that quarterly regional nowcasts are consistent with the national number up to a small measurement error (Barbaglia et al., 2024).

The Ecuador HFEI is a single-factor Bayesian Dynamic Factor Model estimated on a pseudo-weekly grid. Weekly variables load contemporaneously on a latent weekly factor $x_t$, monthly variables load on the average of the last four latent weekly states, and quarterly variables load on the average of the last twelve latent weekly states:

$$y^m_{j,t} = \frac{\lambda_j}{4}\sum_{s=0}^{3}x_{t-s} + u^m_{j,t}, \qquad y^q_{i,t} = \frac{\lambda_i}{12}\sum_{s=0}^{11}x_{t-s} + u^q_{i,t}.$$

The factor follows an AR(1) with stochastic volatility, and the log-variance evolves as a random walk; idiosyncratic variances are modeled analogously (García-Albán et al., 10 Jul 2025). This architecture is intended to accommodate dynamic heterogeneity and heightened uncertainty, including COVID-19.

The ADS Index adopts a closely related state-space logic at daily frequency. It is the Kalman-smoothed estimate of a one-dimensional latent factor $f_t$ in an observation equation $y_t=\mu+\Lambda f_t+\epsilon_t$ and state equation $y_t^*$0, where mixed-frequency indicators are observed at their native release dates and missing on other days (Diebold, 2020). Monthly and quarterly flows are embedded in a “flow” state-space framework that aggregates them into daily units via summation matrices (Diebold, 2020).

Alternative constructions depart from direct latent-factor estimation. MF-AGL specifies a non-linear predictor

$y_t^*$1

with low-frequency supervision imposed through an aggregate loss over areas and periods (Toda et al., 2021). The imputation framework of Ng and Scanlan instead targets a low-frequency diffusion index that is already available, treats high-frequency values as missing, and reconstructs them using multiple factors estimated from the high-frequency data (Ng et al., 2023).

3. Temporal aggregation, synchronization, and identification

A central technical problem is reconciling frequency mismatch without introducing aggregation artifacts. In the regional GVA model, annual releases are linked to quarterly latent factors through the triangular weights that arise from converting quarter-on-quarter rates into annual rates, following Mariano and Murasawa. For each latent factor, the last seven quarterly values enter with weights $y_t^*$2, $y_t^*$3, $y_t^*$4, $y_t^*$5, $y_t^*$6, $y_t^*$7, and $y_t^*$8 (Barbaglia et al., 2024). In quarters when regional GVA is not published, the observation is treated as missing.

The Ecuador framework addresses a different synchronization problem. Direct use of calendar weeks produces time-varying aggregation matrices and irregular missing-data patterns, so the model defines four pseudo-weeks per calendar month: days 1–7 map to Week 1, 8–14 to Week 2, 15–21 to Week 3, and 22–End to Week 4 (García-Albán et al., 10 Jul 2025). For stock variables such as credit and deposits, the within-pseudo-week average is used; for flow variables such as new firm registrations and trade flows, the pseudo-weekly sum is used (García-Albán et al., 10 Jul 2025). The result is exactly 48 pseudo-weeks per year and 12 per quarter, making the loading matrix time-invariant.

The ADS approach handles mixed frequencies through native missingness: lower-frequency series are simply absent on most days and are handled by the Kalman filter and smoother (Diebold, 2020). The imputation literature uses another alignment device: if a low-frequency period contains $y_t^*$9 high-frequency observations, then the low-frequency factor is defined as the average $$y_t^* = \Lambda_t f_t + \epsilon_t, \qquad f_{t+1} = \Phi f_t + \eta_t,$$0, and the observed low-frequency diffusion index is linked to this aggregate factor (Ng et al., 2023).

Identification is likewise framework-specific. In the regional Bayesian model, factor normalization is implicit: no loadings are fixed to unity; instead, a tight prior on the measurement-error variance anchors the triangular restriction, and the national GVA equation pins down scale and sign (Barbaglia et al., 2024). In MF-AGL, identification comes from spatial aggregation constraints: only aggregate labels are observed, but model outputs at the small-area level must reproduce the large-area target under known weights $$y_t^* = \Lambda_t f_t + \epsilon_t, \qquad f_{t+1} = \Phi f_t + \eta_t,$$1 (Toda et al., 2021).

4. Estimation, filtering, and real-time updating

HFEIs are typically estimated with Kalman-based state inference combined with Bayesian or likelihood-based parameter learning. In the regional mixed-frequency model, missing regional observations are naturally accommodated via the Kalman filter, and posterior computation proceeds by Gibbs-style MCMC. The procedure draws VAR coefficients and loadings one row or equation at a time under Horseshoe priors, draws $$y_t^* = \Lambda_t f_t + \epsilon_t, \qquad f_{t+1} = \Phi f_t + \eta_t,$$2 and $$y_t^* = \Lambda_t f_t + \epsilon_t, \qquad f_{t+1} = \Phi f_t + \eta_t,$$3 from inverse-Gamma conditionals, conducts a rejection-sampling step to enforce stationarity of $$y_t^* = \Lambda_t f_t + \epsilon_t, \qquad f_{t+1} = \Phi f_t + \eta_t,$$4, and then runs the Kalman filter and smoother to sample the latent factors (Barbaglia et al., 2024). The prior structure is explicit: local parameters satisfy $$y_t^* = \Lambda_t f_t + \epsilon_t, \qquad f_{t+1} = \Phi f_t + \eta_t,$$5, the diagonal elements of $$y_t^* = \Lambda_t f_t + \epsilon_t, \qquad f_{t+1} = \Phi f_t + \eta_t,$$6 satisfy $$y_t^* = \Lambda_t f_t + \epsilon_t, \qquad f_{t+1} = \Phi f_t + \eta_t,$$7, and measurement errors satisfy $$y_t^* = \Lambda_t f_t + \epsilon_t, \qquad f_{t+1} = \Phi f_t + \eta_t,$$8 (Barbaglia et al., 2024).

The Ecuador BDFM is also estimated by Gibbs sampling. Static parameters and stochastic-volatility innovations are collected in $$y_t^* = \Lambda_t f_t + \epsilon_t, \qquad f_{t+1} = \Phi f_t + \eta_t,$$9, the latent factor and idiosyncratic components are drawn jointly via the Carter–Kohn or Durbin–Koopman simulation smoother, loadings are drawn by GLS, the AR coefficient is drawn from a conditional Normal posterior subject to $f_t = (\tilde f_t', z_t')'$0, and stochastic-volatility paths are updated using the Kim–Shephard–Chib algorithm (García-Albán et al., 10 Jul 2025). The Kalman filter within each iteration handles the missing pseudo-weekly observations implied by monthly and quarterly series (García-Albán et al., 10 Jul 2025).

Real-time updating protocols are integral to the HFEI concept. In the European regional application, each new batch of national indicators triggers a new Kalman filtering step to produce the one-step-ahead predictive distribution of the next three quarterly regional values, which correspond to the remaining quarters of the current year; when annual GVA for year $f_t = (\tilde f_t', z_t')'$1 is released, the model is re-estimated to backcast $f_t = (\tilde f_t', z_t')'$2 and nowcast $f_t = (\tilde f_t', z_t')'$3 (Barbaglia et al., 2024). The ADS Index is updated within two hours of any new data release and maintains every real-time vintage produced since December 2008, enabling vintage-honest nowcasting analysis (Diebold, 2020). MF-AGL adopts a different operational logic: once the parameters are trained up to a given low-frequency period, new high-frequency data can be fed through the network to obtain a partial nowcast, with fine-tuning deferred until new low-frequency labels arrive (Toda et al., 2021).

5. Empirical implementations and performance

The European regional model is estimated on annual NUTS2 regional GVA published by Eurostat, covering 162 regions in 12 countries from 2001–2021, together with quarterly national GVA growth and optional additional national indicators (Barbaglia et al., 2024). Evaluation holds out annual regional GVA for 2016–2021 and respects the real-time release calendar. Predictive accuracy is measured by the Continuous Ranked Probability Score (CRPS), aggregated with GVA-share weights within each country, against a random-walk benchmark on regional annual GVA growth (Barbaglia et al., 2024). Reported results are strong: optimal numbers of factors range from 1 to 7 and satisfy roughly $f_t = (\tilde f_t', z_t')'$4; average CRPS reductions versus the random walk are up to 60% in Greece, about 40% in Slovakia, and about 15–20% in Austria, Belgium, and France; on 97% of regions the model beats the random walk; and the advantage is stable over time, including the COVID shock (Barbaglia et al., 2024).

The Ecuador HFEI uses 4 quarterly variables, 5 monthly variables, and 8 weekly variables over 2004–2024, with publication lags up to 105 weeks for GDP (García-Albán et al., 10 Jul 2025). Model comparison by DIC over eight specifications favored “stochastic volatility in the factor only + dynamic heterogeneity” (García-Albán et al., 10 Jul 2025). The resulting HFEI tracks four major downturns—the 2008–09 oil-price crisis, the 2015–16 oil shock, the COVID-19 collapse, and the 2024 slowdown—and responds within 1–2 pseudo-weeks of observed shocks (García-Albán et al., 10 Jul 2025). In the second-stage regime-switching model, recession probabilities exceed 0.8 at each trough and fall below 0.2 during expansions, while a 0.5 threshold aligns well with known episodes (García-Albán et al., 10 Jul 2025).

The ADS Index provides a daily U.S. benchmark for high-frequency real-activity monitoring. During the 2020 pandemic recession, real-time vintages first under-reacted, then plunged ADS to historic lows, and later converged as new information arrived. Later-vintage ADS bottomed near $f_t = (\tilde f_t', z_t')'$5 in early April, returned to zero by early May, and climbed to $f_t = (\tilde f_t', z_t')'$6 by late June; real-time vintages hit $f_t = (\tilde f_t', z_t')'$7 on March 26, $f_t = (\tilde f_t', z_t')'$8 on April 2, and $f_t = (\tilde f_t', z_t')'$9 on May 8, before turning positive on June 5 after the strong May payrolls release (Diebold, 2020). The later-vintage daily ADS series is strongly negatively correlated with HP-smoothed daily COVID-19 deaths led by 20 days, with absolute correlation exceeding 0.8 over March–June 2020 (Diebold, 2020).

The imputation literature reports a clear ranking across reconstruction strategies. In held-out data, static matrix-completion imputation has RMSE about 0.18, dynamic single-equation AR(1) imputation about 0.14, and the full state-space smoother about 0.12 (Ng et al., 2023). When aggregated back to monthly frequency, the dynamic smoother’s high-frequency imputations are almost indistinguishable from the official CFNAI, with correlation greater than 0.98 (Ng et al., 2023). The weekly imputed series reveals variability around the 2014–15 oil-price collapse that is masked in the monthly index (Ng et al., 2023).

6. Extensions, alternatives, and methodological issues

The literature presents several non-equivalent strategies for HFEI construction. One path is direct mixed-frequency latent-factor modeling with measurement restrictions and Kalman filtering, as in the regional European and ADS frameworks (Barbaglia et al., 2024, Diebold, 2020). A second path is mixed-frequency Bayesian factor modeling with stochastic volatility and dynamic heterogeneity on a harmonized pseudo-week grid (García-Albán et al., 10 Jul 2025). A third path is imputation: target a low-frequency diffusion index, treat high-frequency values as missing, and reconstruct them with factors and dynamic idiosyncratic structure (Ng et al., 2023). A fourth path is aggregate learning, in which low-frequency labels supervise a high-frequency predictor that is trained at a more granular spatial scale (Toda et al., 2021).

Several methodological issues recur. First, real-time honesty matters. The ADS work emphasizes the preservation of every vintage, and the European regional evaluation respects the real-time release calendar; both choices are intended to prevent ex post information from contaminating assessment (Diebold, 2020, Barbaglia et al., 2024). Second, serial correlation in idiosyncratic components is consequential. Static matrix completion that ignores serial correlation yields imprecise estimates of missing high-frequency values irrespective of how the factors are estimated, whereas single-equation and systems-based dynamic procedures yield imputations closer to observed low-frequency aggregates (Ng et al., 2023). Third, temporal harmonization is not innocuous. The Ecuador paper argues that direct use of calendar weeks produces irregular aggregation artifacts, and uses pseudo-weeks specifically to keep the loading matrix time-invariant (García-Albán et al., 10 Jul 2025).

The extension agenda is explicit. Proposed developments for the regional mixed-frequency dynamic factor model include pooling all 162 regions in a single continent-wide MF-DFM with global and country-specific factors, allowing time-varying parameters or stochastic volatility in $\tilde f_t$0, incorporating mixed-frequency MIDAS-type covariates such as daily financial indicators in $\tilde f_t$1, tracing structural shocks and their regional pass-through, and applying the framework to U.S. states, Chinese provinces, or city-level indicators (Barbaglia et al., 2024). A plausible implication is that HFEIs are evolving from stand-alone nowcasting devices into broader empirical infrastructures for regional decomposition, recession dating, and structural analysis under severe data constraints.