Tracking Ecuador’s economy between GDP releases

How do you track an economy when its statistics arrive on different schedules, with different delays? Quarterly GDP provides a broad picture, but it can leave analysts waiting months for confirmation of a downturn. Weekly financial records offer more frequent observations, yet may tell a different story from employment, production or sales.

In the preprint working paper Tracking the economy at high frequency, Yanina Bajaña-Villagómez and colleagues develop a High-Frequency Economic Index, or HFEI, for Ecuador. Their aim is to summarize economic conditions using official statistics rather than rely on any single measure. The study examines historical data through the final quarter of 2024; it is not a reading of Ecuador’s economy today.

Bringing different clocks together

The researchers assemble 17 series spanning 2004–2024, although several begin later. Four are quarterly, including real GDP, household consumption and two credit-survey indexes. Five are monthly, covering sales, employment, expectations and oil production. Eight are weekly, covering business registrations, credit, deposits and trade.

Observation frequency and publication speed are different things. The paper reports an average publication lag of 105 days for GDP and household consumption. Trade records contain daily transaction dates, but are released monthly. Reconstructing weekly trade flows therefore adds historical detail without making those records available sooner.

The authors organize their data into “pseudo-weeks”: days 1–7, 8–14, 15–21, and 22 through the end of each month. Every month consequently contains four periods, every quarter 12 and every year 48. The final period varies in length, so these are not ordinary calendar weeks.

This arrangement simplifies the mathematical connections between weekly, monthly and quarterly observations. Flows, such as trade transactions, are summed within each period; stock measures, such as credit balances, are averaged over available observations.

Finding a shared economic signal

The core method is a Bayesian dynamic factor model. A “factor” is an unobserved signal inferred from movements shared across the indicators. Each series also has its own component, allowing, for example, deposits to behave differently from sales.

The model can accommodate indicators that respond with a delay to the common signal. It can also allow volatility—the size of unpredictable movements—to change over time. Bayesian estimation combines the data with prior assumptions that favor simpler parameter settings unless the evidence supports additional complexity.

Important assumptions remain. The framework uses one common factor and assumes that the series-specific components are mutually uncorrelated. Some input observations are reconstructed: missing historical employment values are estimated using social-security records, while gaps in credit and deposit series are filled using monthly data or neighboring observations.

For presentation, the authors rescale the factor to match the mean and standard deviation of annual real GDP growth. That makes its scale easier to interpret, but does not turn it into directly measured weekly GDP.

What the historical index shows

The preferred specification produces movements the authors associate with four downturns: 2008–09, 2015–16, the pandemic contraction and the weakness of 2024. It places the pandemic trough at the end of May 2020 and shows the 2015 slowdown beginning before the April 2016 earthquake.

The exercise also highlights why combining indicators matters. Credit and deposits did not decline alongside real activity during the pandemic. Deposits also grew strongly during the weak economy of 2024. A financial series alone could therefore give an incomplete picture.

A second-stage regime-switching model converts the estimated factor into recession probabilities. It allows recession and expansion episodes to have different average activity levels. These probabilities are model-based classifications, not official recession declarations.

Better fit is not necessarily a better index

The authors compare eight specifications. An in-sample statistical criterion favors models allowing volatility to change in individual series. However, those models produce a pandemic contraction that the authors judge implausibly mild: unusual movements can be absorbed into individual-series volatility rather than the shared factor.

They instead prefer a specification with changing volatility only in the common factor, plus delayed responses across indicators. This choice emphasizes economic interpretability over the best statistical score.

That trade-off limits how strongly the results can be read. Ecuador has no official recession chronology against which to validate the classifications. The assessment relies mainly on historical plausibility and in-sample comparisons, rather than a reported out-of-sample forecasting contest or systematic replay of information available at each past release date.

The contribution is a framework for organizing fragmented evidence into a frequently updated estimate. Whether it would reliably recognize future downturns earlier remains a separate question.