Model Averaging and its Use in Economics

Scope and contribution

• Model averaging primarily addresses covariate uncertainty in normal linear regression and its extensions, while also treating dependence over time or space and endogeneity.The paper focuses on situations where observations exceed potential covariates and excludes most classification, clustering, density-estimation, nonparametric machine-learning, and data-mining approaches.Source: §1 · pp. 2, 4
• The paper surveys major methodological developments in Bayesian and frequentist model averaging, connecting econometric, statistical, forecasting, shrinkage, and variable-selection literatures.It aims to provide an accessible but formal, up-to-date review for economists, while acknowledging that the survey is not exhaustive.Source: §1 · pp. 2–3

Implementation and recommendations

• BMA weights based on posterior model probabilities are highly sensitive to prior assumptions, and this sensitivity does not disappear as sample size grows.The paper therefore stresses understanding prior effects and recommends prior structures that are easy to elicit and naturally robust.Source: §1 · pp. 3–4
• The paper discusses numerical implementation for Bayesian and frequentist methods, provides freely available computational resources, and offers recommendations for empirical researchers.These resources are intended to let applied researchers try the methods on their own data without a prohibitively large implementation investment.Source: §1 · pp. 3–5

Economic applications

• Model averaging has advanced economic understanding across growth, impulse-response, crisis-warning, production-efficiency, finance, and forecasting applications.It provides a common framework for interpreting different empirical analyses and resolving differences among them.Source: §1 · p. 3

Bayesian and frequentist approaches

• Bayesian model averaging receives primary emphasis because it offers appealing statistical properties, interpretable posterior and inclusion probabilities, and easier computation in large model spaces.Efficient MCMC algorithms make BMA more practical than FMA in large model spaces, while its finite-sample probability framework links naturally to prediction and decision theory.Source: §1 · p. 4

2 Model uncertainty

• Model averaging incorporates evidence across models and adjusts inference for model uncertainty, whereas model selection conditions inference on one chosen model and can understate uncertainty.Averaging is generally appropriate for prediction, covariate effects, and policy questions that are not model-specific; selection is more suited to identifying a true model when the model space contains it.Source: §2 · pp. 6–8

2.1 Why averaging?

• Bayesian model averaging averages first over parameters within each model and then over models using posterior model probabilities, with marginal likelihoods determining model comparisons.This produces mixture posterior or predictive distributions for quantities having a common interpretation across models.Source: §2.1 · pp. 8–10

2.2 Construction of the model space

• The model space must reflect the uncertainty being studied, including relevant theories, specifications, and heterogeneity, while distinguishing M-closed settings from the more realistic M-open economic case.Spaces may be finite or continuous and can range from 20 models for capital-punishment analysis to 32 or 256 models in other applications; omitted models cannot inform averaged inference.Source: §2.2 · pp. 11–13

2.3 Covariate uncertainty in the normal linear regression model

• The canonical framework treats normal linear regression with uncertainty over covariate inclusion, forming a space of K = 2^k subsets that includes null and full models.All models share an intercept and scale interpretation, while standardized regressors make the intercept common across models; nonlinear transformations and interactions can still be included as regressors.Source: §2.3 · pp. 13–14
• Covariate-uncertainty methods include shrinkage, information criteria, cross-validation, Extreme Bounds Analysis, s-values, general-to-specific modelling, Model Confidence Sets, best subset regression, and decision-theoretic Bayesian selection.These methods respectively shrink coefficients, penalize complexity, assess prediction, test robustness, summarize estimate fragility, reduce general models, retain statistically competitive models, search optimal subsets, or optimize utility and loss.Source: §2.3 · pp. 15–16
• Best subset regression found 22 variables using Mallows’ Cp or 23 using BIC in the FLS data, exceeding typical BMA model sizes of 5–10 or average sizes of 10–20.The complete-enumeration runtime simulation found log runtime roughly linear in k, while leaps-and-bounds avoided enumerating every model.Source: §2.3 · p. 17
• Model selection typically conditions inference on one chosen model and can underestimate uncertainty, whereas model averaging retains evidence from alternative models and can also support model selection when posterior model probabilities are available.Most listed approaches target a single best model or a model set; BMA can select the highest-posterior model, but most selection methods do not specify model-space priors.Source: §2.3 · pp. 17–18
• In the UK inflation comparison, Hendry’s selected model contained all variables with very high BACE posterior inclusion probabilities, but its posterior probability was below 0.1 versus 20% for the top model.The Hendry model was the second most probable model in the BACE analysis.Source: §2.3 · p. 18
• Selection procedures with oracle properties remain asymptotic and can underestimate uncertainty in finite samples when the true model is outside the considered model space.This limitation applies to procedures such as SCAD and adaptive LASSO despite their simultaneous variable selection and parameter estimation.Source: §2.3 · p. 18
• Machine-learning ensembles are mainly predictive, combine many often weak base models, generally lack probability-based uncertainty intervals, and trade structural interpretability for nonlinear flexibility.They are therefore not considered in this article, whose model-averaging methods focus on parameter estimation and structural identification.Source: §2.3 · p. 19

3 Bayesian model averaging

• BMA uses exact or approximate posterior model probabilities as averaging weights, integrates out model-specific parameters for prediction, and requires priors for both models and parameters.Source: §3 · p. 19

3.1 Prior Structures

• Priors should reflect interpretable beliefs and be checked for robustness, predictive ability, objective-prior properties, and frequentist performance.Model averaging is often surprisingly sensitive to priors, motivating structures that enhance robustness; predictive checks assess both likelihood and prior components.Source: §3.1 · p. 20, §3.1.1 · p. 21

3.1.1 Priors on model parameters

• For model-specific coefficients, the Fernández et al. (2001a) specification combines Jeffreys-style priors for common parameters with a g-prior, yielding closed-form marginal likelihoods and posterior distributions.Improper priors are permissible for common parameters but not model-specific ones, because arbitrary constants otherwise prevent posterior model probabilities from being determined.Source: §3.1.1 · pp. 21–22

3.1.2 Priors over models

• Hierarchical model-size priors are less sensitive to the chosen mean size than fixed inclusion probabilities, while overly tight small-model priors can reduce predictive performance and model fit.Random-w priors down-weight model sizes near k/2 and account for the number of models at each size, whereas fixed-w priors can strongly favor larger or smaller models.Source: §3.1.2 · pp. 25–26, Figure 2 · p. 26

3.1.3 Empirical Bayes versus Hierarchical Priors

• Because g and w can strongly influence posterior model probabilities, hyperpriors provide a Bayesian way to reduce dependence on a single default choice.The section discusses hierarchical priors on w and g, including hyper-g families and beta priors on the shrinkage factor δ = g/(1 + g).Source: §3.1.3 · pp. 27–28
• Hierarchical priors make marginal likelihoods generally analytically unavailable, motivating Laplace approximations, Gibbs sampling, or Beta shrinkage priors with analytical marginal likelihoods.The robust prior may be problematic when sample size greatly exceeds typical model size; Ley and Steel (2012) recommend a benchmark beta prior with b = c max{n, k2} and c = 0.01.Source: §3.1.3 · pp. 28–29
• Empirical-Bayes procedures use data to select g and w, providing automatic calibration and avoiding the computational complications of hyperpriors, but departing from formal Bayesian prior specification.Local EB estimates g separately for each model, whereas global EB estimates common g and w by maximizing marginal likelihood averaged over models.Source: §3.1.3 · p. 29
• Fully Bayesian and EB procedures both adjust for multiplicity, but EB can assign probability one to the full or null model and may differ substantially across applications; Scott and Berger (2010) recommend fully Bayesian procedures.Cui and George (2008) favor global EB in one known-σ setting, while Liang et al. (2008) find little difference between EB and fully Bayesian procedures when σ is unknown and k < n.Source: §3.1.3 · p. 29
• In a more general GLM setting, local EB performs poorly in null-model simulations, almost always selecting the full model.Li and Clyde (2018) compare EB and fully Bayesian procedures in this broader context.Source: §3.1.3 · p. 30

3.2.1 Consistency and paradoxes

• Model-selection consistency depends on parameter and model-space priors, with several priors consistent under fixed or growing model dimensions, while others fail in specific regimes.For k = O(n^b), consistency ranges from b < 1/2 under uniform model-space priors to b ≤ 1 with a Beta(1,1) hyperprior on w; nonnested-model consistency can depend on pseudo-distance.Source: §3.2.1 · p. 30
• In M-open settings, g-prior BMA selects a model intuitively closest to the true model and attains comparable regression-function estimation loss, while sparse-regression BMA contracts optimally and shrinks spurious coefficients rapidly.These results hold under conditions on g’s growth, including fixed and growing predictor dimensions; nonlocal priors can provide even faster shrinkage.Source: §3.2.1 · pp. 31–32
• Hyperpriors such as Zellner–Siow, hyper-g, and benchmark beta priors resolve the information paradox, whereas fixed g can induce Bartlett and information paradoxes; consistency may still fail near the null model in ANOVA when k is proportional to n.A modified hyper-g/n prior is proposed to address the ANOVA inconsistency region.Source: §3.2.1 · pp. 31–32

3.2.2 Predictive performance

• BMA generally improves probabilistic forecasting over single-model approaches, with results supported by LPS, CRPS, predictive coverage, squared-error loss, and volatility-forecasting applications.CRPS is less sensitive to outliers than LPS, while theoretical results establish at-least-as-good LPS performance and minimized expected squared-error loss under specified predictive assumptions.Source: §3.2.2 · pp. 32–34

3.2.3 Frequentist performance of point estimates

• Under the predictive sampling model, BMA point estimates minimize MSE and its posterior intervals have average coverage at least equal to their nominal probability, but these guarantees need not hold under other data-generating distributions.The paper emphasizes that economic data are unlikely to come from a single model within the candidate space or a shrinking neighborhood around one.Source: §3.2.3 · pp. 34–35

• MCMC methods address large model spaces by constructing chains with the posterior as invariant distribution, using Metropolis-Hastings, Gibbs sampling, or combinations of both.MC3 is a widely used random-walk Metropolis sampler, while SSVS jointly samples parameters and models when marginal likelihoods are unavailable analytically.Source: §3.3 · pp. 35–36
• k = 100 potential covariates create K = 2^100 = 1.26 × 10^30 models, making exhaustive enumeration impractical even with fast processors.Analytical integration of parameters makes pairwise Bayes factors easy to compute, but the model-space size remains the central computational challenge.Source: §3.3 · p. 35
• Correlated covariates can cause MC3 and SSVS to mix poorly, motivating adaptive MCMC, evolutionary Monte Carlo, and Bayesian adaptive sampling without replacement.BAS represents models with a binary tree and updates sampling probabilities using marginal likelihoods from previously sampled models.Source: §3.3 · pp. 36–37
• Strategy i, sampling posterior models and estimating by visit frequencies, can be more precise than strategy ii, which searches good models and renormalizes their probabilities but may be search-biased.Some methods permit both strategies, whereas SSVS with the George–McCulloch (1993) prior requires strategy i and BAS or leaps-and-bounds implementations require strategy ii.Source: §3.3 · p. 37
• Adaptive add-delete-swap proposals improve exploration when standard MC3 overproposes additions; Lamnisos et al. (2013) achieved almost 30 times MC3’s efficiency in a probit problem with n = 62 and k = 1224.More advanced adaptive algorithms can yield orders-of-magnitude efficiency gains for problems with k = 22,576 covariates and n = 60 observations.Source: §3.3 · pp. 37–38
• When model-space MCMC is impractical, analysts can approximate marginal likelihoods or compute them separately using Chib’s method, bridge sampling, or Savage-Dickey ratios.The latter is especially applicable to nested models and is typically easy to calculate and accurate.Source: §3.3 · pp. 38–39

Role of the prior

• Prior distributions can affect posterior model probabilities much more than posterior inference conditional on a model, motivating systematic analysis of g and w.The interplay between g and w was investigated in Ley and Steel (2009b) and Eicher et al. (2011).Source: §3.4 · p. 39
• The posterior-odds decomposition contains prior-odds complexity, Bayes-factor size, and lack-of-fit penalties; the first rises with g and falls with w.The prior-induced complexity penalty may favor either smaller or larger models, whereas Bayes-factor penalties favor smaller and better-fitting models.Source: §3.4 · p. 39
• Hyperpriors on g and w strongly affect induced complexity penalties but do not materially change the impact of relative model fit.Ley and Steel (2012) analyze the three posterior-odds factors separately and relate their log penalties to classical information criteria.Source: §3.4 · p. 39
• Random w and g induce a multiplicity penalty that penalizes models near half the model space, whereas fixed w cannot provide this correction.With equal fit, random hyperparameters yield less extreme posterior odds than fixed values and qualitatively resemble the prior-odds behavior.Source: §3.4 · p. 40, Figure 3 · p. 40
• Marginalising over hyperpriors on w and g produces a much flatter model-size penalty, reducing prior sensitivity and increasing the relative importance of model fit.The choice of prior mean model size matters little with random w and g but substantially with fixed values.Source: §3.4 · p. 40, Figure 3 · p. 40

3.5 Data Robustness

• Revisions across Penn World Table versions 6.0–6.2 produced substantial variation in posterior inclusion probabilities for some growth determinants under both examined methods.Ciccone and Jarociński (2010) applied the methodologies of Fernández et al. (2001b) and Sala-i-Martin et al. (2004).Source: §3.5 · pp. 40–41
• Feldkircher and Zeugner (2012) found that common-country samples roughly halved changes, while a hyper-g prior dramatically reduced instability by adapting g to the data.Fixing g at an unwarranted value amplifies small fit differences; stochastic-g analyses yielded smaller g values.Source: §3.5 · p. 41
• Across three datasets, the data favored g values of 15–50, versus benchmark-prior values of 1681, 4489, and 1190, respectively, making model differences less pronounced.The datasets included two cross-country growth samples and Tobias and Li’s returns-to-schooling data.Source: §3.5 · p. 41
• Rockey and Temple (2016) found that requiring initial GDP per capita and regional dummies enhanced robustness across newer Penn World Table vintages, including 6.3–8.0.Restricting the model space effectively imposed a more informative prior.Source: §3.5 · p. 41

3.6 Collinearity and Jointness

• Jointness measures complement marginal summaries by describing the posterior tendency of covariates to appear together in Bayesian model averaging.Ley and Steel (2007) and Doppelhofer and Weeks (2009) propose measures intended to capture the multivariate posterior distribution more adequately.Source: §3.6 · p. 42
• A satisfactory jointness measure should be interpretable, calibrated, attain maximum jointness when variables always co-occur, and remain defined whenever at least one variable has positive inclusion probability.These are the four criteria formulated by Ley and Steel (2007).Source: §3.6 · p. 42
• Ley and Steel’s measure is a posterior odds ratio comparing joint inclusion with inclusion of either regressor alone, whereas Doppelhofer and Weeks’ measure uses conditional posterior odds.Values above one indicate evidence for jointness, while values below one indicate disjointness; the Ley–Steel measure can be undefined when a variable is always included or excluded.Source: §3.6 · pp. 42–43
• In growth-data applications, important complementary variables showed jointness, while more numerous disjoint relationships indicated substitute regressors that were generally unimportant growth drivers.Disjointness can arise when highly collinear variables act as proxies or substitutes for one another.Source: §3.6 · p. 43
• Different jointness measures can produce substantially different results, whereas results appear robust across prior choices; users should therefore select measures whose properties support the intended interpretation.Composite indicators may combine information across measures but can make interpretation harder.Source: §3.6 · p. 43
• Strong collinearity can create multimodal model posteriors, motivating joint summaries and routine calculation of joint inclusion probabilities alongside marginal inclusion probabilities.Alternative approaches cluster posterior model profiles using latent-class and Dirichlet-process methods, while later work proposes a regularized Yule’s Q jointness measure satisfying extended criteria.Source: §3.6 · pp. 43–44

BIC approximations

• BIC provides an asymptotic approximation to log Bayes factors in regular models and is used when exact marginal likelihoods are unavailable or costly.For the normal linear model with prior (6), Fernández et al. (2001a) link this approximation to g = n, the unit information prior.Source: §3.7 · p. 44

AIC approximations

• Smoothed AIC offers an alternative approximation to posterior model probabilities, justified by Burnham and Anderson (2002) under a different model prior.Pelster and Vilsmeier (2018) apply smoothed AIC weights when assessing credit-default-swap pricing determinants.Source: §3.7 · p. 44

Frequentist–Bayesian hybrids

• Hybrid procedures use BIC-based posterior model weights to average frequentist estimates, including BACE, BAMLE, and classical 2SLS estimates for endogenous-regressor models.Sala-i-Martin et al. (2004) introduce BACE for linear models with a g-prior, Moral-Benito (2012) generalizes it to panel data as BAMLE, and Durlauf et al. (2008, 2012) apply the hybrid approach to endogenous regressors.Source: §3.7 · p. 45

3.8 Prior robustness: illusion or not?

• Different g and w settings can yield similar growth-data results, but this apparent robustness reflects a tradeoff between their opposing model-complexity effects.Fernández et al. (2001b) use w = 0.5 and g = k2, whereas Sala-i-Martin et al. (2004) use w = 7/k and g = n; their combinations lie on nearby equal-probability contours.Source: §3.8 · pp. 45–46
• Figure 4 shows that the Fernández et al. (2001b) and Sala-i-Martin et al. (2004) choices require nearly identical fit increases to offset an additional regressor.For the Sala-i-Martin et al. (2004) setting, the exact value is R2_i = 0.7751, with n = 88, k = 67, k_i = 8, k_j = 7, and R2_j = 0.75.Source: Figure 4 · p. 46, §3.8 · p. 46
• Frequently used hyperparameter combinations may look similar, but results are not generally robust because g and w choices matter crucially.Making prior assumptions implicit, as in BACE, does not eliminate their influence; transparency about prior choices is therefore important.Source: §3.8 · p. 46
• A Beta(1, (k −m)/m) hyperprior on w makes the prior less informative, greatly weakens m-sensitivity, and nearly removes the g–w tradeoff; a hyperprior on g can further robustify analysis.These robustness improvements are illustrated in Figure 4’s right panel and extend the approach discussed for g hyperpriors.Source: §3.8 · pp. 46–47

3.9.1 Generalized linear models; 3.9.2 Probit models; 3.9.3 Generalized additive models

• BMA accommodates generalized linear, probit, and additive models through specialized priors and approximations, with model uncertainty often explored by MCMC.GLMs cover non-normal responses; probit models commonly use data augmentation and reversible-jump methods, while additive models expand the model space over covariate powers or spline specifications.Source: §3.9.1 · p. 47, §3.9.2 · p. 48, §3.9.3 · p. 49

3.9.4 Outliers

• 18 variables were relatively robustly associated with GDP growth from 1960 to 1996 when outliers, heterogeneous measurement error, and data-vintage effects were modeled.Doppelhofer et al. (2016) used eight PWT vintages and JAGS, with a single MCMC run taking about a week even with multiple computers and parallel chains.Source: §3.9.4 · p. 49

3.9.5 Non-normal errors

• Student-t and finite normal-mixture sampling models make inference more robust to outliers and unmodelled heterogeneity, while transformations jointly address non-normality and covariate uncertainty.Doppelhofer and Weeks (2011) used BIC-based posterior model probabilities, so the Student-t specification affected parameter estimates but not model probabilities.Source: §3.9.5 · p. 50

3.9.6 Dynamic models; 3.9.9 Spatial data; 3.9.10 Duration models

• BMA also extends to dynamic, spatially structured, and survival settings, including Cox, hazard, and competing-risks duration models using BIC or Laplace marginal-likelihood approximations.Duration applications study growth spells, firm bankruptcies and defaults, and dropout or delayed graduation while allowing time-dependent covariates or multinomial-logistic outcome structures.Source: §3.9.6 · p. 50, §3.9.9 · p. 54, §3.9.10 · pp. 54–55

3.9.6 Dynamic models

• Dynamic model averaging and selection use time-varying model weights to combine or select forecasts; Koop and Korobilis (2012) found improved inflation forecasts versus usual autoregressive and time-varying-parameter models.Drachal (2016) found no significant DMA superiority over a simple ARIMA model for point-forecast accuracy, while Hoogerheide et al. (2010) reported predictive and economic gains for other combination forecasting schemes.Source: §3.9.6 · p. 51

3.9.7 Endogeneity

• Endogeneity can make BMA coefficient and model-probability inference misleading even as sample size grows, motivating 2SLS, two-stage BMA, IVBMA, and specialized MCMC approaches.These methods average over endogenous and exogenous variables, instruments, or first-stage and second-stage models, but some approaches are computationally costly or rely on restrictive assumptions.Source: §3.9.7 · p. 52

3.9.8 Panel data and individual effects

• Panel-data BMA incorporates individual effects, dynamic dependence, endogenous covariates, instruments, and exogeneity restrictions using likelihood, GMM, BIC, and reversible-jump constructions.The literature includes dynamic panels with fixed effects and weakly exogenous regressors, limited-information BMA, and model spaces that vary in regressors, instruments, and restrictions.Source: §3.9.8 · p. 53

3.9.9 Spatial data

• Spatial BMA has been applied to spatial autoregressive models, but simulation results found SVSS methods better in both in-sample predictive performance and computational efficiency.Applications include income growth, human-capital accumulation, and urban income inequality; spatial eigenvector filtering can reduce computational requirements.Source: §3.9.9 · p. 54

4 Frequentist model averaging

• FMA combines model-specific estimators using simplex weights, and the central methodological problem is choosing those weights without specifying parameter priors.Frequentist methods emphasize repeated-sampling estimator properties, while parameters remain fixed but unknown.Source: §4 · p. 55
• Frequentist model averaging includes information-criterion, likelihood, Mallows, cross-validation, and plug-in weighting procedures with asymptotic optimality established across linear, heteroscedastic, GLM, and semiparametric settings.MMA minimizes large-sample squared predictive error, while JMA remains optimal under heteroscedastic errors and non-nested candidate models.Source: §4 · pp. 56–57

4.1 Estimating the sampling variance of FMA estimators

• For interval estimation, full-model confidence intervals can match FMA intervals asymptotically, whereas simulations show FMA improves finite-sample MSE for point estimation.Model uncertainty must be incorporated because conditional classical intervals are too narrow; Wang and Zhou’s equivalence applies also to varying-coefficient partially linear models.Source: §4.1 · pp. 59–60

4.2 Computational aspects of FMA

• FMA’s computational burden in large model spaces reflects the absence of estimated model probabilities and the costly weight calculations required by methods such as JMA.Researchers reduce the model space through screening, semiorthogonal transformations, or always including selected covariates.Source: §4.2 · pp. 60–61

4.3.1 The forecast combination puzzle

• Equal-weight forecast combinations can outperform data-adaptively estimated weights, creating the “forecast combination puzzle” because weight-estimation uncertainty can reduce mean squared forecast accuracy.When optimal weights are near equal, simple averaging may dominate estimated combinations.Source: §4.3.1 · p. 61

4.3.2 Density forecast combinations

• Density forecast pools choose weights by minimizing Kullback-Leibler distance, and including inferior models can substantially improve prediction.Variable-dependent piecewise-linear weights estimated by optimizing LPS have consistency and asymptotic normality, with advantages over constant-weight combinations in simulations and real data.Source: §4.3.2 · pp. 61–62

5.1 Growth regressions

• Growth regressions face extensive uncertainty because theories offer many potential determinants, measurements, instruments, and possible endogenous variables.Earlier approaches such as EBA and sequential variable addition considered limited model sets, motivating formal BMA and related averaging methods.Source: §5.1 · pp. 62–63
• 2SBMA finds that geography, integration, and institutions all matter for development, revising conclusions based on limited models and specific instruments.The approach averages across many growth determinants and instruments, producing more complete inference; it also downweights potentially inappropriate instruments such as settler mortality when warranted.Source: §5.1.1 · p. 64
• BMA applications identify heterogeneous growth determinants, including energy use, public investment, savings, initial GDP, capital formation, natural resources, and financial-market competition.Energy consumption is important for aggregate GDP growth; public investment may exceed public consumption effects in high-debt countries; savings and capital formation support an Asian development model; oil rents have robust positive effects; and financial-market competition is growth-enhancing.Source: §5.1.2 · p. 65, §5.1.3 · pp. 65–66, §5.1.4 · p. 66
• BMA conclusions can depend on prior structure and modeling choices, as shown by differing effects for government variables and natural-resource interactions.A strong heredity prior makes the main effect of oil rents dominate, whereas treating interactions as additional covariates attributes the effect largely to institutional quality and oil revenues.Source: §5.1.3 · p. 65, §5.1.4 · p. 66

5.2 Inflation and Output Forecasting

• Model averaging is used for inflation and output forecasting with dynamic weights, predictive weights, regime switching, principal components, bridge models, and factor models.Applications include Bayesian averaging of ARIMA and ARFIMA models for real GNP impulse responses, inflation forecasting, GDP forecasting, and joint inflation-output prediction.Source: §5.2 · pp. 67–68
• PCR performs best for deviations of output and inflation from expected paths, whereas BMA performs best for tail events in a 10-year rolling evaluation.The comparison covered 24 BMA specifications and 2 PCR specifications, implying different forecasting choices for risk-neutral and risk-averse policymakers.Source: §5.2 · p. 68
• For Euro-area GDP, BMA bridge models produce smaller forecast errors than a small-scale dynamic factor model and an indirect bridge model.Bridge models connect monthly information to quarterly national-account data for timely nowcasts.Source: §5.2 · p. 68
• A BMA-S SVS-style recession-forecasting method performs well relative to main competitors on monthly U.S. data with 108 potential covariates.The method combines a Gibbs sampler, generalized double Pareto prior, and correlation-based dilution prior over models.Source: §5.2 · pp. 68–69

5.3 VAR and DSGE modelling

• BMA supports Bayesian inference and forecasting in VARs, including structural, stochastic-search, and panel formulations with cross-unit interdependencies.Applications include UK inflation and output forecasting, VARs without exact coefficient zeros, and Panel VARs allowing heterogeneity and interdependence.Source: §5.3 · p. 69
• SVSS-based Bayesian VAR analysis finds significant international spillovers from US demand, supply, and monetary-policy shocks, mainly through financial channels with regional variation.The evidence concerns expansionary aggregate demand and supply shocks and a contractionary monetary-policy shock.Source: §5.3 · p. 69
• US quarterly evidence indicates a model-structure break around 1984, after which technology shocks explain all stochastic trends and investment-specific shocks dominate business-cycle volatility.The finding comes from Bayesian model averaging in DSGE models using data from 1948–2009.Source: §5.3 · p. 69

5.4 Crises and finance

• BMA crisis-warning studies find that inflation aggravates exchange-market pressures, reserves mitigate them, and signal importance varies across banking, balance-of-payments, exchange-rate, and recession crises.No single early-warning signal explains all dimensions of the 2008 crisis; accounting for theory uncertainty changes earlier consensus conclusions.Source: §5.4 · pp. 69–70
• Pre-crisis credit and real-activity growth predicted weaker resilience, while hierarchical BMA identifies financial policies and trade linkages as especially relevant to crisis performance.Group-level and variable-level uncertainty jointly identify more crisis-related variables than approaches selecting variables individually.Source: §5.4 · p. 70
• Institutional similarity, distance, and political risk emerge as contagion predictors, with sudden surges and stops concentrated in fragile emerging markets.Both WALS and BMA support distance and political risk as robust drivers of contagion.Source: §5.4 · pp. 70–71
• Forecasting and financial-market studies show slightly better exchange-rate forecasts than a random walk, robust CDS pricing links to tail dependence, and difficult out-of-sample prediction of OECD financial stress.Bumping reportedly performs even better than BMA for exchange-rate forecasting, while financial diversity and monetary-policy determinants vary across settings and samples.Source: §5.4 · pp. 71–73
• BMA applications find no clear link between bank competition and financial stability, while rule of law robustly predicts financial development and trust predicts stock-market depth and liquidity.The bank-competition result persists after correcting for publication bias and misspecification; rule-of-law effects hold across financial-development measures.Source: §5.4 · pp. 71–73

5.5 Production modelling

• Stochastic frontier models face substantial uncertainty over inefficiency-error distributions, which can materially affect inference on firm efficiencies when observations per firm are scarce.Exponential, gamma, generalized gamma, and mixture distributions can imply different efficiency patterns; the exponential choice may cluster firms near high efficiency.Source: §5.5 · p. 73
• Bayesian and frequentist model averaging address uncertainty over inefficiency distributions, frontier covariates, and variance specifications while averaging elasticities, returns to scale, and efficiency measures.Parmeter et al. (2016) develop nested-structure averaging and J-fold cross-validation for nonnested inefficiency distributions; Sickles (2005) uses simple averages for US banks.Source: §5.5 · p. 73
• BMA applications also study productivity dynamics and electricity-distribution efficiency, using model spaces that address covariates, firm inefficiencies, or complex marginal-likelihood calculations.McKenzie (2016) examines railroad productivity growth, while Makieła and Osiewalski (2018) analyze Polish electricity cost efficiency without running a chain over model space.Source: §5.5 · pp. 73–74

5.6 Other applications

• BMA and FMA meta-analyses find that consumption estimates vary with demographic context, publication bias, liquidity constraints, and study characteristics, while supporting the permanent income hypothesis for average consumers.Havranek and Sokolova (2016) analyze 2,788 estimates from 133 studies, and Havranek et al. (2017) find broadly similar results with BMA and FMA for habit formation.Source: §5.6 · p. 74
• Other applications identify heterogeneous spatial mobility, spatial spillovers in democracy, institutional predictors of corruption, and limited support for historical, cultural, and geographic corruption determinants.These studies span commuter zones, municipal inequality, endogenous regressors, and spatial models, including Kourtellos et al. (2016), Oberdabernig et al. (2018), and Jetter and Parmeter (2018).Source: §5.6 · pp. 74–75
• Education and export-diversification studies use BMA variants to identify robust relationships, including near-zero tuition-fee effects on enrollment after publication-bias adjustment.Primary-school enrollment raises export diversification, natural-resource shares lower it, and tuition effects are similar under BMA and FMA; male and private-university enrollment declines with tuition fees.Source: §5.6 · pp. 74–75
• Model averaging is applied to globalization, resource efficiency, and social acceptance, linking country-specific effects or resource abundance to informal activity, sustainability, and cross-country heterogeneity.Hartwell (2016) concludes that improved property rights can make resource use more sustainable in resource-abundant countries, while Cohen et al. (2016) explain acceptance differences across 27 countries.Source: §5.6 · pp. 75–76
• Dynamic model averaging significantly improves Chinese house-price forecasts in recursive and rolling modes, while predictor importance changes over time and trade studies show model uncertainty, dynamics, and endogeneity jointly alter conclusions.Wei and Cao (2017) find Google searches recently outperform traditional macroeconomic predictors; Chen et al. (2018) identify trade persistence, exchange-rate effects, and agreement-related creation and diversion.Source: §5.6 · p. 76

Availability and adoption

• Free software is important for methodological adoption, and early BMA implementations include code by Raftery et al. (1997) and Fortran code used by Fern´andez et al. (2001a).The Raftery et al. (1997) contribution is now published as an R package in Raftery et al. (2010).Source: §6 · p. 76

Software coverage

• Publicly available resources cover BMA, FMA, DMA, model confidence sets, information-theoretic selection, and specialized settings such as endogenous regressors, nonlinear effects, spatial models, and hyperpriors.Resources span R packages including BMS, BAS, BayesVarSel, ivbma, glmBfp, hypergsplines, eDMA, MCS, MuMIn, and AICcmodavg, alongside Matlab, gretl, and WinBUGS implementations.Source: §6 · pp. 76–77

Package comparisons

• Major R packages generally provide similar, reliable inference within about 10 minutes for problems with up to k = 100 covariates, while BAS is fastest but memory-intensive.BAS is recommended with model-visit-frequency estimation; BayesVarSel is preferred for small or moderate n when memory matters, whereas BMS is preferable when n is large.Source: §6 · p. 77

Online resources

• A dedicated BMA resource page by Feldkircher and Zeugner offers free software and introductory material, complementing other publicly available code and data repositories.Additional resources include BAS materials from Clyde, MMA and JMA code from Hansen, WALS implementations from Magnus, and BMA software or data from Raftery and Steel.Source: §6 · pp. 77–78

7 Conclusions

• Model uncertainty is pervasive in economics, while ignoring alternatives or conditioning inference on one selected model can produce overconfident, underestimated, and biased conclusions.Model selection is valid only in limited circumstances that are unlikely in economic applications.Source: §7 · p. 78
• Model averaging across appropriately chosen model spaces can help resolve persistent theoretical and empirical debates by encompassing competing explanations rather than privileging investigators’ preferred models.The author views growing adoption of model averaging as evidence of the gains from principled approaches to model uncertainty.Source: §7 · p. 78
• Bayesian and frequentist model averaging are complementary approaches, but BMA is especially attractive for its finite-sample interpretation, posterior inclusion and model probabilities, and scalability to very large model spaces.The choice between BMA and FMA depends partly on the investigation’s aims, while BMA also permits model selection when required.Source: §7 · p. 79
• BMA results depend on priors, so researchers should state assumptions explicitly, assess prior sensitivity, and use hyperpriors to make choices more robust and adaptive to the data.Hyperpriors can be elicited through quantities such as prior mean model size; under weak data they tend to favor smaller g, avoiding unwarranted precise distinctions between models.Source: §7 · p. 79
• Sensitivity analysis across priors and sampling models is indispensable for mapping assumptions to results and capturing the range of reasonable uncertainty in non-experimental economic research.The passage presents this explicit mapping as the key practical message of model averaging.Source: §7 · p. 79
• Model averaging is becoming a practical economics toolbox, aided by freely available software, and can improve communication by explaining differences in findings through model spaces, BMA priors, or FMA weights.It can also support more informed policy conclusions by formally accounting for model uncertainty.Source: §7 · pp. 79–80