Aller au contenu principal

Cookies sur Macro by Mark.Nous utilisons des cookies essentiels pour la connexion, la langue et la sécurité. Avec votre accord, nous ajoutons des cookies optionnels de préférences, d’analyse et de mesure publicitaire avec le pixel Meta (Facebook). Lisez notre politique relative aux cookies.

Macro by MarkMacro by Mark
  • Accueil
Chargement…
Chargement…
← Modèles
MacroConceptsPolitiqueModèles

Bridge equations Model

Simple regressions that bridge from monthly indicators to a quarterly target like GDP, refreshed as each indicator lands.

Vue d’ensembleGraphiqueMathématiquesRéférences

How do you exploit high-frequency monthly indicators to produce a GDP forecast before the quarterly figure is published?

Background

GDP is released quarterly with a delay of 4-8 weeks. Monthly indicators - industrial production, retail sales, employment, PMI surveys, consumer confidence - arrive weeks or months earlier. Bridge equations connect these two frequencies: they regress quarterly GDP growth on monthly indicators aggregated to the quarterly level, then plug in available monthly data (partially observed for the current quarter) to generate a 'nowcast.' The term 'bridge' refers to the frequency gap the model spans: monthly information is bridged to a quarterly target.

The technique dates to the late 1990s when central banks began formalizing their short-term forecasting processes. The Banque de France (Bec and Mogliani 2015), the Bundesbank (Drechsel and Scheufele 2012), and the ECB (Runstler and Sedillot 2003) all documented their bridge equation frameworks. The New York Fed's Staff Nowcast (Bok et al. 2018), by contrast, is a pure state-space dynamic factor model with a news decomposition - it does not use bridge equations. The appeal of bridges is mechanical simplicity: each bridge equation is a standard OLS regression. The forecast updates automatically as new monthly data arrives, making it ideal for real-time tracking of the economy within the quarter.

A typical bridge equation setup for GDP has 3-5 indicator-specific regressions plus an aggregation rule. One equation bridges industrial production to GDP manufacturing. Another bridges retail sales to GDP consumption. A third bridges employment to GDP services. Each equation is estimated separately, and the GDP nowcast is either a simple weighted average or an aggregate built from the component forecasts. Weights are fixed (based on national accounts shares) or estimated (using historical forecast accuracy as inverse-variance weights).

The key operational challenge is the 'ragged edge': at any point within a quarter, different monthly indicators are available for different months. January industrial production may be published, but January retail sales may not yet be available. The bridge equation must handle this partial information set by forecasting the missing monthly values (typically with univariate AR models) before aggregating to the quarterly frequency. The forecast updates as each new monthly release arrives, producing a sequence of progressively more accurate nowcasts.

Connections to explore

Math ↗The structure and its conditions, step by step.References ↗Academic sources and guidance for using the model.

How the Parts Fit Together

yQy_QyQ​

Quarterly target variable

The variable to be nowcast, typically quarter-on-quarter GDP growth or a GDP component growth rate. Published with a 4-8 week lag after the quarter ends.

Inputs are two-layered. The quarterly layer is the target variable - typically GDP growth or a major GDP component (consumption, investment, government spending, net exports). The monthly layer consists of k indicator variables, each chosen for its leading or coincident relationship with the quarterly target. Selection criteria: availability (published before GDP), predictive power (significant in-sample), and stability (coefficient estimates do not shift dramatically across subsamples).

The model has two stages. Stage 1 forecasts missing monthly values for each indicator using a univariate model (AR, exponential smoothing, or a simple random walk). Stage 2 aggregates each monthly indicator to the quarterly frequency (usually by averaging the three monthly values) and plugs the result into the bridge regression: yQ=c+β1x1,Q+…+βkxk,Q+eQy_Q = c + \beta_1 x_{1,Q} + \ldots + \beta_k x_{k,Q} + e_QyQ​=c+β1​x1,Q​+…+βk​xk,Q​+eQ​, where yQy_QyQ​ is quarterly GDP growth and xj,Qx_{j,Q}xj,Q​ is the quarterly average of monthly indicator jjj. The quarterly forecast is the fitted value from this regression using the most recent available information.

Aggregation from monthly to quarterly follows one of two conventions. Flow variables (industrial production, retail sales) are averaged over the three months of the quarter. Stock variables (employment, inventories) use the end-of-quarter (third month) value. Survey data (PMI, confidence indices) are averaged. The aggregation choice matters: using the wrong convention introduces a temporal misalignment that biases the bridge coefficient. National statistical offices publish both monthly and quarterly series; the bridge equation must use the aggregation that matches the quarterly national accounts definition.

Assumptions

Testable

Stable quarterly relationship

The regression of yQy_QyQ​ on xj,Qx_{j,Q}xj,Q​ has coefficients that are stable over time. No structural breaks in the bridge between indicators and GDP.

If violated:If the GDP composition shifts (e.g., services overtake manufacturing), bridge equations calibrated to the old structure produce biased nowcasts. Rolling-window estimation partially addresses this.

Testable

Temporal aggregation is correct

The monthly-to-quarterly aggregation matches the national accounts convention. Flow variables are averaged; stock variables use end-of-quarter values.

If violated:Wrong aggregation introduces a timing misalignment that attenuates the bridge coefficient and reduces forecast accuracy.

Testable

Missing monthly values are predictable

The AR or naive model used to fill missing months produces unbiased forecasts of the monthly indicator.

If violated:Biased monthly fill-in propagates directly into the quarterly nowcast. Systematic over- or under-prediction of missing months creates a persistent nowcast bias.

Testable

Indicators are leading or coincident

Each monthly indicator is published before the quarterly target and has contemporaneous or leading predictive power for GDP growth.

If violated:Lagging indicators do not help with nowcasting - they arrive too late. Including them adds noise without improving timeliness.

Testable

Linear relationship between indicator and target

The quarterly regression is linear: yQ=c+βxQ+ey_Q = c + \beta x_Q + eyQ​=c+βxQ​+e. No threshold effects, no asymmetric response in expansions versus recessions.

If violated:Nonlinear bridges (e.g., GDP responds asymmetrically to positive vs. negative PMI surprises) underperform linear bridges in normal times but can improve recession-era nowcasts.

Testable

No multicollinearity among indicators

When multiple indicators enter the same bridge equation, they are not highly correlated. Otherwise, individual coefficients are unstable and forecast variance increases.

If violated:High correlation among indicators inflates standard errors and makes the bridge equation sensitive to small data revisions. Remedy: use principal components of the indicators, or estimate separate single-indicator bridges and combine forecasts.

Applications

The Banque de France operates one of the best-documented bridge equation systems in the world. Their ISMA (Indicateur Synthétique Mensuel d'Activité) model uses about 10 monthly indicators - industrial production, business surveys, consumption of manufactured goods, construction starts - to nowcast French GDP. Each indicator feeds a separate bridge equation. The GDP nowcast is a weighted average of the individual bridge forecasts, with weights based on past relative forecast accuracy. The system produces a new nowcast each time a monthly indicator is released, typically 3-4 updates per month during the quarter.

The Federal Reserve Bank of Atlanta's GDPNow model is the most prominent real-time bridge equation tracker in the United States. It bridges 13 subcomponents of GDP using monthly and weekly source data. Each subcomponent has its own bridge equation or accounting identity. The model updates after every significant data release and publishes the running estimate publicly. Its accuracy at 1-2 months into the quarter rivals professional forecaster consensus.

Bridge equations serve as building blocks in larger forecasting frameworks. The ECB's suite combines bridge equations for GDP components with a dynamic factor model that extracts common factors from 100+ monthly indicators. The bridge equations handle the aggregation from monthly to quarterly; the factor model handles the dimension reduction and missing-data problem. In this role, bridge equations are not standalone models but infrastructure components. The New York Fed's Staff Nowcast, by contrast, is a pure state-space dynamic factor model that handles the monthly-to-quarterly mapping inside the state space, without bridge equations.

Bridge equations break down during structural breaks and economic crises. The COVID-19 recession produced indicator readings far outside historical ranges: PMI values dropped to levels never seen before, unemployment claims spiked by an order of magnitude. Bridge equations estimated on pre-COVID data produced wildly inaccurate nowcasts in March-April 2020 because the historical relationship between indicators and GDP was nonlinear in the extreme tail. Intercept corrections, regime-switching bridges, and manual judgment overrides were needed to restore reasonable nowcasts.

Extensions and open questions

Variants

  • Single-indicator bridge: one monthly indicator per bridge equation. Simple and transparent. The GDP nowcast combines multiple single-indicator bridges.
  • Multi-indicator bridge: multiple monthly indicators in a single regression. More efficient if indicators are not collinear, but less robust to data revisions.
  • Factor-augmented bridge: replace individual indicators with principal components (factors) extracted from a large panel. Handles the curse of dimensionality when many indicators are available.
  • Nonlinear bridge: allow the bridge coefficient to differ in recessions versus expansions, or above/below a threshold PMI level. Improves crisis-era nowcasting at the cost of more parameters.
  • Mixed-frequency bridge (MIDAS-style): instead of aggregating monthly to quarterly before regression, use the monthly observations directly with MIDAS lag polynomials. Avoids the information loss from temporal aggregation.

Open questions

  • Optimal indicator selection: with hundreds of potential monthly indicators, which subset produces the best nowcast? Automated variable selection (LASSO, elastic net) competes with judgment-based selection. No consensus on which approach dominates in real-time forecasting.
  • Combining bridge nowcasts: equal weights, inverse-MSE weights, Bayesian model averaging, or machine learning combination? The forecast combination literature suggests equal weights are hard to beat, but this finding may not extend to the ragged-edge nowcasting setting.
  • Handling extreme events: should bridge equations include nonlinear terms, dummy variables for crises, or regime-switching coefficients? Adding flexibility improves crisis-era fit but risks overfitting in normal times.

Keep going

Read next

Continue with a related concept, model, or example.

Theory model

IS-LM Model

a well-connected theoretical model two steps from here

New angle: theoretical model →
Macro by Mark
Aide
Politique de confidentialitéConditions d’utilisationÉthique et conformité

© 2026 Mark Jayson Nation LLC