Macroéconomie Modèles
Choisissez un modèle selon la question économique qui vous intéresse. Découvrez ses hypothèses, ce qu’il peut éclairer et ses limites.
Une partie de cette page est actuellement disponible uniquement en anglais.
Empirical Forecasting Models
Use forecasting models when the observed path, data vintage, error behavior, and model comparison matter more than a single theoretical channel.
- Frame
- Define the target series, frequency, release lag, and revision exposure.
- Assumption
- Does the model assume stable relationships across regimes?
- Evidence
- Does the model beat a simple benchmark out of sample?
Univariate time-series
One macro series, its own history, and the forecast errors that remain.
- ARIMABaseline forecast from one series: lags, differencing, and residual checks.
- SARIMAARIMA with seasonal structure for recurring monthly or quarterly patterns.
- ETS (exponential smoothing)Forecasts built from error, trend, seasonality, and smoothing states.
- HP filter and Beveridge-Nelson decompositionQuestions where the object is not the next forecast but the split between trend, cycle, and transitory movement in one series.
- Unobserved components modelEstimating latent trend and cycle paths when a direct observable is noisy or mixes several frequency components.
Multivariate time-series
Small systems where output, prices, rates, and labor move together.
- VARA small macro system where variables forecast each other from their lags.
- Bayesian VAR (BVAR)VAR dynamics with priors that shrink noisy coefficients toward stable behavior.
- Structural VAR (SVAR)VAR forecasts plus restrictions that identify structural shocks.
- VECMCointegrated variables with short-run adjustment back to long-run equilibrium.
- ARDL / bounds testingTesting whether a long-run equilibrium relation is plausible without requiring every variable to have the same integration order.
- Threshold VARShock-transmission questions where recession, inflation, leverage, or slack regimes plausibly change the coefficients.
- TVP-VAREvidence questions where the same policy shock may not propagate the same way in the 1980s, 2000s, and post-pandemic data.
- Markov-switching VARBusiness-cycle questions where observed series appear to alternate between persistent regimes rather than move around one fixed law.
- Smooth transition regressionNonlinear macro relationships where the move between regimes is gradual rather than an abrupt cutoff.
Nowcasting and indicator models
Current-quarter reads from mixed-frequency releases and incoming news.
- Bridge equationsMonthly indicators mapped into a quarterly target before the official release.
- Dynamic factor modelMany indicators compressed into latent factors for nowcasting.
- Nowcasting with newsForecast changes attributed to the releases that moved the estimate.
- MIDASMixed-frequency regressions without first aggregating every predictor.
- Kalman filterReal-time tracking problems where the state is unobserved and the information set updates observation by observation.
Semi-structural / macroeconometric forecasting
Estimated macro blocks tied together by identities and policy rules.
Panel / cross-country empirical macro
Country, regional, or sector panels where cross-section variation matters.
Machine learning / AI forecasting
High-dimensional predictors, shrinkage, selection, and forecast combination.
- Ridge regressionMany predictors with shrinkage that keeps coefficients stable.
- LASSOVariable selection and shrinkage when the predictor set is wide.
- Forecast ensembleModel forecasts combined and weighted by observed performance.
- Elastic NetWide macro predictor sets where correlated groups matter and pure LASSO would choose one variable too arbitrarily.
- Random forestForecasting with many predictors when interactions and nonlinear thresholds matter more than a single coefficient table.
- Gradient boostingHigh-dimensional macro forecasting when the gain comes from nonlinear interactions but stopping discipline matters.
- LSTMSequence-forecasting problems where lag structure is long, nonlinear, and hard to summarize with fixed AR lags.
Network and input-output empirical models
Input-output links, sector exposure, and spillovers across networks.