On DSGE Models

The authors’ thesis is that “assessing the effects of a systematic policy change has to involve the use of a model” (p. 113), and that DSGE models are “the leading tool for making such assessments in an open and transparent manner.” Because DSGE models are built from explicit microfoundations, their assumptions are stated plainly enough to be criticized and checked against micro data – in the authors’ words, “micro data break the observational equivalence that was the bane of macroeconomists” (p. 114). They argue that smaller partial-equilibrium models can build intuition but “cannot be a substitute for DSGE models themselves,” because combining insights across separate small models requires informal judgment and hidden assumptions that a single general-equilibrium model makes explicit. Paraphrasing a remark by Fischer (2017) about Samuelson and Solow, they summarize their view as: “we’d rather have Stanley Fischer with a DSGE model than without one” (p. 114).

Real business cycle models (Kydland-Prescott 1982; Long-Plosser 1983) treated business cycles as efficient responses to shocks with no role for stabilization policy, but the authors say this view “crumbled under the impact of three forces”: micro data contradicting frictionless-labor-market assumptions (Chetty et al. 2011), failure to match aggregate facts such as hours volatility, the equity premium, and weak real-wage/hours co-movement (Christiano-Eichenbaum 1992; King-Rebelo 1999), and the fact that money plays no role at all, at odds with monetary interpretations of historical episodes. New Keynesian DSGE models grafted nominal frictions in labor and goods markets onto the RBC “chassis” – the authors suggest it “would be just as appropriate to refer to them as Friedmanite DSGE models,” since they embed Friedman’s (1968) view that money is neutral in the long run but affects real activity in the short run through sticky prices and wages. These models are required to satisfy the Fisherian property (a permanent policy change moves inflation and the nominal rate roughly one-for-one in the long run) and the anti-Fisherian property (a transitory policy change moves the nominal rate and inflation in opposite directions in the short run). The benchmark empirical target, reproduced from Christiano-Trabandt-Walentin (2010) using US data 1951Q1-2008Q4 under a recursive (Cholesky) identification of the monetary policy shock (in the tradition of Sims 1986, Bernanke-Blinder 1992, and Christiano-Eichenbaum-Evans 1996/1999), is that an expansionary funds-rate shock produces hump-shaped expansions in consumption, investment, and output alongside “relatively small rises in real wages and inflation” – a pattern the authors report as robust across lag lengths of one to five quarters and sample starts from 1951Q1 through 1985Q4, and one that high-frequency-identification (HFI) studies (Kuttner 2001; Gürkaynak-Sack-Swanson 2005; Gertler-Karadi 2015) are said to reach as well.

Matching the hump-shaped consumption, investment, and output responses requires habit formation in consumption (so consumption rises slowly rather than jumping and then falling), investment adjustment costs on the rate of change of investment (rather than on the capital stock, which would instead produce a monotone, wrong-shaped response), and Calvo (1983) nominal price and wage rigidities combined with features – sticky nominal wages and variable capital utilization – that keep real marginal cost “nearly acyclical,” which the authors call “critical for dampening movements in the inflation rate.” In the Christiano-Eichenbaum-Trabandt (2016) Bayesian impulse-response-matching re-estimation of the Christiano-Eichenbaum-Evans (2005) model, the posterior mode implies firms change prices about once every 2.3 quarters, households change wages about once a year, the habit coefficient is 0.75, and the elasticity of investment to a one percent temporary rise in the price of installed capital is 0.16. The authors stress that this fit is not robust to removing any of the three ingredients: a flexible-wage counterfactual “deteriorates drastically,” and habit formation and investment adjustment costs are each described as “critical” to reproducing the hump shapes and the muted real-wage and inflation responses.

Pre-crisis DSGE estimation relied on log-linear approximations around the steady state – justified as accurate for the size of shocks typically considered and as giving access to standard linear time-series tools, at a time when fully nonlinear estimation was computationally infeasible. Within that log-linear framework, the literature used several parameter strategies: calibration; limited-information methods such as moment-matching and impulse-response matching (as in Christiano-Eichenbaum-Evans 2005, Altig et al. 2011, Iacoviello 2005, and Rotemberg-Woodford 1991, including variants using external instruments or high-frequency identification); and full-information Bayesian estimation (following Smets-Wouters 2003 and later work). The authors describe Bayesian priors as serving as “a vehicle” for bringing micro-level evidence – for example, evidence on how often firms actually reprice – to bear on macro estimation, and for penalizing parameter values that are implausible on other grounds.

The authors’ preferred account of the crisis (shared with Bernanke 2009) is that it was a rollover crisis in a large, highly levered shadow-banking sector that funded long-term assets with short-term debt: a housing-price decline beginning in mid-2006 (after prices had roughly two-and-a-half times their 1991 level on the S&P/Case-Shiller index) devalued shadow-bank assets and triggered fire sales. Facing this, they accept the first criticism of DSGE modeling – that it failed to signal the sector’s vulnerability – as “correct,” but argue this “reflected a broader failure of the economics community” rather than a defect specific to the DSGE approach. They answer the second criticism, that pre-crisis DSGE models included insufficient financial frictions, by pointing to the pre-crisis empirical record: postwar US recessions had not been closely tied to financial disturbances (even the late-1980s savings-and-loan crisis stayed geographically and sectorally localized), and the financial-accelerator mechanism (Bernanke-Gertler-Gilchrist 1999) that some pre-crisis models did include is reported to have had only “a modest quantitative effect” on estimated model dynamics (citing Lindé-Smets-Wouters 2016, Kocherlakota 2000, and Brzoza-Brzezina-Kolasa 2013).

Post-crisis DSGE work has added financial frictions inside financial institutions, frictions on the borrower side, zero-lower-bound (ZLB) nonlinearities, state-dependent fiscal multipliers, an unresolved “forward guidance puzzle,” and heterogeneous-agent structures. On institutions, Gertler-Kiyotaki (2015) and Gertler-Kiyotaki-Prestipino (2016) model rollover crises and fire sales, illustrated with a stylized shadow-bank balance sheet in which roughly a 10-unit decline in asset value can turn crisis-state net worth negative and make a self-fulfilling rollover crisis possible. On the borrower side, Christiano-Motto-Rostagno (2014) introduce “risk shocks” – time-varying cross-sectional dispersion in firms’ returns – which the essay reports account for about 60 percent of the variance of US business cycles, versus roughly 13 percent for technology shocks (contrasted with the smaller role found by Justiniano-Primiceri-Tambalotti 2010); related work covers housing (Iacoviello-Neri 2010; Liu-Wang-Zha 2013). On the ZLB, which bound in December 2008, Christiano-Eichenbaum-Trabandt (2015) attribute most of the Great Recession to financial frictions interacting with a binding ZLB, with the fall in total factor productivity and a rise in the cost of working capital said to explain the surprisingly small observed drop in inflation; Lindé-Trabandt (2018) explore nonlinear price/wage setting as an alternative mechanism, and Gust et al. (2017) solve a fully nonlinear ZLB model. On fiscal policy, the government-spending multiplier is described as “larger the more binding the zero lower bound,” with Christiano-Eichenbaum-Rebelo (2011) reporting a multiplier “much larger than one” at the ZLB and “substantially below one” away from it. On forward guidance, standard models are said to make it “implausibly powerful” (the “forward guidance puzzle,” per Del Negro-Giannoni-Patterson 2012 and Carlstrom-Fuerst-Paustian 2015), an effect that is muted under sticky information (Kiley 2016), heterogeneous agents, or departures from full rational expectations (Gabaix 2016; Woodford 2018). Finally, heterogeneous-agent New Keynesian (HANK) models (Kaplan-Moll-Violante 2018; McKay-Nakamura-Steinsson 2016) suggest monetary policy works less through representative-household intertemporal substitution and more through a multiplier process operating among a subset of “intertemporal adjuster” households.

DSGE models are used in the Federal Reserve’s Tealbook A/B forecasting process (per Fischer 2017) via the SIGMA and EDO models, and by the European Central Bank, the IMF, the Bank of Israel, the Czech National Bank, the Riksbank, the Bank of Canada, the Swiss National Bank, and Norges Bank. As evidence of usefulness, the authors note that the New York Fed’s DSGE model (Cai et al. 2018, a variant of Christiano-Motto-Rostagno 2014) forecast the slow pace of the post-crisis recovery better than the FOMC’s median projection, concluding that “policymakers have voted with their collective feet” in continuing to build and rely on these models.

The authors rebut Stiglitz (2017) on four specific points, concluding that his “criticisms are not informed” (p. 135). First, on econometrics, they argue modern DSGE estimation does not rely on HP-filtered data as Stiglitz suggests, but instead uses full-information maximum likelihood or GMM, and that endogenous-growth DSGE variants (Comin-Gertler 2006) can handle nonstationary data directly. Second, on financial frictions, they point out that pre-crisis DSGE models did in fact incorporate them, citing Galí-López-Salido-Vallés (2007), Carlstrom-Fuerst (1997), Bernanke-Gertler-Gilchrist (1999), Christiano-Motto-Rostagno (2003), and Iacoviello (2005). Third, on interest-rate spreads, they offer counterexamples in which a spread is a central endogenous variable in the model rather than absent from it. Fourth, on heterogeneity, they argue Stiglitz’s characterization is at odds with the active heterogeneous-agent (HANK) literature the essay itself surveys in Q6.

The authors explicitly state that DSGE models “will not reliably predict the next crisis,” describing DSGE modeling instead as “an organic process that involves the constant interaction of data and theory.” They characterize the pre-crisis vintage of models as having real shortcomings that the crisis exposed, and post-crisis progress (financial frictions, heterogeneity, departures from rational expectations) as a response to those shortcomings rather than a closed, finished program. Their closing claim is a comparative rather than an absolute one: “there is simply no credible alternative to policy analysis in a world of competing economic forces operating on different parts of the economy” (p. 136).