Macroeconomic Dynamics, 18, 2014, 338–368. Printed in the United States of America. doi:10.1017/S1365100512000417 CREDIT CONSTRAINTS, LEARNING, AND AGGREGATE CONSUMPTION VOLATILITY DANIEL L. TORTORICE Brandeis University This paper documents three facts. First, the volatility of consumption growth relative to income growth rose from 1947 to 1960 and then fell dramatically by 50% from the 1960s to the 1990s. Second, the correlation between consumption growth and personal income growth fell by about 50% over the same time period. Finally, the absolute deviation of consumption growth from its mean exhibits one break in U.S. data, and the mean of the absolute deviations has fallen by about 30%. A standard dynamic stochastic general equilibrium model is unable to explain these facts. I examine ability of two hypotheses to account for these facts: a fall in credit constraints and changing beliefs about the permanence of income shocks. I find evidence for both explanations. The beliefs explanation is more consistent with the data. Keywords: Consumption, Business Fluctuations, Learning, Liquidity Constraints 1. INTRODUCTION This paper establishes three facts about aggregate U.S. consumption. First, after rising from the 1950s to the 1960s, the standard deviation of consumption growth relative to income growth fell 40% from the 1960s to the 1970s and then another 30% from the 1980s to the 1990s. Finally, it rose about 25% during the most recent recession. Second, the correlation between consumption growth and income growth has fallen 50% over this same period. Finally, there is a structural break in the size of the absolute deviation of consumption growth from its mean. The break is in 1977, and the total fall in the absolute deviation of consumption growth from its mean is 30%. These facts complement the Great Moderation facts by showing that consumption volatility has fallen more than might have been expected based on the fall in income volatility alone. A previous version of this paper was titled “Income Process Uncertainty, Learning and Aggregate Consumption Volatility.” I thank Robert Barro, John Campbell, Gita Gopinath, George Hall, David Laibson, N. Gregory Mankiw, Catherine Mann, Blake LeBaron, James Stock, Aleh Tsyvinski, two anonymous referees, and attendees of the Harvard macroeconomics lunch seminar, the Brandeis faculty workshop, and the Society for Economic Dynamics 2010 Annual Meeting for helpful comments, suggestions, and advice. I am grateful to the Jacob K. Javits Fellowship Program at the U.S. Department of Education, Harvard University, and the Brandeis University Theodore and Jane Norman Fund for financial support. All errors are mine. Address correspondence to: Daniel L. Tortorice, Department of Economics, Brandeis University, Mailstop 021, 415 South St., Waltham, MA 02454-9110, USA; e-mail: tortoric@brandeis.edu. c 2012 Cambridge University Press  1365-1005/12 338 Downloaded from http:/www.cambridge.org/core. University of Florida, on 25 Dec 2016 at 16:13:41, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms. http://dx.doi.org/10.1017/S1365100512000417 CREDIT, LEARNING, CONSUMPTION VOLATILITY 339 This paper joins a growing literature suggesting a changing relationship between consumption and income. Stock and Watson (2003) estimate structural breaks in consumption. Dynan et al. (2006, 2009) note that the marginal propensity to consume has fallen substantially. This paper, however, is the first to document all three outlined facts and study two different hypothesis: a reduction in credit constraints and changing beliefs about the productivity process as explanations. I model credit constraints as in Ludvigson (1999), in which impatient consumers face a binding debt limit. I estimate a fraction of credit-constrained consumers that varies over time and use this model to simulate a consumption series. I examine how well this series replicates the stated consumption facts. It is not obvious that reduced credit constraints reduce the volatility of consumption. If individuals want to smooth income shocks, then reducing financial constraints makes consumption smoother. However, if individuals want to borrow to respond more than one-to-one with income shocks—for example, if shocks are permanent—then reduced financial constraints could increase consumption volatility. I therefore model uncertainty as to the permanence of income shocks. As pointed out by Deaton (1992), if income shocks are persistent but transitory, consumption is less volatile than income. However, if shocks are permanent, consumption may be more volatile than income. Cochrane (1988) and Stock (1991) show that distinguishing between these two models in samples the length of the U.S. macroeconomic time series is very difficult. Because these two income processes are difficult to distinguish, one might expect consumers to have uncertainty as to which is the true model. Their beliefs may change over time, resulting in time-varying consumption-to-income volatility.1 To examine the plausibility of this hypothesis, I study a learning model based on Cogley and Sargent (2005). My model is a standard dynamic stochastic general equilibrium model in which output fluctuations are driven by changes in productivity, with one modification. The agent believes the productivity process is nonstationary with some probability and stationary otherwise. At each point in time the individual first updates her beliefs about the parameters of the two models, and then, using Bayes’s rule, updates her beliefs about the probability that each model is true. Based on these beliefs, she chooses her optimal level of consumption. From this model I can simulate a consumption series and examine the model’s ability to capture the previously outlined facts. I find that two benchmark dynamic stochastic general equilibrium (DSGE) models with a known productivity process fail to predict the fall in consumption volatility. The learning model, however, is able to partly explain the rise and fall in consumption volatility early in the sample and also the fall and rise in consumption volatility toward the end of the sample. The credit constraints model is consistent with the early rise and fall in consumption volatility, but cannot explain the magnitude of the fall in consumption volatility. Importantly, only a model with learning about the productivity process can replicate the estimated break. Finally, Downloaded from http:/www.cambridge.org/core. University of Florida, on 25 Dec 2016 at 16:13:41, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms. http://dx.doi.org/10.1017/S1365100512000417 340 DANIEL L. TORTORICE the learning model’s implied probability weights on the different productivity processes have significant explanatory power for consumption changes. The learning model also has an implication for recent economic events. During the current recession, the weight on the nonstationary model has increased substantially; the model implies a large drop in consumption, consistent with the data. This is perhaps one of the most interesting predictions of the learning model. This paper relates to three strands of literature. First, it relates to the literature on learning in macroeconomic models [Evans and Honkapohja (2001); Georges and Wallace (2009)]. It shows how relaxing the strict rational expectations assumption in favor of learning results in a significantly different optimal policy. In my learning model, the two possible productivity processes are difficult to distinguish in small samples. Hence, learning which process is true takes a long time. However, because consumption volatility is sensitive to deviations from stationarity, the processes imply different optimal policies. Learning then introduces a very different choice of consumption than in the no-learning model. As in Waters (2009) and Tillman (2011), the inclusion of uncertainty changes the optimal policy. The paper also relates to some recent research on consumption. Guvenen (2007) finds that introducing learning into a life-cycle consumption model changes many of the model’s predictions. Specifically, optimal consumption choices are very different when one learns about the trend in income than when it is known. The paper also relates to Aguiar and Gopinath (2007). They explain the differences in consumption volatility relative to income volatility across countries. In their model, different countries have different ratios of consumption volatility to income volatility because they have different productivity processes. In my paper, the variance of consumption relative to income varies over time depending on the relative likelihood of two different productivity processes. Finally, this paper relates to the Great Moderation literature. Although an extensive review would be out of place, I note three closely related papers. Cecchetti et al. (2005) show that the reduction in consumption volatility was accompanied by a rise in U.S. debt levels, supporting a financial innovation explanation. Dynan et al. (2006) show that marginal propensities to consume have fallen over time. Cecchetti et al. (2006) relates changes in consumption volatility across countries to changes in estimated fractions of rule-of-thumb consumers. This paper differs for three reasons. First, the question is different. I ask if financial innovation can account for the fact that consumption volatility has declined more than personal income volatility in U.S. data. Second, I simulate a model of credit constraints and test the model directly to see if it is consistent with the decline in consumption volatility. Finally, I solve an innovative learning model, and it explains features of the data that the credit constraints model does not. The rest of the paper proceeds as follows. Section 2 describes the data I use. Section 3 describes the empirical facts I attempt to explain. Section 4 describes the credit constraint and learning models. Section 5 tests these models, and Section 6 provides more intuition for the learning model’s predictions. Section 7 examines robustness to different parameter choices. Section 8 concludes. Downloaded from http:/www.cambridge.org/core. University of Florida, on 25 Dec 2016 at 16:13:41, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms. http://dx.doi.org/10.1017/S1365100512000417 CREDIT, LEARNING, CONSUMPTION VOLATILITY 341 2. DATA Data come from the National Income and Product Accounts (NIPAs). The consumption data are aggregate consumption of nondurable goods and services, and income is the personal income series. Data begin in 1947 and end in 2010:1.2 Series are transformed into real per capita series by deflating with the nondurable consumption deflator, the service consumption deflator, and the GDP deflator, respectively, and dividing by population. I pay particular attention to the distinction between accrued wages and distributed wages. In general, firms distribute most of the wages accrued or earned within a quarter. However, at times firms distribute wages early or late. Occasionally this difference is large. For example, in 1992:Q4 firms distributed 63 billion dollars more in wages than employees earned, and in 1993:Q1 employees earned 72.1 billion dollars more in wages than firms distributed. These discrepancies appear because of increases in income tax rates.3 These distributions are large enough to affect learning about the productivity process. They appear as large shocks to income, though they do not represent uncertainty. As a result, I assign to income all wages earned in a quarter regardless of how much is distributed. However, this modification does not affect the conclusions of the paper. 3. EMPIRICAL FACTS This section establishes three facts about the aggregate U.S. consumption series. First, the volatility of consumption relative to income rose early in the sample and then fell substantially over the last 45 years. Second, the correlation of consumption growth and income growth has fallen over time. Finally, there is a substantial break in the absolute deviation of consumption growth from its mean, falling about 30% over the last 60 years. 3.1. Consumption Volatility Relative to Income Volatility In Figure 1, I plot the rolling standard deviation of consumption growth divided by the rolling standard deviation for income growth. I use a window over the next 10 years and nondurables and services consumption.4 Results are robust to varying the window from 20 to 50 quarters. As one can clearly see, consumption variance relative to income variance rises over time, reaching a peak around 1960, and then falls quite dramatically by 40% from 1960 to 1970. It remains constant for about 20 years before beginning to fall again in 1990 by an additional 30%. During the most recent recession it increased by 25%. I obtained similar results (omitted) with nondurables alone, finding a similar, but more gradual decline beginning in 1960 and ending in 1980. Downloaded from http:/www.cambridge.org/core. University of Florida, on 25 Dec 2016 at 16:13:41, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms. http://dx.doi.org/10.1017/S1365100512000417 342 DANIEL L. TORTORICE 1.2 1 0.8 0.6 0.4 0.2 1950 1960 1970 1980 1990 2000 FIGURE 1. Ratio of the standard deviation of consumption growth to income growth. Dashed lines represent the 95% confidence interval. 3.2. Time-Varying Response of Consumption to Income The data demonstrate that consumption changes appear to have moderated over time relative to income changes. To get a parametric representation of this fact, I estimate a time-varying coefficient in the regression  ln ct = α + mt  ln yt + εt , (1) where ct is the NIPA value of real per capita nondurable and services consumption and yt is real per capita personal income. I let mt take the form mt = m0 + m1 t.5 I found that higher-order terms (quadratic and cubic) were not significant. Consumption and income growth are measured as annualized percentage changes. The results, in column (1) of Table 1, indicate that there is a significant timevarying component to m, as m1 is statistically significant and negative. At the beginning of the sample period mt = 0.55; by the end of the sample period mt is around 0.276. This result again shows a considerable moderation of consumption relative to personal income. Although income has moderated over time, consistent with the Great Moderation facts, consumption has moderated even more. This approach is one way of measuring a marginal propensity to consume. Dynan et al. (2009) estimate a marginal propensity to consume, controlling for potential output, interest rates, the unemployment rate, and the index of consumer sentiment. They also find a substantial fall in the response of consumption to income. Downloaded from http:/www.cambridge.org/core. University of Florida, on 25 Dec 2016 at 16:13:41, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms. http://dx.doi.org/10.1017/S1365100512000417 CREDIT, LEARNING, CONSUMPTION VOLATILITY 343 TABLE 1. Relationship between income growth and consumption growth Benchmark models Data 0.55∗∗∗ [0.08] −0.124∗∗∗  ln(inc.)∗ (time/100) [0.043] Constant 1.17∗∗∗ [0.19] Obs. 221  ln(inc.) Nonstationary Stationary Learning 1.03∗∗∗ [0.04] 0.021 [0.021] −0.11 [0.08] 221 0.50∗∗∗ [0.03] −0.009 [0.017] 1.17∗∗∗ [0.05] 221 Credit Credit + learning 0.70∗∗∗ 0.73∗∗∗ 0.81∗∗∗ [0.04] [0.02] [0.02] −0.073∗∗∗ −0.050∗∗ −0.061∗∗∗ [0.025] [0.024] [0.012] 0.90∗∗∗ 0.76∗∗∗ 0.60∗∗∗ [0.08] [0.07] [0.04] 221 221 221 Note: This table reports the result of regressing annualized consumption growth from the data and the predictions of the different models on the annualized income growth rate, allowing the coefficient to vary by time. Newey–West standard errors with five lags in brackets. ∗ Significant at 10%; ∗∗ Significant at 5%; ∗∗∗ Significant at 1%. 3.3. Breaks in the Variance of Consumption Because consumption has moderated more than income volatility, it is natural to examine how much consumption has moderated. To investigate, I follow the methodology of Stock and Watson (2003) and estimate a break in the mean of the absolute value of the residual from the regression of nondurable consumption growth on a constant.6  ln ct = α + ηt , |ηt | = ζ + ζ1 τ + εt . (2) (3) To estimate τ (the break dates), I use the methodology of Bai and Perron (1998) and the algorithms and GAUSS code available in Bai and Perron (2003). The methodology, in addition to providing a feasible way to estimate the structural break model, describes a sequential method for estimating the number of breaks (allowing for the possibility of more than one break). These methods consistently estimate the number of breaks and the proportion of the sample that occurs before the break date. To estimate the number of breaks, Bai and Perron (1998) recommend the sequential application of a test of l breaks versus the alternative of l +1 breaks, called the sup F (l + 1|l) test. The procedure first tests one break versus the alternative of zero breaks, and then two versus one, and so on until the statistic fails to reject the null hypothesis. In addition, Bai and Perron provide a test statistic, called the U D max statistic, that tests the hypothesis of zero breaks against the alternative of k > 0 breaks, where k is unknown. The estimates of the break dates are those that minimize the sum of squared residuals in equation (3). Downloaded from http:/www.cambridge.org/core. University of Florida, on 25 Dec 2016 at 16:13:41, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms. http://dx.doi.org/10.1017/S1365100512000417 344 DANIEL L. TORTORICE TABLE 2. Breaks in consumption volatility Benchmark models Data Nonstationary Stationary Learning Credit Credit + learning 7.29∗ 3.86 supF(1|0) supF(2|1) 7.38∗ 3.44 2.32 5.24 3.2 4.64 9.47∗ 3.71 5.26 4.33 UD Max 7.38 6.07 4.4 10.61∗ 5.26 7.29 Break 90% CI 1977:4 1957:2–1993:1 — — — — 1982:1 1973:2–1995:3 — — 1984:3 1971:2–2000:2 2.65 (0.20) 1.87 (0.17) — — — — — — — — 1.5 (0.11) 0.94 (0.11) — — — — 1.77 (0.12) 1.21 (0.13) δ1 δ2 Note: This table reports the results from testing for a break in the residual of the annualized consumption change regressed on a constant. The means of the residual are given by δ, for before and after the breaks. supF and UDMax are tests for L+1 breaks vs. L and any breaks respectively and the critical values for the tests from Bai and Perron (2003) are given as follows. (star denotes >= 10% significance): Critical values for tests 10% supF(1|0) supF(2|1) supF(3|2) UD Max 5% 1% 7.04 8.58 12.29 8.51 10.13 13.89 9.41 11.14 12.66 7.46 8.88 12.37 Table 2 presents the results from these tests.7 Consumption growth is measured as an annualized percentage change. The consumption series has one break: at 1977:4. The means for each segment are 2.65 and 1.87. This implies a moderation in consumption volatility of 30%. 4. EXPLANATIONS This section outlines the two models I use to explain the fall in consumption volatility relative to income volatility. The first model is a model of credit constraints with a time-varying fraction of credit-constrained consumers. The second is a model of learning in which the true productivity process is unknown and therefore agents’ beliefs about the true process are changing over time. 4.1. Credit Constraints This section of the paper outlines a model with time-varying credit constraints as an attempt to explain the previously described empirical facts. For another example analyzing the interaction of credit constraints and consumption see Benito and Mumtaz (2009). Consumption may be very volatile early in the sample because Downloaded from http:/www.cambridge.org/core. University of Florida, on 25 Dec 2016 at 16:13:41, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms. http://dx.doi.org/10.1017/S1365100512000417 CREDIT, LEARNING, CONSUMPTION VOLATILITY 345 individuals are credit-constrained, but become less volatile as more assets are accumulated or better consumption insurance becomes available. To examine this explanation, I assume that aggregate consumption comes from two types of agents: an unconstrained agent and a constrained agent: ct = λt ctcc + (1 − λt )ctuc . (4) Here λt is the fraction of the population that is credit-constrained, ctcc is the change in consumption of the credit-constrained consumers, and ctuc is the consumption change of the unconstrained consumer. I model the constrained consumer as in Ludvigson (1999). In this model the agent chooses consumption to maximize  1−γ  ∞  Ct+i Et βi (5) 1 −γ i=0 subject to Dt+1 = (1 + r)(Dt + Ct − Yt ), where Dt is consumer debt, r is the real interest rate, and Yt is labor income. Labor income evolves according to  ln Yt = g +ηt −ψηt−1 , where ηt is i.i.d. Consumers are credit-constrained: they can borrow only up to D t+1 = w1 Yt exp(ξt ), where ξt+1 = φξt + ut+1. Therefore, −Dt (1+r) + Yt . maximum consumption at time t is given by Xt = Dt+1(1+r) zt+1 = Yt+1 /Yt , we can solve for the policy If we define wt = Xt /Yt and  function θt (wt , ηt , ξt ) = Ct /Yt , as a function of the state variables, by iterating on the Euler equation. Defining the marginal-utility-of-consumption function −γ pt (wt , ηt , ξt ) = v(θt ) = θt , Ludvigson (1999) shows that   −γ zt+1 p(wt+1 , ηt+1 , ξt+1 ) , (6) pt (wt , ηt , ξt ) = max v(wt ), β ∗ Et −1 t+1 −ψηt +ξt+1 −ξt +g)−(1+r) zt+1 (1+r)[wt −v −1 pt (wt , ηt , ξt )]+ exp(η where wt+1 = 1+ (1+r) exp(ηt+1 −ψηt +g+ln ω−ξt ) ∗ and β = (1 + r)β. I solve for pt (wt , ηt , ξt ) by iterating on (6) using five grid points for η and ξ . I discretize the i.i.d. ηt and the AR(1) ξt using the method of Tauchen and Hussey (1991).8 Finally, I use 100 grid points for w and linear interpolation to calculate the function between grid points of w. To calculate the predicted aggregate consumption change I need to estimate λt . To do so let ctuc = μ + σ rt + εt , where rt is the ex post real interest rate and εt is i.i.d., because the unconstrained consumer does not respond to predictable p variation in income, and let ctcc = α + σ c rt + mpct yt + νt where vt is p i.i.d., yt is the predicted change in income, and mpct is the credit-constrained individual’s marginal propensity to consume out of predictable income estimated from the model.9 Then p or ct = λt (α + σ c rt + mpct yt + νt ) + (1 − λt )(μ + σ rt + εt ) (7)  p ct = μ + σ rt + λt mpct yt + σ c − σ rt + α − μ + ηt . (8) Downloaded from http:/www.cambridge.org/core. University of Florida, on 25 Dec 2016 at 16:13:41, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms. http://dx.doi.org/10.1017/S1365100512000417 346 DANIEL L. TORTORICE 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 1950 1960 1970 1980 1990 2000 2010 FIGURE 2. Estimated fraction of credit-constrained consumers. p To estimate this equation, I first calculate yt as the fitted value of a regression of yt on the lagged income growth rate, the lagged change in the S&P 500 index, and the lag ex post real interest rate; then I multiply by the estimated mpct .10 Because the ex post real interest rate is also correlated with ηt , I instrument its value with the same instruments used for income growth to obtain rt . The final equation I estimate is  p rt + λt mpct yt + σ c − σ rt + α − μ + ηt . ct = μ + σ (9) I estimate the time-varying λt using the Kalman filter [see Hamilton (1994, pp. 399–402)].11 Figure 2 plots the resulting series for the estimated fraction of consumers who are credit-constrained. Credit constraints rise in the 1960s, with a peak of 80% credit-constrained. Credit constraints fall nonmonotonically to a low of 10% in 2010, though most of that fall is in the last few years. As late as 2005, 40% of the population was still estimated to be credit-constrained. I follow Campbell and Mankiw (1990) and estimate the fraction of creditconstrained consumers using data from 1955 onward. This avoids the Korean War period and some large income swings due to government transfer payments. To estimate credit constraints pre-1955, I assume that credit constraints do not change from 1947 to 1955. This assumption is only for completeness and does not affect the results. Downloaded from http:/www.cambridge.org/core. University of Florida, on 25 Dec 2016 at 16:13:41, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms. http://dx.doi.org/10.1017/S1365100512000417 CREDIT, LEARNING, CONSUMPTION VOLATILITY 347 4.2. The Learning Model This section describes a model economy in which the representative agent maximizes consumption while also investing in a capital stock. The agent combines capital and labor to produce output. Productivity grows at a constant rate g, but is subject to stochastic shocks. The agent is learning if the correct model for the productivity shocks is one with permanent or temporary deviations from trend, and also about the parameters of each productivity model. Model description. mizes For the learning model, the representative agent maxi- max {ct+j ,lt+j ,it+j }∞ j =0 Et ∞  j =0 βj {ct+j [1 − (θ lt+j )φ ]}1−γ 1−γ (10) subject to, with probability p = ps,t , kt+1 = (1 − δ)kt + it , ktα [A0 (1 + g) e lt ] t zt 1−α = ct + it , zt = θ1S zt−1 + ... + θpS zt−p + εts , (11) (12) (13) and with probability p = pns,t = 1 − ps,t , kt+1 = (1 − δ)kt + it , ktα [A0 (1 + g) e lt ] t zt 1−α = ct + it , zt = θ1NS zt−1 + ... + θpNS zt−p+1 + εtns . (14) (15) (16) Here kt is beginning-of-the-period capital stock, yt is output, ct is consumption, it is investment, lt is labor supply, and A0 (1 + g)t ezt is the productivity level. With probability ps,t , the log deviation of productivity from its trend is given by (13) and is therefore stationary around the time trend; with probability pns,t , the log deviation of productivity from its trend is given by the unit root process (16) and hence exhibits permanent deviations from the trend.12,13 Updating of beliefs. To choose consumption, the agent each period must update his beliefs concerning the probability that each model is true. I update beliefs as in Cogley and Sargent (2005). There are two models of the productivity process, indexed i = s, ns, which can  be written in regression form as yi,t = xi,t θi,t + εi,t . For the stationary model the regression equation is (13); for the nonstationary model the equation is (16). t represent all observations on yt and xt up to time t, the agent’s prior Letting Z beliefs on the parameters for each model are given by −1 t−1 ) = N θt−1 , σ 2 Pt−1 , (17) p(θ |σ 2 , Z p(σ 2 |Z t−1 ) = IG(st−1 , vt−1 ). (18) Downloaded from http:/www.cambridge.org/core. University of Florida, on 25 Dec 2016 at 16:13:41, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms. http://dx.doi.org/10.1017/S1365100512000417 348 DANIEL L. TORTORICE −1 θt−1 is the parameter estimate based on t − 1 data, σ 2 Pt−1 is the estimate of the variance–covariance matrix of θt−1 , st−1 is the residual sum of squares, and vt−1 is the degrees of freedom for estimating variance of the residuals (t − k). IG is the inverse gamma distribution. I follow Cogley and Sargent (2005) and use the normal and inverse gamma distributions because these allow one to update the parameters recursively and use analytical formulas to update the posterior beliefs. Stepping out of the normal–inverse gamma family would involve computationally intensive simulations at each point in time to calculate the likelihood (23) of each model. Bayesian updating implies that the parameters are updated recursively according to Pt = Pt−1 + xt xt , (19) θt = Pt−1 (Pt−1 θt−1 + xt yt ),   (20)  st = st−1 + yt yt + θt−1 Pt−1 θt−1 − θt Pt θt , (21) vt = vt−1 + 1. (22) Next the agent updates the probability weights on each model. For each model, the marginalized likelihood is  t p ys |xs , θi , σi2 p θi , σi2 dθi dσi2 mit = (23) s=1 and the probability weight is wit = mit pi,0 , where pi,0 is the prior probability on model i. To calculate the marginalized likelihood, Cogley and Sargent (2005) note that Bayes’s rule implies, for any θi , σi2 , t  t = p θi , σi2 |Z s=1 mit t  mit = p(ys |xs , θi , σi2 )p(θi , σi2 ) s=1 , (24) . (25) p(ys |xs , θi , σi2 )p(θi , σi2 ) t ) p(θi , σi2 |Z Therefore, wi,t+1 mi,t+1 pi,0 = , wi,t mi,t pi,0 = p(yt+1 |xt+1 , θi , σi2 ) (26) t p θi , σi2 |Z . t+1 p θi , σi2 |Z (27) Downloaded from http:/www.cambridge.org/core. University of Florida, on 25 Dec 2016 at 16:13:41, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms. http://dx.doi.org/10.1017/S1365100512000417 CREDIT, LEARNING, CONSUMPTION VOLATILITY 349 As Cogley and Sargent (2005) show, this expression can be evaluated analytically, because  p yt+1 |xt+1 , θi , σi2 = N (yt+1 − xt+1 θi , 0, σi ), −1 , st+1 , vt+1 ), p θi , σi2 |Zt+1 = NIG(θi , σi ; θt+1 , Pt+1 (28) (29) where N and NI G indicate the normal and the normal-inverse gamma probability density functions, respectively. There are analytical expressions for both distributions, which are available in Cogley and Sargent (2005). Finally, to get the actual probabilities for each model, one normalizes the weights to sum to one: pNS,t = wNS,t , wNS,t + wS,t (30) pS,t = 1 − pNS,t (31) Model solution. Now that we can calculate the agent’s beliefs at each point in time, I can solve for the optimal consumption, investment, and labor supply plans. I first normalize the trending variables by the nonstochastic level of productivity and then take a linear–quadratic approximation to the utility function around the nonstochastic steady state. Subsequently, I cast the model as a stochastic optimal linear regulator problem and solve it using standard methods [e.g., of Ljungovist and Sargent (2004, Ch. 5) and Cogley and Sargent (2005)].14 For a given variable st , letting st = st /(1 + g)t (normalizing A0 = 1) and substituting the resource constraint into the utility function, we can rewrite the objective as max {lt+j , it+j }∞ j =0 Et ∞  j =0 β ∗j φ α {[ kt+j (ezt lt+j )1−α − it+j ][1 − θ lt+j ]}1−γ 1−γ , (32) where β ∗ = β(1 + g)1−γ , and rewrite the capital evolution equation as kt + it . (1 + g) kt+1 = (1 − δ) (33) Next, I take a linear–quadratic approximation to the utility function. See the Appendix for details. Following Cogley and Sargent (2005), I write each submodel (i.e., the model for each possible productivity process) as a linear regulator problem and then stack the model matrices to cast the complete model as an optimal linear regulator. For the stationary model, the state variables are xtS = [1 zt kt zt−1 ... zt−p+1 ] and the control variables are ut = [lt it ] . For the nonstationary model, the state variables are xtNS = [1 zt kt zt zt−1 ... zt−p+2 ] and the control variables are ut = [lt it ] . Then for model i = {s, ns}, I can write the problem in the linear Downloaded from http:/www.cambridge.org/core. University of Florida, on 25 Dec 2016 at 16:13:41, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms. http://dx.doi.org/10.1017/S1365100512000417 350 DANIEL L. TORTORICE regulator format, Uti = max∞ Et {ut+j }j =0 ∞  i i i i β ∗j xt+j R xt+j + ut+j Qut+j + 2xt+j W i ut+j , (34) j =0 subject to i i xt+1 = Ait xti + B i ut + C i εt+1 . i i i (35) , Ait , B i , C i The matrices R , Q , W are given in the Appendix. For the complete model, the agent seeks to maximize15 max ps,t Et Uts + (1 − ps,t )Et Utns . (36) {ut+j }∞ j =0 We can write this as an optimal control problem by stacking the matrices of the submodels, max∞ Et {ut+j }j =0 ∞     β ∗j (xt+j Rxt+j + ut+j Qut+j + 2xt+j W ut+j ), (37) j =0 subject to xt+1 = At xt + But + Cεt+1 , (38) where xt = [xts xtns ] ,  0 ps,t R S , R= 0 pns,t R ns   S 0 At At = , 0 ANS t   ps,t W s W = , pns,t W ns    Bs B= , B NS   Cs C= . C NS 5. TESTING THE MODELS EMPIRICALLY 5.1. Credit Model Simulation To simulate a consumption series from the credit constraint model, I first need to parameterize the model. I use the Ludvigson (1999) calibration adjusting for 1 quarterly data: ψ = 0.44, r = 0.03/4, β = 1+0.15/4 , φ = 0.6, ω = 15, γ = u y 2, σ = 0.0025, and σ = 0.025. I next need to simulate the time path for household income. I first assume an economy with 5,000 households, all described by the credit constraint model. As in Ludvigson (1999), the household income process is decomposed into an aggregate shock and two idiosyncratic shocks, as in h h h + ε3,t − ε3,t . Here, ε1 is the aggregate shock and ηt − ψηt−1 = ε1,t + χ ε1,t−1 + ε2,t h h ε2 and ε3 are household-specific shocks. We need two household shocks to match observed positive autocorrelation in aggregate income and negative autocorrelation in household income. I estimate χ and the ε1 time series from an ARIMA(0,0,1) model on the real per capita personal income growth series. The ARIMA estimate Downloaded from http:/www.cambridge.org/core. University of Florida, on 25 Dec 2016 at 16:13:41, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms. http://dx.doi.org/10.1017/S1365100512000417 CREDIT, LEARNING, CONSUMPTION VOLATILITY 351 also gives the growth rate of income g = 0.0052. I then estimate the variances of the idiosyncratic shocks by matching the variance of the household income series and its autocorrelation. I simulate the household shocks as random draws from a normal distribution. Finally, to get a time series for ξt (the shock to the debt limit), I draw a random chi from the grid according to the probabilities given by the method of Tauchen and Hussey (1991). Once I have each household’s consumption series, I can sum across households to obtain an aggregate consumption series. I match this to the data by scaling up the mean of this series to the mean of real per capita nondurables and services consumption in the data. Then I combine consumption changes from this scaled series with consumption changes from the learning model, using equation (4) to generate the finalized consumption series from the credit constraint model. I generate two series: a credit series that combines consumption from the credit constraint model with consumption from the learning model when individuals believe that the productivity process is stationary, and a credit and learning series that combines consumption from the credit constraint model with consumption from the learning model when individuals’ beliefs change over time. 5.2. Learning Model Simulation The calibration of the learning model is fairly standard: β = 0.99, α = 13 , δ = 0.025, and γ = 2. A higher γ leads to more volatile consumption, and I choose γ = 2 to better match the volatility of consumption. Finally, I choose θ and φ so that the agent inelastically spends 1/3 of his time working. This assumption, although not necessary to generate the observed learning dynamics, allows me to choose productivity so that income in the model exactly matches income in the data, as I describe later. It also allows us to interpret lt as potential labor supply, and therefore an increase in unemployment would be seen as a fall in productivity. Hence, in the current recession, model productivity is below trend, whereas some measures of labor productivity in the data are above trend. This assumption is also realistic, I believe, in the sense that consumers view increases in unemployment as the economy using its resources inefficiently. To construct the productivity series that the agent learns about, I take the real per capita personal income series16 and regress ln yt = a0 + gt + εt . Growth in the model is set to g and I calculate yt = ln yt − a0 − gt. I then choose zt so yt )y ss .17 This ensures that income in the model that model output ytmod = (1 + matches income in the data, allowing me to combine this model easily with the credit constraint model.18 One way to set the prior beliefs would be to run a regression on the first few quarters of data and set the priors on θ0 and P0 based on the regression estimates. However, this approach leads to unstable estimates of the stationary model, where productivity is predicted to go off to infinity. Instead, I take a different approach and assume that the agent has some knowledge of the productivity process, even though the quarterly data only begin in 1947. Downloaded from http:/www.cambridge.org/core. University of Florida, on 25 Dec 2016 at 16:13:41, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms. http://dx.doi.org/10.1017/S1365100512000417 352 DANIEL L. TORTORICE 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 1950 1960 1970 1980 1990 2000 2010 FIGURE 3. Probability weight on the nonstatationary model. I set θ0s = [0.99 0 0 0] and make this prior as diffuse as possible while still maintaining finite forecasts of productivity. This requires me to set P0S = 0.15×I. I set θ0ns = [0 0 0] and set P0NS = P0S . I set σ02 = 0.012 (which is the full sample estimate of the variance of the regression errors) and v0 = 1. To interpret these priors, remember that (17) implies that the standard deviation of the initial θ0 = 0.026. 19 Finally, I set k0 = k ss , ps,0 = 0.25, and lag = 4. Robustness to the choice of initial lag and prior probability in shown in Section 7. Finally, to compare model consumption to consumption in the data, I multiply model consumption by (1+g)t and scale consumption by the mean of income in the data divided by the mean of income in the model. From the learning model, I generate three consumption series: one based on learning about the parameters of the two income processes and the likelihood of each model, one based on setting ps,t = 0 for all t (nonstationary model), and one based on setting ps,t = 1 for all t (stationary model). For the latter two models, I set θ S and θ NS to the O.L.S. estimates of the complete productivity time series. They do not change over time. These last two series are my benchmark no-learning models. 5.3. Dynamics of the Relative Variances of Consumption and Income Growth Figure 3 plots the agent’s probability weight on the nonstationary productivity model implied by the learning model. Recall that the initial prior is 0.75 in 1947. As is evident from the graph, the probability falls and then rises, reaching a peak in Downloaded from http:/www.cambridge.org/core. University of Florida, on 25 Dec 2016 at 16:13:41, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms. http://dx.doi.org/10.1017/S1365100512000417 CREDIT, LEARNING, CONSUMPTION VOLATILITY 353 1.2 1 Data Learning Nonstationary Stationary 0.8 0.6 0.4 0.2 1950 1960 1970 1980 1990 2000 FIGURE 4. Ratio of standard deviation of consumption growth to income growth (models). 1962. It remains at that level for about 10 years and then falls, nonmonotonically, putting more and more weight on the stationary model as time passes. In the last few years the probability on the nonstationary model has risen substantially, as income has not returned to trend. This probability pattern suggests that consumption variance relative to income variance would rise early in the sample and fall later on. We saw that this was in fact true in U.S. data (Figure 1), so now we examine how well the model captures this pattern. Figure 4 plots the predictions for the variance of consumption relative to income from the models with and without learning. The no-learning, nonstationary-come model predicts that the ratio should have no trend. It also overestimates the level of the ratio. The no-learning, stationary model again predicts that this ratio should be roughly constant and substantially underestimates its level, especially early in the sample. In contrast, the learning model, shown in the middle of Figure 4, does much better than these benchmark models. It predicts a rise and fall in consumption volatility. There are a clear rise and fall in consumption variance relative to income variance, peaking at the same time as the peak in the data. However, the model predicts too little variance at the peak of consumption volatility and from 1975 to 1990. (It is possible to raise the overall variance of consumption in the learning model by increasing the prior weight on the nonstationary model and by increasing γ ; however, I prefer to remain with my calibration to show that one gets good results with a very reasonable calibration. Perhaps one of the reasons the model overestimates the speed at which consumption volatility falls is that actual individuals learn more slowly than a Bayesian learner.) It is Downloaded from http:/www.cambridge.org/core. University of Florida, on 25 Dec 2016 at 16:13:41, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms. http://dx.doi.org/10.1017/S1365100512000417 354 DANIEL L. TORTORICE Credit and Nonstationary Model 1 Data M odel 0.8 0.6 0.4 1950 1960 1970 1980 1990 2000 Credit and Stationary Model 1 Data M odel 0.8 0.6 0.4 1950 1960 1970 1980 1990 2000 FIGURE 5. Ratio of standard deviation of consumption growth to income growth (credit models). important to underscore that the learning model replicates the rise and fall in the variance of consumption without using information on consumption data. It is quite remarkable that the model replicates this distinctive consumption pattern using probabilities calculated only from income data. Another interesting feature of the model is that it replicates the rise in consumption volatility at the end of the sample, coinciding with the most recent recession. Because income has been so far below trend for so long, the model puts extra weight on the nonstationary model, and this increases consumption volatility. It may be puzzling that the learning model line and the stationary model line cross. Recall that the learning model is updating the parameters on the productivity process every period, whereas the stationary model uses the full sample estimates. Hence, when the learning model crosses the stationary model, it is putting most of the weight on the stationary model, and its parameters for the stationary model imply less persistent deviations from trend than in the no-learning stationary model. This explains why the two lines cross. Also, volatility in the nonstationary model is quite high because the full sample estimates imply substantial autocorrelation in productivity growth. Figure 5 examines the ability of the credit constraints model with and without learning to match this fact. The credit constraint model combined with the nonstationary model generates a slight increase in consumption volatility over time. We see this because the volatility of consumption in the nonstationary model is higher than that in the credit constraint model. As consumption becomes slightly less Downloaded from http:/www.cambridge.org/core. University of Florida, on 25 Dec 2016 at 16:13:41, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms. http://dx.doi.org/10.1017/S1365100512000417 CREDIT, LEARNING, CONSUMPTION VOLATILITY 355 1.1 Data M odel 1 0.9 0.8 0.7 0.6 0.5 0.4 1950 1960 1970 1980 1990 2000 FIGURE 6. Ratio of standard deviation of consumption growth to income growth (credit + learning model). constrained, consumption volatility rises. When mixed with the stationary model, the credit constraints model delivers a rise and fall in consumption volatility consistent with the estimated rise and fall in credit constraints. But the magnitude is small because the estimated change in credit constraints is small. Mixing the credit constraints and the learning model (Figure 6) gives results similar to the credit constraints with stationary consumption model. The only difference is that the rise and fall of consumption volatility is slightly higher with the learning model because the learning model generates a larger rise and fall in consumption. Both models, though, are unable to generate the low volatility of consumption at the end of the sample, because estimated credit constraints have not fallen enough. Remember, these are 10-year averages, so the relevant measure of credit constraints is that average over the next 10 years. And although credit constraints were low in 2010, they were quite high on average from 2000–2010. The last figure also highlights an important contribution of the learning model. If loosening of credit constraints is an important contributor to the reduction in consumption volatility, it is important to have a good model of consumption choice absent credit constraints. The credit constraints model is sensitive to the choice of income process. If the income process is stationary, the credit constraints model gives the correct qualitative result. However, if the income process is nonstationary, the reverse is true—consumption becomes more volatile as credit constraints loosen. One can view the learning model as completing the credit constraints model. It allows one not to impose an income process in the face of uncertainty as to which process is correct, and to bring the model a bit closer to the data. Downloaded from http:/www.cambridge.org/core. University of Florida, on 25 Dec 2016 at 16:13:41, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms. http://dx.doi.org/10.1017/S1365100512000417 356 DANIEL L. TORTORICE 5.4. Time-Varying Response of Consumption to Income In Section 3 I found that the correlation of consumption and income has fallen over time. In the equation  ln ct = α + mt  ln yt + εt , mt fell over time (mt = m0 + m1 t with m0 = 0.55 and m1 = −0.124). Table 1 reports the ability of the different models to match this fact. We first see that the benchmark models without time-varying credit constraints or learning (those with a known stationary or nonstationary productivity process) cannot match this fact (m1 is not significantly different from zero). Both the learning model and the credit constraints model improve over the benchmark models; these models predict a fall in m over time. For the learning model m0 = 0.7 and m1 = −0.073, and for the credit constraint model m0 = 0.73 and m1 = −0.05. Adding the time-varying credit constraints to the learning model gives similar results. In this case, m0 = 0.81 and m1 = −0.06. All models predict a smaller fall than is in the data; however, the learning model comes closest to matching the data. 5.5. Breaks in the Variance of Consumption In Section 3 I measured breaks in the absolute deviation of consumption growth from its mean. Table 2 presents the results from this estimate. The nondurables series has one break at 1977:4. The means for each segment are 2.65 and 1.87. The simulated consumption series from the learning model is estimated to have one break occurring in the first quarter of 1982.20 Although a few years from the break estimated in the data, it is within the estimated confidence interval for the break date in the data. The respective means for each segment from the consumption model with learning are 1.5 and 0.94. The model matches both the number of breaks and the relative magnitude of the change. However, the break dates do not line up exactly. None of the consumption measures from the benchmark nolearning models show evidence of any breaks. Consumption data from the model with credit constraints show no breaks. When the credit constraints model is combined with the learning model, there is one break in 1984, quarter 3, and the mean falls from 1.77 to 1.21. 5.6. Relation between Consumption Changes and Probability Changes The last few subsections argued that a learning model with changing probability weights on the nonstationary model is consistent with the observed change in the variance of consumption over time. This subsection extends that argument. In the data, large changes in consumption are associated with large changes in the estimated probability that the nonstationary model is true. I find that this fact is replicated by the learning model. To motivate this section’s analysis, imagine we want to investigate the hypothesis that changes in beliefs influence consumption choice. Then to test this theory, we might want to identify the points in time that resulted in the largest changes Downloaded from http:/www.cambridge.org/core. University of Florida, on 25 Dec 2016 at 16:13:41, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms. http://dx.doi.org/10.1017/S1365100512000417 CREDIT, LEARNING, CONSUMPTION VOLATILITY 357 in beliefs and see if consumption changed more at those times than would be predicted based on changes in income. Because my model identifies how much beliefs should change at each point in time, I can use my estimated beliefs to examine if changes in beliefs help explain consumption changes, controlling for other factors that influence consumption. To this end, I estimate the following regression on the consumption data:  ns      cs − ct−1   + γ Xt + εt , (39)  ln ct −  ln c = α + β ps,t − ps  ∗  t−1   ct−1 where the bar notation denotes the mean, ct is the NIPA measure of nondurables and service consumption, ps,t is the estimated probability that the stationary model is true (note that this is a function only of the income data and not of the consumption data), and Xt is a vector of control variables, including a time variable21 and cs −cns | ln yt −  ln y|, where yt is personal income. | t−1ct−1 t−1 | is the absolute size of the difference in the consumption choice implied by the learning model if the agent puts probability one on the stationary model versus putting probability zero on the stationary model, as a percentage of nondurables and service consumption. The model implies that consumption changes are proportional to this term (see Section 6.2 for more details). The results in Table 3 indicate that the absolute value of the probability change is a significant predictor of the absolute value of the consumption deviation from its mean. The coefficient β = 0.027. In other words, large changes in the estimated probability are associated with large changes in consumption controlling for changes in income and time trends. In estimation I multiply both |ps,t − ps | cs −cns and | t−1ct−1 t−1 | by 100. Hence, this coefficient means that if consumption under the stationary model differs by 10% and the probability changes by 0.01, the deviation in consumption will increase by 0.27 percentage points, compared to a mean of 1.4 percentage points. Table 3 reports the ability of the different models to replicate this fact. The benchmark models without learning or time-varying credit constraints are unable to replicate this fact. The model with learning naturally replicates this fact with β = 0.067. However, the coefficient is too high. On the other hand, the model with credit constraints predicts that β should be negative. Putting the credit constraints model together with the learning model gets a coefficient that is close to correct, β = 0.035. 6. MECHANISM OF THE LEARNING MODEL To better explain the internal mechanisms of the learning model, I discuss two implications of the model in more detail. First, I examine the largest movements in the probability that the stationary model is true and show that they roughly accord with intuition. When log productivity tends to be persistently above its steady state value (0), we move toward the nonstationary model, and when productivity Downloaded from http:/www.cambridge.org/core. University of Florida, on 25 Dec 2016 at 16:13:41, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms. http://dx.doi.org/10.1017/S1365100512000417 358 DANIEL L. TORTORICE TABLE 3. Effect of belief changes on consumption growth Benchmark models Data Nonstationary Stationary Learning Credit |(pst )−E[(ps)]|∗ 0.027∗ |(cst−1 −cnst−1 )/ct−1 | [0.015] 0.24∗∗∗ |ln(yt )−E[ln(y)]| [0.06] t/100 −0.09 [0.14] Constant 0.96∗∗∗ [0.21] Obs. 221 0.004 [0.011] 1.01∗∗∗ [0.04] 0.02 [0.07] 0.15 [0.11] 221 Credit + learning −0.004 0.067∗∗∗ −0.019∗∗ 0.035∗∗∗ [0.006] [0.007] [0.009] [0.006] 0.46∗∗∗ 0.50∗∗∗ 0.65∗∗∗ 0.66∗∗∗ [0.02] [0.03] [0.03] [0.03] −0.02 −0.18∗∗∗ −0.14∗∗∗ −0.16∗∗∗ [0.04] [0.04] [0.04] [0.04] 0.14∗∗ 0.29∗∗∗ 0.33∗∗∗ 0.31∗∗∗ [0.06] [0.08] [0.09] [0.09] 221 221 221 221 Notes: Column (1) reports the results of regressing the absolute value of annualized consumption growth minus its mean on the estimated change in the probability that the stationary model is true times the difference in consumption predicted by the stationary and nonstationary models, controlling for the changes in income (y) and time trends. Columns (2)–(6) redo the analysis of column (1) using simulated data from the nonstationary, no-learning model, from the stationary, no-learning model, from the learning model, from the credit constraint model, and from the credit constraint with learning model. Newey–West standard errors with five lags in brackets. ∗ Significant at 10%; ∗∗ Significant at 5%; ∗∗∗ Significant at 1%. returns toward its steady state value, we move toward the stationary model. Second, I decompose the consumption variance in the learning model into variance due directly to shocks from productivity and that due to changes in the estimated probability weights on each productivity model. 6.1. Largest Movements in the Productivity Process Probability Recall, from (27), that the dynamics of the productivity process probability are given by NS wNS,t  εt+1 wNS,t+1 S = , (40) wS,t+1  εt+1 wS,t ti i p θi , σi2 |Z i i , = p yt+1 (41) |xt+1 , θi , σi2  εt+1 i t+1 p θi , σi2 |Z pS,t = 1 1+ wNS,t wS,t . (42) NS S )/(εt+1 ) is lowest, and the five dates when Table 4 lists the five dates when (εt+1 it is highest (and the data from the previous quarter), in order to build intuition for the mechanisms of the model. These are the dates that provide the best evidence for the stationary model and the nonstationary model, respectively. Additionally, the table lists the percentage deviation of productivity from its steady state value Downloaded from http:/www.cambridge.org/core. University of Florida, on 25 Dec 2016 at 16:13:41, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms. http://dx.doi.org/10.1017/S1365100512000417 CREDIT, LEARNING, CONSUMPTION VOLATILITY 359 TABLE 4. Largest movements in the nonstationary probability Year Quarter PNS ( NS )/( S ) Productivity NS Forecast S Forecast 2001 2001 1 2 Movements toward the stationary model 0.07 1.10 5.44 5.20 0.05 0.66 3.50 5.42 4.82 5.02 1974 1974 3 4 0.50 0.41 0.85 0.70 3.96 1.51 5.01 3.90 4.60 3.49 1949 1950 4 1 0.76 0.70 1.04 0.70 −8.83 0.74 −7.83 −8.82 −7.87 −7.76 2001 2001 2 3 0.05 0.03 0.66 0.72 3.50 1.72 5.42 3.42 5.02 3.08 1959 1959 1 2 0.55 0.47 0.89 0.73 −6.84 −4.88 −7.91 −6.83 −7.03 −5.89 2008 2009 4 1 Movements toward the nonstationary model 0.08 1.56 −9.02 −7.08 0.19 2.74 −13.45 −9.21 −6.80 −8.96 2008 2008 2 3 0.03 0.05 1.09 2.05 −3.83 −6.95 −3.35 −3.83 −3.10 −3.58 2009 2009 2 3 0.18 0.28 0.93 1.72 −13.28 −15.14 −13.73 −13.45 −13.37 −13.06 2004 2005 4 1 0.02 0.03 0.81 1.59 −2.81 −4.95 −4.18 −2.75 −3.92 −2.51 2008 2008 3 4 0.05 0.08 2.05 1.56 −6.95 −9.02 −3.83 −7.08 −3.58 −6.80 Note: This table lists the largest movements in the nonstationary probability along with the percentage deviation of productivity from its steady state and its predictions. (1) and the nonstationary and stationary forecasts of this productivity deviation based on information from the previous period. At each date the stationary model predicts that productivity from the previous period should move toward zero in the next period. The nonstationary forecast always predicts that productivity should be farther from zero than the stationary forecast predicts.22 The periods, then, that give the best evidence for the stationary model are periods when productivity is far from zero and moves substantially toward zero. In contrast, the periods that give the best evidence for the nonstationary model are the periods where productivity is far from zero and moves even further away in the next period. For example, in 2001 quarter 2, productivity moves from 5.4% above trend to 3.5% above trend. This quick reversion to the mean generates evidence for the stationary model. Similarly, during the latest recession, in quarter one of 2009 Downloaded from http:/www.cambridge.org/core. University of Florida, on 25 Dec 2016 at 16:13:41, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms. http://dx.doi.org/10.1017/S1365100512000417 360 DANIEL L. TORTORICE productivity falls from 9% below trend to 13% below trend, providing much evidence for the nonstationary model. This section helps illustrate that although this model may seem to be based on a difficult econometric problem, too far detached from the decision making of individuals, it can capture important features of actual consumption decisions. First, imagine a situation such as the 1990s, in which the economy begins to grow faster. Two views of the world may emerge. The first, the new era view, assumes that the gains are permanent. A second, more pessimistic view would be that the gains are temporary and the economy will eventually return to trend. If uncertainty about these two views is important, when the recession of 2000 comes along and gives evidence against the new era view, one can imagine an additional effect on consumption from changes in beliefs, over and above the direct effect of the productivity shock. I find support for this interpretation in Tortorice (2012). This paper finds that past changes in unemployment affect the extent to which individuals expect unemployment to mean revert. Second, note that the United States has done comparatively well over the last 50 years, returning often to a stable trend. One could imagine that, after seeing this repeated pattern, agents will become more confident in the U.S. economy, and react less to recessions, seeing them as temporary deviations from trend. 6.2. Decomposition of Consumption Variance Figure 7 decomposes the variance of consumption growth relative to income growth from the learning model into variance due directly to income shocks and variance that comes from changes in the probability weight on each model. I confirm numerically for the learning model that ctl = ps,t cts + (1 − ps,t )ctns , (43) where ctl is the consumption choice of the agent at time t from the learning model, cts is the optimal consumption choice if ps,t = 1, and ctns is the optimal consumption choice if ps,t = 0. This equation then implies that s ns ctl = ps,t cts + (1 − ps,t )ctns + (ps,t − ps,t−1 ) ct−1 . − ct−1 (44) Therefore we can decompose the change in consumption into changes due directly to income shocks and changes due to changes in beliefs: ct = ct1 + ct2 , ct1 = ct2 = (ps,t (45) + (1 − s ns . − ps,t−1 ) ct−1 − ct−1 ps,t cts ps,t )ctns , (46) (47) Downloaded from http:/www.cambridge.org/core. University of Florida, on 25 Dec 2016 at 16:13:41, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms. http://dx.doi.org/10.1017/S1365100512000417 CREDIT, LEARNING, CONSUMPTION VOLATILITY 361 0.7 0.6 0.5 0.4 0.3 Model Variance Variance from Shock Variance from Belief Change 0.2 0.1 0 1950 1960 1970 1980 1990 2000 FIGURE 7. Variance decomposition—plot of the learning model’s prediction of cov(ct1 , ln ct ) var(ln yt ) cov(ct2 , ln ct ) and its decomposition into and var(ln . Here ct1 = yt ) cnst−1 s pst ) × ct−1 , where cst is the consumption choice if pt = 1 and cnst t−1 ) . choice if pst = 0. ct2 = (pst − pst−1 ) × (cst−1c−cns t−1 pst × cst ct−1 var( ln ct ) var( ln yt ) + (1 − is the consumption A standard variance decomposition gives  ct Var ct−1   ct1 ct = Cov , ct−1 ct−1   ct2 ct + Cov , ct−1 ct−1  . (48) The first term, ct1 , represents the consumption response to a productivity shock today. You are hit with a shock today, you put some weight on the shock being permanent and some weight on it being temporary, and you adjust your consumption accordingly. But because you revised your beliefs today, you also believe that you made a mistake in your consumption choice last period. You put too much weight on one of the models. And therefore you adjust your consumption even more today, based on how much your revised beliefs indicate that last period’s consumption choice was in error. This effect is captured by the ct2 term. Figure 7 contains the results from this variance decomposition, normalized by t the variance of the income growth rate. Note that I report the variance of cc t−1 23 predicted by the model, divided by the variance of income. There is a rise and fall in the second covariance term early in the sample (marked x). This drives about 15% of the rise in consumption volatility early in the sample. However, by 1975, as the volatility of the probability changes has fallen substantially, most of the variance of consumption is given directly by the response to changes Downloaded from http:/www.cambridge.org/core. University of Florida, on 25 Dec 2016 at 16:13:41, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms. http://dx.doi.org/10.1017/S1365100512000417 362 DANIEL L. TORTORICE Lag 2 ps = .25 Lag 4 ps = .05 1 1 0.8 0.8 0.6 0.6 0.4 1960 1980 2000 0.4 Lag 4 ps = .25 1 0.8 0.8 0.6 0.6 1960 1980 2000 0.4 Lag 6 ps = .25 1 0.8 0.8 0.6 0.6 1960 1980 2000 1960 1980 2000 Lag 4 ps = .45 1 0.4 1980 Lag 4 ps = .25 1 0.4 1960 2000 0.4 1960 1980 2000 FIGURE 8. Robustness to different choices of lag length and prior: model (x); data (–). in productivity, and the continued fall in the variance of consumption is due to the increased weight put on the stationary model. Finally, during the most recent recession, the weight on the nonstationary model rises dramatically and the second covariance term again contributes about 15% of the observed volatility of consumption. 7. ROBUSTNESS Figure 8 explores the sensitivity of the results to different choices for the prior probability. I plot the consumption standard deviation ratio from Figure 4 for the learning model with an initial prior on the stationary model of 0.25 and for a lag choice of 2, 4, and 6. Then I plot the same figure for priors on the stationary model of 0.05, 0.25, and 0.45 with a lag length of 4. As seen in column (1) of Figure 8, the initial rise in consumption variance relative to income variance and the subsequent fall are present for all lag lengths. Moreover, the fit gets better as the lag length increases; the fit is good for lags 4 and 6. The fit deteriorates slightly for a lag length of 2. The same rise and fall pattern is evident, but the magnitude of the fall is muted. Similarly, the initial rise and fall in consumption variance relative to income variance is visible for all priors. Downloaded from http:/www.cambridge.org/core. University of Florida, on 25 Dec 2016 at 16:13:41, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms. http://dx.doi.org/10.1017/S1365100512000417 CREDIT, LEARNING, CONSUMPTION VOLATILITY 363 Decreasing the prior increases the mean level of the consumption variance and increasing the prior decreases the mean variance, but the basic dynamic pattern of a rise and fall in consumption volatility is not changed. 8. CONCLUSION This paper studies three empirical observations. First, after an increase early in the sample, the standard deviation of consumption growth relative to income growth fell by 50%. Similarly, the correlation between consumption growth and personal income growth has also fallen by about 50%. Finally, the consumption series has one estimated break in the absolute deviation of consumption from its mean. The change in mean deviation is about 30%. I examine two explanations of these facts. The first is that a fall in the fraction of credit-constrained consumers has lead to a reduction in consumption volatility relative to income volatility. Although this model captured some of the fall in consumption volatility, it was unable to match the full magnitude of the decline in consumption volatility. It also left open the question of how an unconstrained consumer would respond to an income shock, because if productivity shocks are permanent, relaxing credit constraints could make consumption more volatile. To address these shortcomings, I studied a model with an agent who learns whether income shocks are permanent or transitory. This model was consistent with the overall shape of the relative volatility of consumption to income, the declining correlation of income and consumption over time, and the estimated break in the consumption growth series. The learning model in the paper has many applications. Perhaps one promising route is to use learning about the specification of an exogenous process to add to the literature that uses learning to explain large swings in asset prices [e.g., Bullard (2010)]. NOTES 1. Section 6.1 provides more motivation as to why consumers share this econometric uncertainty. 2. As usual in the literature, consumption analysis is carried out using only data from 1955 onward, avoiding the Korean War. Also, as Campbell and Mankiw (1990) point out, the National Service Life Insurance (NSLI) benefits paid to WWII veterans in 1950:1 distort tests of the permanent income hypothesis. In theory 1950:1 could affect learning about the productivity process. To investigate this possibility, I resimulated the model removing personal current transfer receipts, the component of income including the insurance payment. Results were very similar. Because the 1950:1 observation does not influence the results, I use all the income data to estimate the processes. Because beliefs about the productivity process are central to this paper, I tie them to the data as much as possible. 3. I thank Kurt Kunze, an economist at the Bureau of Economic Analysis, for this insight. 4. Standard errors are calculated using the GMM delta method formulas. See, for example, Cochrane (2001,  pp. 207). In short, let μ = [E( ln ct )2 E( ln ct ) E( ln yt )2 E( ln yt )] and φ(μ) = E( ln ct )2 − [E( ln ct )]2 / E( ln yt )2 − [E( ln yt )]2 . Then var(φ) = j =∞ 1  [dφ/dμ] xt−j )[dφ/dμ], where xt = [( ln ct )2 ( ln ct )( ln yt )2 ( ln yt )]. I j =−∞ cov(xt , T estimate the covariance matrix using a Newey–West estimator with 5 lags. Downloaded from http:/www.cambridge.org/core. University of Florida, on 25 Dec 2016 at 16:13:41, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms. http://dx.doi.org/10.1017/S1365100512000417 364 DANIEL L. TORTORICE 5. Another approach, estimating mt with the Kalman filter, results in a similar conclusion. mt declines approximately linearly from 0.44 to 0.21. I prefer this deterministic approach because I need to make less restrictive assumptions about the nature of the error terms, and statistical inference concerning mt is straightforward. 6. I chose not to include lag regressors in the regression equation that follows, so that the interpretation of the residual is clearer. I am looking to see if deviations in consumption about its mean have changed in magnitude over time. 7. Estimated standard errors and confidence intervals allow for heterogenity and autocorrelation in the residuals, using automatic bandwidth with AR(1) approximation and a quadratic kernel [Andrews (1991)]. The residuals are AR(1) prewhitened and the variance–covariance matrix is allowed to vary across segments. See Bai and Perron (2003) for details. 8. I use the Matlab code provided by Adda and Cooper (2003) to implement this method. i=t 9. To calculate mpct , I simulate a series of ηi=0 and an alternative series that is the same except p  . Letting c = c |η   that ηt−1 = ηt−1 t t t−1 , ct = ct |ηt−1 , yt = Et [yt |ηt−1 ], and yt p  ], the mpc = Et [yt |ηt−1 t ct −ct p p yt −yt = . 10. The ex post real interest rate is measured as the three-month T-Bill rate minus the log change in the GDP deflator. 11. The consumption change is measured in logs. 12. Modeling productivity as a deviation from an HP-filter trend would render productivity clearly stationary, and the agent would learn that productivity is stationary very quickly. 13. Although the productivity process is nonstationary with some probability, the stability conditions for the solution method still hold. See Section 4.2.3 for details. 14. The solution method requires that the eigenvalues of A − Bf , where the optimal policy is given by u = −f x (where x is the state vector), are less than √1β in modulus. The computer program checks this condition and finds it to hold. The matrices A and B are defined hereafter. 15. I take the approach of Kreps (1998) and Sargent (2001) of assuming that the agent treats its probability weights and productivity process estimates as constants when calculating the optimal policy. This “anticipated utility” approach is the most common in the learning literature and necessary to have a tractable solution to the problem. 16. I use the personal income series instead of GDP because the empirical facts are based on the personal income series. 1 17. zt = 1−α [ln((1 + yt )y ss ) − α ln kt − (1 − α) ln(l ss )]. Recall that kt is chosen at time t − 1. To initialize the model I need lag values of kt . I assume that kt = kss for t = 1, . . . , lag. 18. I have confirmed that the results for the learning model hold with more conventional measures of productivity (e.g., TFP); however, the volatility of the income series in the model differs enough from the income series in the data that when I combine the learning model with the credit constraints model, it is unclear what income series to use to calculate, for example, fact 1, the standard deviation of consumption growth to income growth. I would get different results using income from the learning model or using income from the credit constraints model that matches income in the data. Similar issues would arise in analyzing the other facts as well. The current setup removes this ambiguity, ensuring that the learning model generates income exactly equaling income in the data. 19. I have taken care to make sure the results do not depend on the prior choice. Increasing P0NS to 1.5 × I did not affect the results, nor did reducing both P0S and P0NS to 0.015 × I , as long as I did not update the A matrix for the first 10 periods to avoid unstable estimates of the stationary model. Also, increasing θ ns = [0.05 0], lowering θ s = [0.95 0], and lowering P0NS to 0.015 × I caused no noticeable change in the results except a larger increase in pns in the last few years. The results were also insensitive to the choice of σ02 . 20. One can think of these breaks as “measured breaks” because the simulated data do not have discreet breaks in consumption variance. 21. Higher-order time variables were not significant. Downloaded from http:/www.cambridge.org/core. University of Florida, on 25 Dec 2016 at 16:13:41, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms. http://dx.doi.org/10.1017/S1365100512000417 CREDIT, LEARNING, CONSUMPTION VOLATILITY 365 22. Whether productivity is predicted to increase or decrease in the nonstationary model depends on the lag values of zt and varies by observation. 23. This measure is not directly comparable to Figure 4, which plots the ratio of standard deviations; however, the standard deviation, being nonlinear, does not decompose easily. REFERENCES Adda, Jerome and Russell W. Cooper (2003) Dynamic Economics: Quantitative Methods and Applications. Cambridge, MA: MIT Press. Aguiar, Mark and Gita Gopinath (2007) Emerging market business cycles: The cycle is the trend. Journal of Political Economy 115, 69–102. Andrews, Donald W. K. 1991. Heteroskedasticity and autocorrelation consistent covariance matrix estimation. Econometrica 59(3), 817–858. Bai, Jushan and Pierre Perron (1998) Estimating and testing linear models with multiple structural changes. Econometrica 66(1), 47–78. Bai, Jushan and Pierre Perron (2003) Computation and analysis of multiple structural change models. Journal of Applied Econometrics 18(1), 1–22. Benito, Andrew and Haroon Mumtaz (2009) Consumption excess sensitivity, liquidity constraints, and the collateral role of housing. Macroeconomic Dynamics 13(3), 305–326. Bullard, James, George W. Evans, and Seppo Honkapohja (2010) A model of near-rational exuberance. Macroeconomic Dynamics 14(2), 166–188. Campbell, John Y. and N. Gregory Mankiw (1990) Permanent income, current income, and consumption. Journal of Business and Economic Statistics 8(3), 265–279. Cecchetti, Stephen G., Alfonso Flores-Lagunes, and Stefan Krause (2005) Assessing the sources of changes in the volatility of real growth. In Christopher Kent and David Norman (eds.), The Changing Nature of the Business Cycle, pp. 115–138. Sydney: RBA Annual Conference Volume, Reserve Bank of Australia. Cecchetti, Stephen G., Alfonso Flores-Lagunes, and Stefan Krause (2006) Financial Development, Consumption Smoothing, and the Reduced Volatility of Real Growth. AEA Conference Papers 2007. Cochrane, John (2001) Asset Pricing. Princeton, NJ: Princeton University Press. Cochrane, John H. (1988) How big is the random walk in GNP? Journal of Political Economy 96(5), 893–920. Cogley, Timothy and Thomas J. Sargent (2005) The conquest of US inflation: Learning and robustness to model uncertainty. Review of Economic Dynamics 8(2), 528–563. Deaton, Angus (1992) Understanding Consumption. Oxford, UK: Oxford University Press. Dynan, Karen E., Wendy Edelberg, and Michael Palumbo (2009) The effects of population againg on the relationship among aggregate consumption, saving and income. American Economic Review: Papers and Procedings 99(2), 380–386. Dynan, Karen E., Douglas W. Elmendorf, and Daniel E. Sichel (2006) Can financial innovation help to explain the reduced volatility of economic activity? Journal of Monetary Economics, 53(1), 123–150. Evans, George and Seppo Honkapohja (2001) Learning and Expectations in Macroeconomics. Princeton, NJ: Princeton University Press. Georges, Christophre and John C. Wallace (2009) Learning dynamics and nonlinear misspecification in an artificial financial market. Macroeconomic Dynamics 13(5), 625–655. Guvenen, Fatih (2007) Learning your earning: Are labor income shocks really very persistent? American Economic Review 97(3), 687–712. Hamilton, James (1994) Time Series Analysis. Princeton, NJ: Princeton University Press. Kreps, David (1998) Anticipated utility and dynamic choice, a 1997 Schwartz Lecture. In Donald Jacobs, Ehud Kalai, and Morton Kamien (eds.), Frontiers of Research in Economic Theory, pp. 242–274. Cambridge, UK: Cambridge University Press. Downloaded from http:/www.cambridge.org/core. University of Florida, on 25 Dec 2016 at 16:13:41, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms. http://dx.doi.org/10.1017/S1365100512000417 366 DANIEL L. TORTORICE Ljungovist, Lars and Thomas Sargent (2004) Recursive Macroeconomic Theory. Cambridge, MA: MIT Press. Ludvigson, Sydney (1999) Consumpiton and credit: A model of time-varying liquidity constraints. Review of Economics and Statistics 81(3), 434–447. Nakajima, Makoto (2007) Solving RBC Models with L–Q Approximation. Mimeo, UIUC. Sargent, Thomas (2001) The Conquest of American Inflation. Princeton, NJ: Princeton University Press. Stock, James H. (1991) Confidence intervals for the largest autoregressive root in U.S. macroeconomic time series. Journal of Monetary Economics 28(3), 435–459. Stock, James H. and Mark Watson (2003) Has the business cycle changed and why? In Mark Gertler and Kenneth Rogoff (eds.), NBER Macroeconomics Annual 2002, pp. 159–230. Cambridge, MA: National Bureau of Economic Research. Tauchen, George and Robert Hussey (1991) Quadrature-based methods for obtaining approximate solutions to nonlinear asset pricing models. Econometrica 59(2), 371–396. Tillman, Peter (2011) Parameter uncertainty and non-linear monetary policy rules. Macroeconomic Dynamics 15(2), 184–200. Tortorice, Daniel L. (2012) Unemployment expectations and the business cycle. B.E. Journal of Macroeconomics (Topics) 12(1), 1–47. Waters, George (2009) Learning, commitment and monetary policy. Macroeconomic Dynamics 13(4), 421–449. APPENDIX: LEARNING MODEL A.1. LQ APPROXIMATION I approximate U (w) = {[k α (ez l)1−α −i][1−(θl)φ ]}1−γ 1−γ with w = [z k l i] about the steady state  w = [z k l i] . These calculations are based on Nakajima (2007) and Ljungovist and Sargent (2004). We can take a second-order Taylor approximation as 1 U (w) ≈ U (w) + (w − w) J + (w − w) H (w − w) , 2 where J is the Jacobian matrix and H is the Hessian matrix, both evaluated at the steady state. This expression can be manipulated into this form: ⎡ U (w) − w J + 12 w H w U (w) ≈ [1w ] ⎣ 1 (J − H w) 2 1 (J 2 ⎤ − H w)  1  ⎦ . w 1 H 2 Let ⎡ U (w) − w J + 12 w H w M=⎣ 1 (J − H w) 2 1 (J 2 − H w) ⎤ ⎦ 1 H 2 Downloaded from http:/www.cambridge.org/core. University of Florida, on 25 Dec 2016 at 16:13:41, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms. http://dx.doi.org/10.1017/S1365100512000417 CREDIT, LEARNING, CONSUMPTION VOLATILITY 367 and let x = [1 z k ] and u = [l i] . We partition M as  U (w) ≈ [1w ] [M] 1 w  =      x R W x u W Q u = x  Rx + u Qu + 2x  W u. A.2. MODEL MATRICES For the stationary model the state variables are xtS = [1 zt kt zt−1 . . . zt−p+1 ] and the control variables are u = [lt it ] . So we have   R 0 s R = , 0 0   W Ws = , 0 Qs = Q. The analogs for the nonstationary model are exactly the same. For the nonstationary model the state variables are xtNS = [1 zt kt zt zt−1 . . . i = zt−p+2 ] and the control variables are ut = [lt it ] . So for the laws of motion (xt+1 i i i i i At xt + B ut + C εt+1 ), we have (for clarity I have written it for four lags, but it is easy to generalize) ⎡ ⎤ 1 0 0 0 0 0 ⎢ 1 2 3 4 ⎥ ⎢0 θs,t ⎥ 0 θs,t θs,t θs,t ⎢ ⎥ ⎢ ⎥ 1−δ 0 0 0 0 0 ⎢ ⎥ S 1+g At = ⎢ ⎥, ⎢ ⎥ ⎢0 1 ⎥ 0 0 0 0 ⎢ ⎥ ⎣0 0 0 1 0 0⎦ 0 0 0 0 1 0 ANS t ⎡ 1 ⎢ ⎢0 ⎢ ⎢ ⎢0 =⎢ ⎢0 ⎢ ⎢ ⎢ ⎣0 0 0 0 0 0 1 0 1 θns,t 2 θns,t 0 1−δ 1+g 0 0 0 0 1 θns,t 2 θns,t 0 0 0 0 1 0 0 1 B s = B NS ⎡ 0 ⎢0 ⎢ ⎢0 =⎢ ⎢0 ⎢ ⎣0 0 0 ⎤ ⎥ 3 ⎥ θns,t ⎥ ⎥ 0 ⎥ ⎥, 3 ⎥ θns,t ⎥ ⎥ ⎥ 0 ⎦ 0 ⎤ 0 0 ⎥ ⎥ 1 ⎥ 1+g ⎥ , 0 ⎥ ⎥ 0 ⎦ 0 Downloaded from http:/www.cambridge.org/core. University of Florida, on 25 Dec 2016 at 16:13:41, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms. http://dx.doi.org/10.1017/S1365100512000417 368 DANIEL L. TORTORICE ⎡ ⎤ ⎡ ⎤ 0 0 ⎢1⎥ ⎢1⎥ ⎢ ⎥ ⎢ ⎥ ⎢ 0 ⎥ NS ⎢ 0 ⎥ ⎥ ⎢ ⎥ Cs = ⎢ ⎢0⎥ C = ⎢1⎥. ⎢ ⎥ ⎢ ⎥ ⎣0⎦ ⎣0⎦ 0 0 Note that the optimal policy is independent of C. Because the objective is linear–quadratic, the optimal policy exhibits certainty equivalence. Downloaded from http:/www.cambridge.org/core. University of Florida, on 25 Dec 2016 at 16:13:41, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms. http://dx.doi.org/10.1017/S1365100512000417