The full text on this page is automatically extracted from the file linked above and may contain errors and inconsistencies.
orKing raper series C o n s u m p t i o n - B a s e d Modeling of Long-Horizon Returns K e n t D . D a n ie l a n d D a v id A . M a r s h a ll 3 ■) W o r k in g P a p e r s S e r ie s R e s e a r c h D e p a r tm e n t F e d e r a l R e s e r v e B a n k o f C h ic a g o D e c e m b e r 1 9 9 8 (W P -9 8 -1 8 ) FEDERAL RESERVE B A N K OF CHICAGO December 24, 1998 Consumption-Based M odeling of Long-Horizon Returns* K e n t D . D a n ie l* a n d D a v id A . M arshall** Abstract Numerous studies have documented the f i u e of consumption-based pricing mod alr e s to explain observed patterns in stock and bond returns. This f i u e has some l alr times been attributed to f i t o s transaction costs or d r b l t . I such f i t o s rcin, uaiiy f rcin axe important, they should primarily a f c the higher frequency components of fet asset returns. The long-swings, or lower-frequency comovements should be l s a es f fe t d Consequently i transaction costs are important, t s s of the consumption ce. f et based asset pricing model which concentrate on lower-frequency components may be more s c e s u . ucsfl We investigate t i hypothesis using a variety of diagnostic t s s We f r t use hs et. is coherence analysis and bandpass f l e i g analysis to show t itrn hat, while there i s a complete lack of correlation between asset returns and consumption growth at frequencies higher than about 0.7 years-1 (swings longer than 1 4 ye r ) the coher. as, ence/correlation between the two s r e at lower frequencies i above 60%. We per eis s form Hansen and Jagannathan (1991) bounds t s s x 2 t s sofmoment r s r c i n , et, et etitos and Hansen and Jagannathan (1997) specification t s softhree consumption-based et models of the asset-pricing k r e : Time-separable preferences with power u i i y enl tlt, the Abel (1990) “ Catching up with the Joneses” preferences, and Constantinides (1990) habit-formation preferences. While none ofthese models perform well at the quarterly horizon, the performance ofthe Abel and Constantinides models improves strikingly at the two-year horizon. tW e thank Steve C ecchetti, Larry C hristiano, John C ochrane, G eorge C onstantinides, Larry E pstein, Lars H ansen W alter Torous and participants o f the N BER A sset P ricing Group M eeting and UCLA F inance Sem inar for helpful discussions, and we acknow ledge D enise Duffy and G lenn M cAfee for superlative research assistance. D aniel gratefully acknowledges research support from the C enter for R esearch in Security Prices (C R SP) at th e U niversity o f C hicago. T he opinions expressed in th is paper are th ose o f th e authors, and do not reflect the view s o f the Federal R eserve Bank o f C hicago or the Federal R eserve System . ’ K ellogg G raduate School o f M anagem ent, N orthw estern U niversity, E vanston, IL 60208, 847-491-4113, kent@ kent.kellogg.nw u.edu, h ttp ://k en t.k ellogg.n w u .ed u / ’ ’ Federal R eserve Bank of C hicago, 312-322-5102, dm arshall@ frbchi.org 1 Introduction Most research on consumption-based a sset pricing focuses on short-horizon returns. The start ing point i the familiar intertemporal Euler equation s 1= ( 1) E t [m T+Tr l +T] t where rJ+T denotes the gross real cumulative asset return from date t to date t + r and m J+T denotes the intertemporal marginal rate of substitution (IMRS) between wealth at date t and wealth at date t + r The unconditional version of th s Euler equation i . i s 1= ( 2) E [mJ+TrJ+T] . In the literature on consumption-based asset p i i g m j i modeled as a function of aggregate rcn, s consumption, and the implications of ( or ( are tested f r return horizons r equal to one 1) 2) o month or one quarter. The l t r t r typically r j c s the conditional Euler equation ( ) ieaue eet 1. Presumably, the unconditional Euler equation ( should be easier to ftto the data. Equa 2) i tion ( only requires that the IMRS correctly price equity returns ”on average”. Yet, Cochrane 2) and Hansen (1992) show how d f i u t i i to even ft t i unconditional Euler equation. The ifcl t s i hs problem i that, to ft equation ( , one needs substantial covariation between m J+T and r[+ r . s i 2) Formally, equation ( ) implies 2 E [rj+ r ~ rf t ] E = -cm (m[+T,rt = T +T) (3) - o m a r p m>r where r f ( denotes the gross re l r s - r e rate from date t to date t + r a m and a ikfe , ar denote the standard deviations ofm l +T and rJ+T r s e t v l , the correlation c e f c e t between m[+T epciey ofiin and rJ+T i denoted pm>r,and we use the implication of equation ( ) that £[mj+T]= s 1 E [ l/ r f{ \ . The left-hand side of equation ( ) i the mean equity premium, ’discounted” by the mean 3 s ’ inverse r s - r e r t . To generate a large mean equity premium, a model of m j+ r must display ikfe ae substantial negative covariation with equity returns. Within the context ofconsumption-based asset pricing, th s would require substantial positive covariation between equity returns and i consumption growth. Empirically, the contemporaneous correlation between quarterly returns and quarterly con sumption growth i small ( e s than 0 1 ) and the la g s correlation at any lead/lag (when s ls .5, ret 1 returns lead consumption growth by two quarters) i l s than O.2.1 Cochrane and Hansen s es (1992) c l t i low correlation between the return on market proxies and consumption growth al hs the “ correlation puzzle.”2 A number of factors have been proposed to account for the low correlations between stock returns and aggregate consumption growth at short-horizons, in cluding uninsurable cross-sectional heterogeneity,3 fixed costs ofadjusting consumption,4 costs of portfolio adjustment,5 and even small deviations from perfect rationality.6 While these fac tors could substantially a f c the co-movements of asset returns and aggregate consumption fet at high frequencies, they should be l s disruptive to the theory at longer horizons. Simple es correlations between consumption growth and the V W index return suggest that there may be merit in this argument. While the contemporaneous correlation between consumption growth and returns at the one-year horizon i about the same as f the one-quarter horizon, the s or correlation between one-year consumption growth and one-year returns lagged by two quarters i 0 3 , almost twice as high as the maximal correlation found between quarterly returns and s .5 quarterly consumption growth I consumption-based pricing operators perform better at pricing long-horizon a s t , t i f ses hs would provide indirect evidence that the basic intuitionunderlying the equilibrium asset pricing theories of Lucas (1978), Breeden (1979), and Grossman and S i l r (1982) i sound; the wel hle s l documented f i u e of this theory at matching high-frequency data could be attributed to alrs transaction c s s market imperfections, and uninsurable heterogeneity. This analysis would ot, not be a substitute f r formal modeling of these f i t o s but would at l a t suggest that o rcin, es t i approach i on the right t a hs s r ck. I, however, consumption-based pricing proves useless f at a l frequencies, th s would represent a substantial challenge to equilibrium pricing theory. l i In a l such models, financial assets are vehicles f r transferring consumption across time and l o 1 T hese num bers were calculated using real non-durable and services consum ption d ata, and returns on the CRSP value-w eighted index, deflated by the ND& S deflator. 2Even if the correlation betw een consum ption grow th and equity returns is sm all, equation (3) suggests an alternative way o f generating a large m ean equity premium: increase th e variability o f the IM RS (for exam ple, by assum ing a high degree of risk aversion). C ecchetti, Lam , and M ark (1993) try th is second strategy. W hile they can m atch the m ean equity prem ium , they have difficulty m atching both the first and second m om ents o f the equity and risk-free return data. 3See, e .g ., C onstantinides and D uffie (1996), Lucas (1994), H eaton and Lucas (1995) 4G rossm an and Laroque (1990), M arshall (1994), M arshall and Parekh (1998). ®Luttmer (1996), He and M odest (1995). 6See C ochrane (1989). 2 random states. It is not clear what sort of model can capture this intuition without implying a theoretical linkage between aggregate consumption and asset returns at some horizon. In this study, we investigate the a i i y of of consumption growth to explain asset returns blt at low frequencies using a variety of diagnostic t s s We f r t use multivariate spectral analysis et. is to characterize to co-movements of consumption growth and excess equity returns. We find that, while the coherence between these variables i i s g i i a t at high frequencies (above 0.7 s ninfcn years-1) at lower frequencies th s coherence i high and s a i t c l y s g i i a t We confirm , i s ttsial infcn. the coherence analysis using bandpass f l e i g analysis similar to that suggested by Baxter itrn and King (1994). We find no s a i t c l y s g i i a t correlation between the two s r e at high ttsial infcn eis frequencies. However, at lower “ business-cycle” frequencies, we find a correlation of over 50%. We then turn to more formal t s s of consumption-based asset pricing at longer horizons. et We look at standard time-separable power u i i y the Abel (1990) “Catching up with the Jone tlt, ses” preferences, and the Constantinides (1990) habit-formation preferences. Our investigation uses three diagnostic t o s f r t we apply the Hansen and Jagannathan (1991) mean-variance o l : is, analysis, modified to take into account the unconditional correlation between the pricing kernel and asset returns; second, we apply the standard x 2 ofthe moment r striction in ( ) The third e 2. diagnostic we use follows Hansen and Jagannathan (1997). They note that some preference specifications may result in an extremely v l t l m t, but one which does not r a l “ i ” the oaie e l y ft data better. However, because of the increased v l t l t , such a preference specification may oaiiy s i lyield a low x 2- Hansen and Jagannathan (1997) and Hansen, Heaton, and Luttmer (1995) tl suggest a specification t s which i immune to t i problem. We employ t i t s as a further et s hs hs et diagnostic t o . ol According to our empirical r s l s none of the consumption-based models performs well at eut, the quarterly horizon. However, when the horizon i lengthened to two years, versions of the s Abel (1990) and Constantinides (1990) models ofpreferences perform f i l w l . In particular, ary el no signi i a t violations of the Hansen-Jagannathan r s r c i n are found, the x 2 s a i t c are fcn etitos ttsis insignificantly d f e e t from z r , and the Hansen and Jagannathan (1997) diagnostic reveals ifrn eo rather small discrepancies from equation ( . 2) Other papers in the l t rature have asked whether the r s r ie e t ictions of economic theory are 3 better s t s i d with longer-horizon returns than with monthly or quarterly returns. Brainard, aife Shapiro, and Shoven (1991) find that stock returns correspond more closely to the theoretical return to invested capital at longer horizons. Cochrane and Hansen (1992), however, find that the performance of time-separable preferences deteriorates as the horizon lengthens. F inally, Daniel and Marshall (1997) use a vector A R C H model to estimate the conditional covariance between equity returns and the pricing kernel implied by various models. All models do poorly at the quarterly horizon. At the two-year horizon, however, they find that the mean and variance of this conditional covariance s r e approximate the corresponding moments of eis the conditional equity premium as long as s f i i n time-nonseparability i incorporated into ufcet s preferences. Furthermore, the variation through time in th s conditional equity premium i to i s some extent reflected in variation in the conditional covariance s r e . eis The conditional evidence i important. Although we find evidence ofa strong unconditional s relationship here, the high unconditional correlations between consumption growth and asset returns at business cycle frequencies could be caused by co-movement of the expected growth and return s r e as opposed to co-movements of the innovations. There i strong evidence eis s of variation in expected returns across the business cycle.7 Thus a t s of the conditional et model might f i when the business cycle variables are used as instruments, even though the al unconditional t s does not r j c the model. The r sults in Daniel and Marshall (1997) suggest et eet e that this should not be the c s . ae The remainder of the paper i organized as f l o s In section 2 we present evidence that s olw: , although consumption-growth i virtually uncorrelated with asset returns at high frequencies, s i displays substantial covariation with asset returns at business-cycle frequencies. In section 3 t we characterize the preference specifications to be explored in the remainder of the paper; we discuss the model diagnostics to be used; and we evaluate particular parameterizations of these models at various horizons. In that s c ion, we also discuss problems in modeling long horizon et returns that were pointed out by Cochrane and Hansen (1992). We argue that these problems can be resolved, in p i c r n iple, by the use of time-nonseparable preferences as advocated by Abel 7For exam ple, Fam a and French (1989) show th a t th e term spread and th e default spread forecast future long horizon returns, and Chen (1991) show s th at th ese sam e variables forecast future econom ic activity. If th ese sam e variables forecast future consum ption grow th, th is could result in an uncondi tion al correlation, even if the conditional correlation is zero. 4 (1990) and Constantinides (1990). In section 4 we provide a more formal analysis of the Abel , and Constantinides models in which the model parameters are chosen to minimize the Hansen and Jagannathan (1997) specification c i e i n Section 5 summarizes. rtro. 2 L o w - F r e q u e n c y Correlation of Asset R e t u r n s a n d C o n s u m p tion G r o w t h In th s section, we ask whether the Cochrane and Hansen (1992) correlation puzzle i a phe i s nomenon at a ll frequencies, or just at high frequencies. We use two techniques to explore t i hs question: multivariate spectral or coherence ana y i , and an analysis of the correlations of lss bandpass f l e e consumption growth and a itrd sset returns. 2.1 Coherence Analysis To assess the frequency-by-frequency breakdown ofthe correlation between consumption growth and a market proxy, we perform coherence analysis of the consumption growth and excess V W index return s r e . Ess n i l y what the coherence analysis does i to s l t each of the two eis etal, s pi s r e up into a s t of Fourier components at d f e e t frequencies, and then to determine eis e ifrn the correlation of a s t of Fourier components f r the two s r e around each frequency. The e o eis method we use f r generating the coherence s r e i described in the appendix. In addition, o eis s the method yields the phase relation between the two s r e , which i a measure of how f r the eis s a s r e must be shifted to maximize the correlation of the s t of Fourier components. eis es In Figure 1, the upper panel gives the coherence between the two s r e , as a function of eis the frequency. Approximate 95% confidence i t r a s f r both coherence and phase are plotted nevl o as dashed l n s in the figure.8 We also plot the 95% confidence interval that the coherence i ie s above z r , which i the dot-dashed l n i the f g r . eo s ie n iue Figure 1 shows that the coherence i low a high frequencies. However, f r lower frequencies, s t o but not the very lowest, the coherences are r l t v l l r e and are considerably more than eaiey ag, two standard errors away from z r . This suggests that any correlation that a i e between eo rss 8T hese are calculated follow ing Bloom field (1976). For the coherence series, th e standard error is inversely proportional to the level of the coherence, and in calculating these confidence intervals we have assum ed th a t th e true coherence is equal to the estim ator. 5 Coherence — Consumption Growth and excess V W returns — Phase — Consumption Growth and excess V W returns — 19 47-1997 1947— 1 997 T his figure show s, in th e upper panel, the coherence betw een quarterly real non-durable and services consum ption grow th and quarterly real returns on th e CRSP VW index over th e period 1947:1-1994:1. T he low er panel show s th e phase relationship, in degrees, betw een th e tw o series a t th e sta ted frequency. A ham m ing w indow and a sm oothing param eter ( n / m ) o f 16 is used. D etails o f th e coherence and phase calculations are given in the A ppendix. Figure 1 Coherence and Phase plots for consumption Growth and excess V W index : Returns the consumption growth and asset return s r e i due to the co-movement at these lower eis s frequencies. The lower panel of Figure 1 displays the phase as a function of frequency. Notice that at a frequency of z r , the phase i z r , and then decreases approximately linearly with frequency eo s eo up to a frequency of about 1 year""1. This suggests a constant-length lead/lag relationship between the two variables, since the approximate lead/lag length equals the phase multiplied by (1 / f r e q u e n c y * 360). Performing t i calculation, Figure 1 t l s us that return s r e leads hs el eis the consumption growth s r e by approximately two quarters. To verify t i , in Figure 2 we eis hs lag the V W return s r e by 2 quarters and rerun the coherence a a ysis. The coherence values eis nl increase sli h l at most frequencies, except a frequency zero where i increases dramatically. gty t t The phase i now approximately zero f r frequencies l s than one year”1. s o es 6 Coherence — Consumption Growth and excess V W returns — Phase — Consumption Growth and excess V W returns — 19 47— 1997 1947— 1 997 T he construction o f th is figure was identical to th at for Figure 1, except th a t the VW index return was lagged by tw o quarters before calculating the coherence. Figure 2 Coherence and Phase plots for Consumption Growth and Lagged excess : V W index Returns In summary, the coherence analysis suggests while there i l t l relation between con s ite sumption growth and asset returns at high frequencies, there i a strong relationship at low s frequencies, but that the relationship i not contemporaneous: consumption growth lags the s market return by about two quarters. This might t e inwell with a f i t o s s o y I there i an i rcin tr: f s extra cost of adjusting consumption quickly rather than s o l , then th s i exactly the sort of lwy i s relationship one might expect to observe. Alternatively, i agents’preferences exhibited habit f formation ( . . Constantinides (1990)) and there were a u i i y cost to rapid consumption eg, tlt adjustment, consumption would respond slowly to change in asset p i e . rcs 7 Figure 3 Bandpass Filter Frequency Response : T his figure show s th e gain o f the three filters used in the bandpass filtering analysis over the frequency range from 0 to 1 years” 1. The gain from 1 to 2 years " 1 is not show n, but is approxim ately 1 for th e high filter and 0 for th e other tw o. 1.4 M Tr©ndMf l e itr -- "Cycle- f l e itr ----High-Pass- fl e itr 1.2 0.8 0.6 - \ / \/ \‘ / / ' /' •/ . 04 .- /* . 0.2 - / \ / \ / * / .... ✓.■’ *** ' •. * \ 0.4 O.S 0.6 0.7 Frequency (1/years) . 0.1 2 .2 0.2 0.3 . .\ 0.8 0.0 Bandpass Filtering Analysis To verify the coherence r s l s we also conduct a time-series bandpass f l e i g analysis, in eut, itrn which we break the consumption growth and asset returns s r e into d f e e t components eis ifrn using a s of moving-average (MA) f l e s To do t i , we u i i e a bandpass f l e i g analysis et itr. hs tlz itrn similar to that suggested by Baxter and King (1994), and u i ized by Baxter (1994). tl The M A f l e swe u i i ehere are each designed to pick out a certain range offrequencies in itr tlz the data. The f l e s are symmetric two-sided f l e s and consequently do not introduce phase itr itr, distortions into the data. Though the design of the f l e s uses spectral analysis techniques, itr the f l e s themselves are simple M A f l e s The process of f l e i g a data s r e i conducted itr itr. itrn eis s by simply convolving the data s r e with the s t of f l e weights, as with any M A f l e . eis e itr itr The f l e design i done using the FIRLS ( i i e impulse response l a t squares) function itr s fnt es in M ATLAB? This function takes as an input a desired piecewise linear spectral response function, and then finds the length n (here 31 quarters) M A f l e which provides a spectral itr response which i as close as possible to the desired spectral shape ( n a weighted least-squares s i 9T his function is taken from M A T L A B s th e Signal P rocessing T oolbox. 8 sn e se). We constructed three f l e s To conform to the business cycle l t r t r , we c l these the itr. ieaue al tre n d , c y c le and h ig h f l e s The itr. tren d f l e i designed to eliminate a l swings in the data itr s l shorter than 8 years ( . . with a frequency of higher than 0.125 years-1.. The te, ) cy cle fle i itr s designed to pass a l components with wavelength between 1 5 and 8 years, and the h ig h f l e l . itr i designed to eliminate a l swings longer than 1 5 years. The spectral gain obtained f the s l . or three f l e s are shown in Figure 3.10 While we would l k these f l e s to have a gain of 1 itr ie itr over the bandpass region and zero elsewhere, t i i not achievable with a f n t length f l e .11 hs s iie itr Instead, th s figure shows that the f l e frequency responses are about as good as i achievable, i itr s given the constraints. The motivation f r eliminating high frequencies from the data has already been discussed. o The motivation f r eliminating the correlations at very low frequencies i that there may be o s long-run changes in the structure of the economy that a f c the co-movement of consumption fet and returns on financial a s t . (Changes in the l g l and institutional framework of the ses ea pension-fund industry i one example.) At a more basic l v l since other macroeconomic s ee, variables move together at business cycle frequencies, and not elsewhere, th s may also be true i f r consumption. o The correlations between various measures of consumption growth and V W return s r e , eis f l e e using each of the three f l e s in turn, are presented in Table 1 We consider s x itrd itr . i alternative measures of consumption: consumer durables, consumer non-durables, consumer s r i e , nondurables plus s r i e , consumer services excluding the implied service flow from evcs evcs owner-occupied housing, and nondurables plus services excluding the service flow from owneroccupied housing. The procedure f r constructing each of these correlations was to f l e each o itr of the two s r e , and then calculate the maximum correlation p max between the two s r e eis eis over a lead-lag range of eight quarters, that i s 10We also demean the set of M A c e f c e t output from the FIRLS procedure f r the cycle and high ofiins o f l e weight s r e . Since these are f n t length f l e s i the sum of the M A c e f c e t i zero than itr eis iie itr, f ofiins s the f l e s w l eliminate any integrated (1(1)) component present in the data. itr i l 11 Baxter and King (1994) discuss the reasons that an “ideal spectral shape” i not achievable in s practice. 9 Table 1 The Correlation of Bandpass Filtered Returns and Consumption Growth : Series T he quarterly real C R SP VW index returns and the real non-durable and services consum ption grow th series (1947:1-1997:4), were each filtered using one o f the three filters, as described in th e te x t. T he m axim um correlation coefficient betw een th e tw o filtered series, pmax, is com puted as in equation (4 ). T he value o f p max as w ell as the m axim izing value o f r are given in the Table for each filter. T he M onteC arlo determ ined p-values are the probability o f obtaining a number at least as large as th e sam ple value assum ing th a t th e returns are i.i.d. norm al, and independent o f consum ption grow th. Six m easures o f consum ption are used: consum er durables (”D ur.”); consum er non-durables (”N D ”); consum er services (”Serv.”); nondurables plus services (”ND& S”); consum er services excluding th e im plied service flow from ow ner-occupied housing (”SxH”); and nondurables plus services excluding th e im plied service flow from ow ner-occupied housing (”ND&SxH”). Dur. Pmax Tmax ND Pmax Tmax Serv. Pmax Tmax ND&S Pmax Tmax SxH Pmax Tmax ND&SxH Pmax Tmax / Pmax 1 Fle itr trend cycle -0.1053 (0.718) 0.5343 (0.0 2 0) 0 2 -0.0521 (0.657) 0.5092 (0.003) 2 5 -0.0148 (0.619) 0.4733 (0.020) 8 2 -0.0884 (0.703) 0.5497 (0.0 2 0) 2 8 -0.1427 (0.747) 0.4248 (0.052) 2 8 -0.1504 0.5360 ( 002) 0. 8 2 \ / 1 high 0.2053 (0.192) -3 0.2735 (0.019) 1 0.1449 (0.634) 4 0.1662 (0.443) 1 0.1344 (0.698) 4 0.1739 ( . 8 03) 1 \ m ra{T,T+r} min{ r+T} T- — max £ T (a(r)a(Ac)) (t - t ) t= m a x {l,l+ r} ^ 8 < r < 8. (4) where r and Ac* are demeaned returns and consumption growth over period t respectively, * , and a ( r ) a (A c) are the corresponding sample standard deviations f r the s r e . o eis We use Monte-Carlo analysis to determine the significance l v l ofthe correlations in Table ees 1 both in order to account f r the f c that we axe reporting the maximum of 17 c e f c e t , o at ofiins and to account f r the s r a correlation induced in the two s r e by the f l e i g operations. o eil eis itrn The p -v a lu e s reported in Table 1 are therefore the probabilities ofobtaining a c e f c e t at l a t ofiin es as large as the one found i the quarterly returns s r e were i.i.d . normal, and independent of f eis consumption growth. 10 As predicted by the coherence a a y i , the maximum correlation between the two s r e nlss eis a t r f l e i g using the cycle f l e i high. For example, the maximized correlation c e f c e t fe itrn itr s ofiin i 54.97% f r ND&S consumption, with a one-tailed p-value of 0. s o 2%. Also, as predicted by the coherence analysis, the correlation i maximized when the returns s r e i lagged by two s eis s quarters. We find l t l correlation between the returns s r e and any of the high-pass f l e e con ite eis itrd sumption growth s r e . The maximized correlation c e f c e t i about 20% f r most measures eis ofiin s o of consumption, and i ins g i i a t y d f e e t from z r . An exception i the non-durable s s infcnl ifrn eo s e r e , which i signi i a t at the 1.9% (one-tailed) value. To t s whether the correlation i is s fcn et s d f e e t at the high and business cycle frequencies, we conduct an additional Monte-Carlo ifrn simulation. Note that the correlation between the business-cycle f l e e s r e i approxi itrd eis s mately 0.5 f r a l measures of consumption. In our simulation e e c s , we impose the null o l xrie hypothesis that the correlation between the high-pass f l e e s r e i also 0 in population. itrd eis s .5 We then compute the probability of obtaining a maximum observed correlation of 0.27 (the highest correlation in the l s column of Table 1 . We find that th s probability i l s than at ) i s es 0.001. We conclude that the correlations between high-pass f l e e s r e are si n f c n l l s itrd eis g i i a t y es than the correlations between business-cycle f l e e s r e . i trd eis I t r s i g y the maximum correlation of the trend components of the two s r e i also neetnl, eis s i s g i i a t at - 8% (one-tailed p-value of 70.3%) f r non-durable and s r i e . A Monteninfcn, 8. o evcs Carlo simulation shows that the probability ofgetting a measured trend-correlation ofl s than es zero i 4.8%, when the true correlation between the two s r e equal to 0 at a l frequencies. s eis .5 l Thus, i appears that the correlation i indeed lower at trend frequencies than a business cycle t s t frequencies. The normalized, cycle -itrd a f l e e sset returns and consumption growth s r e s r plotted e i s ie in Figure 4.12 We have also lagged the V W index return s r e by two quarters. The plot eis shows the very strong relationship between the two s r e in t i frequency range from the 50’ eis hs s through the early 80’. The l t 80’ are the only period in the plot where the relation i not s ae s s extremely strong, perhaps because of the 1987 crash. 12To normalize the s r e , we divide each s r e by i s sample standard deviation. eis eis t 11 Figure 4 Filtered Asset Returns and Consumption Growth Series : This figure plots the business-cycle filtered real consumption growth (Total) and filtered, two-quarterlagged, VW returns series. The filtering method is described in the text. The period is 1947:1-1994:1, but the first and last four years are truncated in the filtering process. Each series is normalized by dividing by its standard deviation. Business Cycle Components 3 Unconditional M o m e n t s Tests of the Asset Pricing Relation Section 2 documents that consumption growth i more highly correlated with equity returns at s business cycle horizons than at short horizons, and indeed that there i in i n f c n correlation s sgiiat at frequencies of l s than 0.67 years-1. This suggests that there may be f i t o s which es rcin “ de-link” consumption and asset price movements at high frequencies. I s , t s s of the f o et consumption based asset pricing model may be considerably more successful at pricing longerhorizon returns. For the remainder ofthe paper, we explore th s conjecture at various horizons. i 3.1 Preference Models 3.1.1 Time-Separable Preferences The most widely-studied (and widely-rejected) preference specification in the consumptionbased pricing literature i time-separable power u i i y In this s e i i a i n agents solve the s tlt. pcfcto, 12 following maximization problem: oo max U {* j J o c+ ) i 1-7 (5) = E t Y B j -^ ~ jjtS Zo A _T 1 ' subject to the usual budget constraint. The r-period IMRS i: s ( 6) An important treatment of long-horizon returns with time-separable preferences can be found in Cochrane and Hansen (1992). They use the method of Hansen and Jagannathan (1991) to look at the implications of time-separable u i i y at horizons ranging from one quarter to f v tlt ie years. They find that the performance of time-separable u i i y actually tlt d e te rio ra te s as the horizon lengthens. This f i u e of time-separable preferences at long horizons i largely caused alr s by a high implied r s - r e r t . Aggregate consumption i a ( ikfe ae s stochastically) growing s r e . In eis the time-separable model, agents seek to transfer some of the high future consumption to the present by borrowing. A counterfactually high r s - r e rate i needed to discourage t i bor ikfe s hs rowing. (Recall that net borrowing must equal zero in equilibrium.) However, in a stochastic model this e f c i partially countered by the precautionary motive f r saving: agents might fet s o wish to insure against the possi i i y of consumption downturns. As the horizon lengthens, blt th s precautionary demand becomes weaker, since the empirical probability of a consump i tion downturn i smaller f r the longer horizons. (For example, Cochrane and Hansen (1992) s o note that there i no fiv - e r period in post-wax US data over which aggregate consumption s eya d c i e . As a r s l , the equilibrium r s - r e rate implied by the time-separable model i elns) eut ikfe s counterfactually high f r longer-horizon data. o What i needed, then, i a reason why agents would be willing to save at low i t r s r t s s s neet ae, even though they know that their future consumption i l k l to grow. One possible reason s iey i that agent’ within-period utility-of-consumption changes through time. Preference s e i i s s pcf cations with t i property include the Constantinides (1990) habit-formation preferences and hs Abel (1990) “ catching-up-with-the-Joneses” preferences. In these spe i i a i n , the marginal cfctos u i i y of a given l v l of consumption grows through time. Agents refrain from borrowing to tlt ee increase current consumption because they know that they w l need the consumption more il in the future. (As Weil (1989) has pointed out, the same e f c could be induced by having a fet 13 subsistence point that grows deterministically in time.) To put t i another way, the problem hs discussed by Cochrane and Hansen (1992) cam be resolved, in p i ciple, i agents are f a f l rn f eru, not of a decline in consumption, but of a decline in consumption p o in t, re la tiv e to so m e re fe re n ce where the reference point i s l grows at the same rate that consumption grows. To tef resolve the r s - r e rate puzzle in t i way, we w l u i i e the Abel (1990) and Constantinides ikfe hs il tlz (1990) models in our empirical tests.13 In the following sect o s we formalize these preference in, specif c t o s iain. 3.1.2 The Abel (1990) ” Catching-Up-With-the-Joneses” Preferences Let ct denote the per-capita consumption at date t. The agent s olves ma x U = E t y {c*+>)r=o () 7 ] 1 -7 fr'o subject to the usual budget constraint, where ht = f >Vi, r, > 0, 0 < S < 1. ( 8) i= l The interpretation i that agents compare their consumption to the consumption of their s neighbors (the "Joneses”) in the recent past. In the formal model, the neighbors’consumption i represented by c , and agents behave as i they have a subsistence point equal to t] times a s * f weighted average of the per-capita consumption l v l over the past m periods. Notice that ees agents treat h t as exogenous: the marginal u i i y of a fixed l v l of consumption inherits the tlt ee upward trend in ct, but agents cannot a t r the ht process by t le heir own actions. (Of course, in equilibrium c t = ct. ) With Abel preferences, (ct+r h t+ r) mr _ t+T " ( t - h t r* c Let the value function V ( W ,h ) ^ (9) be defined as the maximum value of the objective function that can be attained given i i i l wealth W and an i i i l subsistence point h . We define the nta nta c e f c e t of relative r s aversion (denoted R R A t ) by ofiin ik rV W W { W ,h ) R R A t = -W V w ( W ,h ) • ( 10) 13 We n ote th a t we have also investigated preferences in which u tility is based on consum ption relative to a d eterm inistic trend w ith sim ilar results for th e unconditional m om ent tests. 14 For Abel preferences, RRAt = (1 1) ct — ht Notice that the c e f c e t of r l t v r s aversion i time-varying, and everywhere exceeds 7 . ofiin eaie ik s The r s - r e rate puzzles w l be partially resolved by the Abel s e i i a i n s n e empirically, ikfe il pcfcto, ic, declines in ( * — c 3.1.3 h t) are observed more frequently than declines in c. * Constantinides (1990) Habit-Formation Preferences Constantinides (1990) models agents as maximizing an objective function of the same form as ( ) with the following alternative specification f r h t : 7 o ^ = 7 > 0 >0 < ^ < 1 ? i=l The difference between ( and ( 2 i that in ( 2 the stochastic subsistence point 8) 1) s 1) function of the agent’ own consumption s ct, (12) ht i a s rather than the per-capita consumption. The marginal rate of substitution i now s m r _ o r M U t+ r t+ T ~ P ‘M (3 1) U t where the marginal u i i y of consumption M U t i defined by tlt s M U t = ( c t ~ h tH - £ W t =1 * E t [ct+i ~ fti i + )]-7 (4 1) With habit-formation preferences, agents consider the e f c of their current consumption on fet future values of h t . The presence of conditional expectations in equation ( 14) r f e t t i elcs hs f c . These conditional expectations must be computed when we construct rri[. We do t i as at hs f l o s F r t define the variable D t by: olw. is, „ _, D, »(1 - = 1- S) E2.1 W * W---- ' s ^ n ( 15) The variable D t behaves as a stationary stochastic process. Equations (1 ) and (15) imply 3 that, in the Constantinides model, m,t+T i c t+ r h t+ T) (c t — h t) 15 7 D t+ r D t+ r + m 7 E tD t+ m (6 1) Since D t is stationary, we can fit an autoregressive time-series model for this variable, and use the fitted values as our estimate of E t D t + m . For most models, the likelihood ratio statistics testing n lags against n —1 lags in the autoregression for four lags. We estimate a fourth-order autoregression in (for Dt D t, n between 1 and 5) favor and project the fitted regression m periods into the future. We consider Constantinides preferences separately from Abel preferences for two reasons. First, it is possible that the behavior of in (13) may differ substantially from that implied by (9). A second, and more important reason, is that habit formation preferences do not accentuate risk aversion in the way that Abel’s preferences do:14 the coefficient of relative risk aversion implied by habit-formation for a given specification of {7 , r?, <, /?} is smaller than 5 that given in (11). From the perspective of the individual agent, Abel preferences are timeseparable, since a change in an individual’s ct does not affect his marginal utility with respect to c t+ i for i / 0. For such preferences, the coefficient of relative risk aversion equals the curvature of the per-period utility function U , as measured by . In contrast, habit formation preferences incorporate true time nonseparability: The marginal utility at date t + t is affected by changes in the state-contingent consumption plan for dates i = l, ...,m. In response to a wealth shock at date i, t, the agent with habit-formation preferences adjusts her state-contingent plans for future consumption so as to optimally adjust h t+ i, i = l , ..., m . This attenuates the impact of a given wealth shock on the objective function, as compared to the Abel specification. In particular, higher 77 does not increase R R A t as much as in (1 1 ). While RRAt cannot be computed analytically for our model of habit-formation, Constantinides (1990) and Person and Constantinides (1991) obtain closed-form solutions for RRAt in the context of a simpler model. They show that, for preference parameters similar to ours, the mean coefficient of relative risk aversion is not too far above 7 . 15 14This point is extensively discussed by Constantinides (1990), Person and Constantinides (1991), and Boldrin, Christiano, and Fisher (1995). 15Boldrin, Christiano, and Fisher (1995) report a similar result. 16 3 .2 Tests of Long-Horizon Pricing Using Unconditional M o ments In Section 3.3 we redo the long-horizon Hansen-Jagannathan analysis of Cochrane and Hansen (1992) using Abel (1990) and Constantinides (1990) preferences, and find that there i indeed s considerable support f r consumption based asset pricing models at long horizons, and very lt o i t esupport at short horizons. In th s section we describe the long-horizon Hansen-Jagannathan l i t s s and t s - t t s i s we u i i e et etsaitc tlz. 3 .2.1 The Hansen-Jagannathan Analysis Cochrane and Hansen (1992) derive the Hansen and Jagannathan (1991) mean-variance bounds using a linear projection of a candidate stochastic discount factor on the space of portfolio p y f s In th s s c i n we describe an alternative derivation that i helpful in pointing out the aof. i eto, s importance of the unconditional correlation between the pricing kernel and the mean-variance e f c e t p r f l o In vector form, Equation ( i: fiin o t o i . 1) s E [ m T R*T]= t 1. (7 1) where R [ denotes the vector consisting of the gross-returns between period t and period t + r on a s t of n a s t , and 1 i an n-vector of ones. Equation (1 ) holds f r a l r s , in the e ses s 7 o l , o remainder of th s s c i n we suppress the superscript r Equation (17) implies tha : i eto, . t co v (m , R) = B[mR] — E[m]E[R] 1 - E[m]E[R] = (8 1) Rearranging, we obtain f r any return r o : E [r) - E lm } - 1 = c o v (m , _ r E [m ] ~ ~ P m ’r /a mo r \ V^h J ‘ (9 1) where oy and crm are the standard deviations ofr and r , resp c i e y and p m r i the correlation a etvl, s c e f c e t between r and m. Since the correlation c e f c e t — pm r cannot be l s than — 1, ofiin ofiin es we see tha : t or ) 17 ~ E [m ] ( 20) where the ”< ” must hold f r every asset and every f a i l p r f l o Let us define the l f o esbe otoi. et hand side ofequation (2 ) as the asset’ S h a rp e 0 s ra t io .16 Hansen and Jagannathan (1991) derive the strongest bound implied by ( : in e f c , they evaluate the left-hand side of ( at the 20) fet 20) portfolio return with the highest Sharpe r t o which i simply the unconditional mean-variance ai, s e f c e t p r f l o This portfolio has a Sharpe ratio of fiin o t o i . s r * = v/(^[R] - 1 .E M - i y r r ^ R ] - 1 •%]->) < where f i the covariance matrix ofreturns. For any given value ofE [ m \ , this gives a bound on l s am which i dependent only on the asset return f r t and second moments, and i independent s is s of preferences. Equation (19) also t l sus something about the importance ofthe correlation between asset el returns and the pricing ke n l This equation can be rewritten: re. &m E [m ] _ ~ SRj pm,Rj' where S R j denotes the Sharpe ratioofassetj . This equation says t hat, f rany asset or portfolio o of assets with SRj > 0 the lower bound on em —► oo as p f 0-: i the pricing operator m , r f has an arbitrarily small negative correlation with some asset with a positive Sharpe r t o this ai, implies an a r b it r a r ily t ight bound on cm . This c early t e in with our results in Section 2: r l is the positive correlation between equity returns and consumption growth suggests a negative correlation between equity returns and the marginal rate of substitution in consumption that serves as the consumption-based proxy f r m. I the correlation between short-horizon equity o f returns and short-horizon consumption growth i positive but small, the implied bound on <jm s i l k l to be much tighter than the standard Hansen-Jagannathan bound. s iey 3.2.2 Projected Hansen-Jagannathan Bounds Tests Following Cochrane and Hansen (1992), we note that i m s t s i s ( 7 , then so w l the f aife 1 ) il projection (denoted m*) of m onto the s t of asset returns and a constant: e m* = a + R' •b, (21) 16The risk-free rate equals Et[mt+i]-1, so our definition of the Sharpe ratio differs from the usual definition by a "Jensen’s inequality” correction. If the risk-free rate were constant, the two definitions would coincide. 18 where m = a + R/ •b + e , and i?[eR] = 0 . If the correlation of m with the elements of R is very small, then the variance of m* will be very small relative to that of m . As in equation (20), this means that a higher value of crm will be required to satisfy the HJ bounds when implemented using m*. This test therefore provides a tighter restriction than the standard HJ test. We present HansenJagannathan tests both for m and m *. We also use the procedure described in section 3.1 of Hansen, Heaton, and Luttmer (1995) to test formally whether the Hansen-Jagannathan mean-variance restrictions are satisfied. 3.2.3 A x2 test of the moment restriction B[mR] = 1 Because we are not estimating model parameters in our test, it is straightforward to directly test the moment restriction £[m R —1] = 0. for the n assets being considered. If we define gt = (mtRf —1), the sample moment estimator is g r = (1 /T) Y j T g - We can construct a test t statistic J t = T g tp S o 1 gr> (2 2 ) where So is a consistent estimator of the spectral density matrix of g* at frequency zero, and where this spectral density matrix is estimated using the Newey-West procedure. 17 As shown in Hansen (1982), Jt is asymptotically distributed x 2 with n degrees of freedom. In our case, we use two distinct asset returns, so n = 2 . For the time-separable model, Cochrane and Hansen show that this x 2 test (for these two assets) strongly rejects the time-separable power utility model for 7 ’s of less than 50.18 3.2.4 The Hansen-Jagannathan (1997) Specification Test As pointed out in Hansen and Jagannathan (1997), x 2 tests of the moment restriction (17), like that discussed in Section 3.2.3, can be misleading in comparing specifications. If a model 17The number of lags used equals 8 , 11, 15, and 19 lags for the quarterly, yearly, two-year, and three-year horizons, respectively. 18See Cochrane and Hansen (1992), Table 2, p. 129. 19 produces a highly v l t l m, the eigenvalues ofthe spectral-density matrix may be huge, so the oaie weighting matrix Sq 1 in equation (22) may be close to singular. This would imply a “ small” X2 s a i t c even when the pricing errors are large.19 ttsi, To help to assess whether observed low x2 t s s a i t c r s from a superior f t of a e t t t s i s e ult i given model, we u i i e a distance measure derived by Hansen and Jagannathan (1997).2 We tlz 0 denote th s measure the ”HJ97 s a i t c This t s constructs an estimator the mean squared i t t s i ”. et distance between the candidate stochastic discount factor and the set ( M ) of valid pricing kernels (that i, the s t of y’ that s t s y E ( y R — 1) = 0 . Formally, this distance measure, s e s aif ) <2,i defined by 5 s 52 = min E ( y t — m t)2 (23) y& M As stated, t i minimization i over a s t of random variables M hs s e . Equation ( in Hansen, 8) Heaton, and Luttmer (1995) show that 8 can be computed more simply as the solution to the conjugate maximization problem. We follow t heirprocedure in computing (the finite-sample analogue to S2) f the various models of m t we study.21 or HJ97 s a i t c axe not directly comparable across d f e e t investment horizons without ttsis ifrn some normalization; here we derive a modified HJ97 s a i t c with appropriate normalization ttsi to allow such a comparison. The r period long horizon misspecification measure ( n our i notation) i: s % = min E(yJ yr 6 M r where /iT = E (y [ — ) and <2 = r va r(yl - m [)2 = p2 + a2 — mj"), where y is the a rg m in ( 24) of (24). If we assume that the “best-fit” pricing kernel y T will not change with horizon, then the r period pricing kernels 19For example, the x 2 test discussed in 3.2.3 (and presented by Cochrane and Hauisen (1992)), strongly rejects for small 7 ’s ( J t = 5.1 for 7 = 40 for our sample), but fails to reject the time-separable model for a 7 ’s over 200 ( J t is equal to 0 .8 8 for 7 = 240). However, the HJ97 statistic 6 2 (defined in equation (23)) is equal to 0.257 for 7 = 0 and 0.416 for 7 = 240. In other words, despite the low x 2 statistic, the m for the 7 = 240 model is fu rth e r from M than the naive (and obviously sub-optimal) specification m t = 1 , for all t. 20See also Hansen, Heaton, and Luttmer (1995) (HHL), Section 1 .2 ). 21 Hansen, Heaton, and Luttmer (1995) show that if m t G M (that is, if S = 0 ), then the limiting distribution of y / T d r is degenerate. For this reason, we do not use the HJ97 statistic to test whether mt is a valid pricing kernel. 20 in (24) are simply products of the single period pricing kernels: IIi=1mJ+i_1. Since, for small r, both yl — II[_1yt1 +i_1 and = and m should be close to one, and since: y N « 1+ n j)Li(i + ^ 2 5i i= 1 we have that: r til - m l = nf=1tit+i-i - n U i m l + i - i « J 2 (tit+ i-1 - W +i-i) ti = Thus, under this approximation, « T/Xi , Hr and a2 — where the variance ratio ( y T — m T) V R (t ) t • a f • V R ( t ), is defined in the usual way as the ratio of the variance of to the variance of (y1 —m1), divided by r: V R ( r ) , ^ . It follows that the HJ97 statistic can be written — n l + a* We now define the ^ r 2 • Ai + r • a i * as m o d ifie d H J 9 7 s t a t is t i c (5 1 )2 = £ T z (25) ’ V R ( t ). + £ = h (26) + <2 -V R (t ) t T For r = 1, (££)2 is equal to the standard HJ97 statistic in (24). To implement (26), we compute y as the a r g m in of the sample version of equation (24) ,22, we calculate the sample flT and dT, and form the square-root of the sample version of our modified statistic, <vT. In the empirical section, we use the modified HJ97 statistic S T to assess the fit of the model at different horizons. The modified HJ97 statistic tells us exactly what we want to know about how the char acteristics of (y —m ) change as the test horizon lengthens: if (y t —m$) is characterized by 22HHL show that the y € M which minimizes the sample version of equation (24) is y t = m t — R 'a, where a is the vector of Lagrange multipliers of the original minimization problem, and where the F.O.C. defining a is given in their equation (10). Therefore, (yt —m t) = a'R. /zr and a T are the sample mean and standard deviation of this series. 21 rapid, negatively autocorrelated movements which are “ washed-out” at longer horizons, then the variance-ratio in (25) w l be small and d\. w l be small f r long horizons. This would t l il il o el us that eliminating high frequency components improves the ft of the model. I, on the other i f hand, (yt — m*) i characterized by positively correlated, long-swings, then the variance-ratio s in ( 5 , and d . will be large f r long horizons, and the modified HJ97 s a i t c will indicate a 2) j o ttsi poorer ft to the data at longer-horizons. i 3.3 Empirical Results We now present the r esults from our empirical analysis of the models presented in Section 3 1 using the t ., ests presented in Section 3 2 We use quarterly data from 1952-1997. Detailed .. description of the data can be found in the Data Appendix. We follow the standard practice in the literature in using consumer expenditures on nondurables and services(”N D & S ”) as our primary measure of consumption. We also perform the analysis f r consumer expenditures on o nondurables ( D ”) F n lly, The data compiled by the Bureau of Economic Analysis data ”N . ia f r consumer expenditures on services includes a s r e measuring the imputed rental value o eis of owner-occupied housing. Rather than being compiled from observed surveys on consumer expenditure, th ss r e i a construct. I has rather d f e e t properties from other consumption i eis s t ifrn s r e . In particular, Marshall and Parekh (1998) note that the covariance between the growth eis rate of this s r e and equity returns i negative. To allow f r the possibility that this s r e i eis s o eis s substantially mis-measured, our third measure of consumption i N D&S minus this imputed s rental value of owner-occupied housing. 3.3.1 Time-Separable Preferences We examine time-separable preferences f r four horizons: one quarter, one year, two years, and o three years. We s t (3 equal to 1 0 and we vary the curvature parameter 7 . The behavior of e ., this model f r N D & S consumption i displayed in Figure 5.23 The s l d “ o s o i U-shaped” curves in Figure 5 plot the Hansen-Jagannathan bounds, calculated using two returns: the real CRSP 23We also conducted this analysis using non-durable consumption and using N D & S minus the service flow from owner-occupied housing. The results are very similar to those fore ND&S, and are not displayed. 22 F ig u re 5: H a n s e n a n d J a g a n n a t h a n b o u n d s t e s t s fo r T im e - S e p a r a b le P r e f e r e n c e s , N o n - D u r a b le a n d S e r v ic e s C o n s u m p t io n The four solid curves in both panels are the H ansen-Jagannathan bounds on the quarterly standard deviation of th e pricing kernel, inferred from the asset return data at the four horizons: one quarter, one year, tw o years, and three years. In the top panel, th e dotted lines p lot th e m eans/ standard d eviations for th e tim e-separable m odel m arginal rate o f substitution for th e follow ing four horizons: one quarter ( “+ ”), one year ( “*” ), two years ( “circle” ), and three years ( “x ”). W e set /? = 1 and we le t th e value o f 7 range from 0 to 200; for each line th e spacing betw een the sym bols is 7 = 5. In the bottom panel, th e dashed lines are generated by running a regression o f th e candidate discount factors on th e set o f real returns (in this case th e VW index and th e T -B ill rate). The m eans and standard d eviations o f th e fitted regressions are plotted for 7 = 0 , ..., 200. C onsum ption is m easured as expenditures on consum er nondurables plus services; th e tim e period is 1947:1-1997:4. Normalized Mean (Quarterly) 23 VW index return and the real 3-month T-Bill return. To facilitate comparison across different investment horizons, we normalize the plots to put both mean and variance of the pricing kernel in quarterly terms. To do this, we divide the resulting lower bound on a m by y /r , where r is the horizon length in quarters (1, 4, 8 or 12 in this plot). The mean is normalized by plotting .E[m]1/T. In the upper panel of Figure 5 we plot the mean-standard deviation loci of m t implied by time-separable preferences with ND&S consumption, for the four horizons and for values of 7 ranging from 0 to 200. In the lower panel, we project the time-separable model’s pricing kernel onto the set of asset returns as described in Section 3.2.2, and plot the mean and standard deviation of m* (defined in equation (21) against the same H-J bounds. Figure 5 confirms the failure of the time-separable model that has been noted by many previous studies. At every horizons from one quarter through three years, the (mean, stan dard deviation) loci of the pricing kernel implied by this model are substantially outside of the Hansen-Jagannathan bound. As in Cochrane and Hansen (1992), the longer horizons are fur ther from the bound than the quarterly horizon. We test formally these Hansen-Jagannathan mean-variance restrictions using the procedure described in Hansen, Heaton, and Luttmer (1995).242 The failure of the time-separable model is generally confirmed. In particular, the 6 5 Hansen-Jagannathan restrictions for the projected kernel are rejected at the 1% marginal sig nificance level for all horizons except when 7 is near zero25 or, for the quarterly horizon, for values of 7 exceeding 20 0.26 One should not, however, take these results as indicating that time-separable preferences can fit the data with 7 ’s near zero or (for the quarterly horizon) near 200. When x2 statistics axe calculated using equation (2 2 ). The statistics reject the model for all values of 7 . In particular, the smallest values of these x 2 statistics are 24.8, 16.8, 14.9, and 12.32 for horizons of one quarter, one year, two years, and three years, respectively. All of these values exceed 10.60, which is the 0.5% critical value for a x2 statistic with two degrees of freedom. The modified HJ97 statistics, given in equation (26), tell the same story. These statistics range between 0.273 and 0.437, implying that the root-mean-squared deviation of the model’s pricing 24See Hansen, Heaton, and Luttmer (1995), section 3.1. 25When 7 = 0 (risk-neutrality), this test has p-values of 0.073, 0.048, 0.074, and 0.056 for horizons of one quarter, one year, two years, and three years, respectively. 26The p-value for the quarterly horizon when 7 = 200 is 0.087. 24 kernel from the set of valid kernels is between 2730 bp and 4370 bp per quarter. 3.3.2 Time-Nonseparable Preferences: A n Overview of the Empirical Results In this section, we conduct an analysis of the Abel and Constantinides models analogous to that done in section 3.3.1 for time-separable preferences. That is, we look at the models’ performance as curvature increases. While the model with time-separable preferences has only two parameters, (3 (the subjective discount rate) and and Constantinides models have five parameters: depreciates), 77 7 (3, 7 (the curvature parameter), the Abel , S (the rate at which the habit stock (the ratio of mean consumption to mean habit stock) and m (the number of lagged consumptions that enter the habit stock). In this subsection, we proceed somewhat informally: We simply fix {f3 ,8 , 77, m} at values where the models perform fairly well27 and we vary 7 , as in figure 5. In section 4, below, we consider the models’ performance when the parameters are chosen optimally. Note that equation (14) for the marginal utility of consumption in the Constantinides model only makes sense if this marginal utility is positive. If the second term in equation (14) exceeds the first term, the expected effect of additional consumption on the future habit stocks would outweigh the effect on current period utility. In that case, it would be optimal for the consumer to reduce consumption via free disposal. In addition, the right-hand side of equations (9) or (14) could be negative or complex if (cj —ht) is negative. We discard any parameterizations of these models where either of these anomalies occurred. 3.3.3 The Abel Model For the Abel model, we set 77 = 0.8; 8 = 0.9; (3 = 1 ; and m = 20. Figure 6 is analogous to Figure 5. The upper panel plots the mean-standard deviation loci of m t implied by Abel with ND&S consumption, for the four horizons, and for various values of the curvature parameter 7 . The lower panel is the analogous plot for the projection of Abel model pricing kernel onto the set of asset returns. The upper panel shows that the model pricing kernels with ND&S consumption satisfy the 27We conduct a preliminary search in which we set (3 — 1.0 and we vary 77 and 8 from 0.5 through 0.9, we vary 7 from 0 through 30, and we vary m from 2 through 20. 25 Figure 6: Hansen and Jagannathan bounds tests for Abel Preferences, for Non-Durable and Services Consumption The four solid curves in both panels are th e H ansen-Jagannathan bounds on th e quarterly standard deviation o f th e pricing kernel, inferred from th e asset return data a t the four horizons: one quarter, one year, tw o years, and three years. In th e top panel, th e d otted lines p lot th e m eans/standard deviations for the A bel m odel m arginal rate o f su b stitu tion , as given in equation (9, for th e follow ing four horizons: one quarter ( “+ ”), one year ( “*”), tw o years ( “circle”), and three years ( “x ”). T he m odel param eters are rj = 0.8, 5 = 0 .9 , m = 20, fi = 1 . T he value o f 7 ranges from 0 to 30; for each line the spacing betw een the sym bols is 7 = 2. In the bottom panel, the dashed lines are generated by running a regression o f th e candidate discount factors on th e set o f real returns (in th is case th e V W index and the T -B ill rate). The m eans and standard d eviations o f th e fitted regressions are p lotted for 7 = 0, ..., 30. C onsum ption is m easured as expenditures on consum er nondurables plus services; th e tim e period is 1947:1-1997:4. 26 HJ bounds at the yearly (“*”) two-year (“circle”), , and three-year (“x”) horizons when 7 = 10-14. The quarterly horizon (“+ ”) requires a somewhat higher value of (over 20). The 7 lower panel, where the pricing kernels are projected onto the space of asset returns, give a somewhat different story. While the two-year horizon comes closest to the HJ bounds, none of the point estimates implied by the model actually satisfy these bounds. The reason is that the correlation between consumption growth and returns is small. However, when the HansenJagannathan mean-variance restrictions are tested formally using the Hansen, Heaton, and Luttmer (1995) procedure, the distance between the HJ bound and the (mean,variance) loci of the pricing kernels is significantly different from zero only for the quarterly horizon (and perhaps for the 3-year horizon) at low values of 7 . In particular, for 7 ’s in the range of 6-24, p-values of above 7 10 % are achieved for the three longer horizons. (For the quarterly horizon, ’s greater than 10 are required.) This means that, for moderately high values of 7 , none of the models are significantly outside the HJ bounds according to this test. The x 2 statistics, calculated using equation (22) give similar results. These statistics are graphed in Figure 7. For each of the four return horizons, the dashed line displays the value of the x 2 statistic as a function of 7 . The scale for this statistic is given at the left-hand side of each graph. The horizontal dotted line indicates the value of 5.99, which is the 5% critical value for a x 2 random variable with two degrees of freedom. This critical value is given only as a point of reference. If the parameters were chosen before looking at the data, the x 2 (2) would be the appropriate distribution under the null hypothesis that the model fits in population, since equation (2) must hold for each of two assets. However, the parameters used in this section were chosen after an informal grid search, so the x 2 (2 ) critical values should only be used as an informal guide. In figure 7, the value of the x 2 statistic drops rapidly for all horizons as 7 increases. For high enough values of 7 , there seems to be little evidence against the model. Note in particular the rapid monotonic decline for the quarterly horizon (panel A). Results like these lead some to argue that models with time non-separabilities are consistent with the behavior of short-term asset returns.28 An alternative interpretation, put forth by Hansen and Jagannathan (1997), is 28See, for example, Ferson and Constantinides (1991) and Campbell and Cochrane (1994). 27 Figure 7: x 2 statistics and H ansen-Jagannathan (1997) statistics for Abel Prefer ences. The figure plots the x 2 s a i t c (dashed l n ) and the HJ97 s a i t c ( o i f n ) implied by the Abel ttsi ie ttsi sld ie model at the quarterly (panel A), one-year (panel B), two-year (panel C), and three-year (panel D) horizon fo values of 7 ranging from 0 to 30 (horizontal a i ) The scale for the x 2 s a i t c i on the r xs. ttsi s left-hand side of the graph; the scale for the HJ97 s a i t c i on the right-hand side of the graph. ttsi s The horizontal dotted l n indicates the value of 5.99, which i the 5% c i i a value for a x 2 random ie s rtcl variable with two degrees of freedom. The model parameters are 77 = 0 8 S = 0 9 m = 20, / = 1 ., ., ? . Consumption i measured as expenditures on consumer nondurables plus s r i e ; the time period i s evcs s 1947:1-1997:4. Panel A: Quarterly Horizon Panel B One-Year Horizon : 28 that the x 2 statistic lacks power. In figure 7, the solid lines plot the modified HJ97 statistics as 7 increases. The scale for these statistics, given at the right-hand side of each graph, is chosen so that the HJ97 statistic and the x 2 statistic are at the same level for 7 = 0 . According to these results, only the longer horizons with 7 around 16-20 perform particularly well. A useful point of comparison for these statistics is the value when 7 = 0 , which implies a constant IMRS of one. The HJ97 statistics for 7 = 0 are therefore the root-mean square (RMS) distance from the set M. of valid pricing kernels to mj = 1. The HJ97 statistics for the quarterly test of the Abel model show that the RMS distance from the model to distance to m* = 1 for any M is not much smaller than the value of 7 . However, for the 1 and 2 year horizons, the statistics are about a factor of six smaller than for the model with constant m*: this indicates that m is indeed getting close to the set M .. The minimal value of the modified HJ97 statistic at the two-year horizon (figure 7, panel C) is 0.037, occurring at 7 = 18. That is, the minimal RMS error is on the order of 3.7%. The RMS error for the three-year horizon is approximately 4.5%, when 7 equals 16. We have conducted a similar analysis of the Abel model using the alternative consumption measures.29 When consumption is measured as purchases of consumer nondurables, the HJ97 statistics tend to be less favorable to the model. Higher values of 7 are required to get a reasonable fit at the two-year horizon, and the three-year horizon does poorly. In contrast, when the service flow from owner-occupied housing is omitted from the standard ND&S measure of consumption, the model tends to perform somewhat better. In particular, rather low values of the HJ97 statistic are found for the two- and three-year horizon when 7 equals 14.30 This is of interest because the service flow from owner-occupied housing is not an observed series, but is a construct that may be fraught with measurement error. 3.3.4 Constantinides Habit Persistence We conclude this section by performing an analogous study of a particular parameterization of Constantinides preferences. We measure consumption as ND&S, and set 5 = 0.9, 77 = 0.8, 29Detailed tabulations of these results are available from the authors. 30With this measure of consumption, the HJ97 statistic when 7 = 14 is 0.038 for the two-year horizon and 0.053 for the three-year horizon. 29 Figure 8 Hansen-Jagannathan bounds tests for Constantinides Preferences, for : Non-Durable and Services Consumption The description for figure 6,applies here a s , except that the preferences here are Constantinides; the lo model parameters are 17 = 0 8 6 = 0 9 m = 2, / = 1 ;the value of 7 ranges from 0 to 18 and the ., ., ? ; spacing between the crosses i 7 = 1 This figure i for Non-Durable and Services Consumption. s . s 30 and m = 2. The Hansen-Jagannathan bounds plots for are presented in Figure the Abel model displayed in Figure 6 8 . Unlike , the point estimates for this parameterization of the Constantinides model satisfy the HJ bounds only at the two-year horizon. When the pricing kernels are projected onto the space of asset returns (lower panel), none of the point estimates satisfy the HJ restrictions. As with the Abel model, the two-year horizon comes closest to the HJ bounds. However, the Hansen-Jagannathan mean-variance restrictions not rejected statistically by the Hansen, Heaton, and Luttmer (1995) test except at the three year horizon. Figure 9 is the analogue for Constantinides preferences to figure 7. The x 2 test appears less favorable to the quarterly horizon than with Abel preferences: In panel A of figure 9 the x 2 statistics exceed the 5.99 critical value for all values of 7 . The three-year horizon does somewhat more poorly according to this criterion with Constantinides preferences, as compared to Abel preferences. However, both the one- and two-year horizons do quite well for values of 7 exceeding 4. Moreover, the modified HJ97 statistics (solid lines) indicate that the x 2 statistics are small because the distance between m and M is indeed shrinking. For example, the RMS distance between the model’s IMRS and the set of valid pricing kernels is only 0.018 for the two-year horizon with 7 equal to 11. As with the Abel model, the Constantinides model does less well when consumption is measured by nondurables only. When housing services are omitted from the data, the results are similar to those for ND&S. To summarize the evidence from this section, the graphical H-J analysis, the x 2 statistics, and the HJ97 statistics suggest that the Abel model performs best at the two-year horizon and the Constantinides model does fairly well at both the one- and two-year horizons. 4 Evaluation of T i m e N o n - S e p a r a b l e Preferences W h e n M o d e l Parameters A r e Ch o s e n Optimally In the previous section, we simulate the Abel and Constantinides models over a grid of pa rameters. We found that, broadly speaking, these models appeared to perform poorly at the quarterly horizon, fairly well at the two-year horizon, and with intermediate performance at the one- and three-year horizons. In this section, we ask how well these models perform when the parameters are chosen optimally. In particular, for each model we wish to see whether 31 Figure 9 x2 statistics and Hansen-Jagannathan (1997) statistics for Constantinides : Preferences. The figure plots the x2 statistic (dashed line) and the HJ97 statistic (solid line) implied by the Con stantinides model at the quarterly (panel A), one-year (panel B), two-year (panel C), and three-year (panel D) horizon for values of 7 ranging from 0 to 30 (horizontal axis). The scale for the x 2 statistic is on the left-hand side of the graph; the scale for the HJ97 statistic is on the right-hand side of the graph. The horizontal dotted line indicates the value of 5.99, which is the 5% critical value for a x2 random variable with two degrees of freedom. The model parameters are r) — 0.8, S = 0.9, m = 2 , f = 1. Consumption is measured as expenditures on consumer nondurables plus services; the time 3 period is 1947:1-1997:4. Panel A: Quarterly Horizon Panel B: One-Year Horizon Panel C: Two-Year Horizon Panel D: Three-Year Horizon 32 there are parameter configurations that set the HJ97 statistic to zero (indicating that the mean residuals from the asset pricing Euler equations equal zero in sample). If a zero value of this criterion is achieved, we wish to see whether the parameter that achieve this value are ’’plausible” in some intuitive sense.31 In conducting this exercise, we note that the model is under-determined. We have only two unconditional Euler equations (one for the 3-month T-bill return, and one for the equity return). However, in each of the two models we have five parameters to choose: We set the quarterly subjective discount factor 0 {/3 , 7 , 77, 8, m } . = .99 (implying a yearly discount rate of approximately 4%). We treat m, the number of lagged consumption in the habit-stock formation, as an approximation to an infinite lag. We do so by letting m depend on set m as the smallest number such that Sm < 0.05. That is, we choose an m 8: we sufficiently big that the discarded lags (those greater than m) all have discount factors less than 5%. This still leaves us with three parameters {7 , 77, <} and only two model restrictions. Since this is a 5 highly nonlinear model, one could find no parameter combinations that set the HJ97 statistic to zero or one could find multiple parameter combinations that achieve this goal. To explore this question we fix 8 at three values {0.6,0.7,0.8), and for each value we minimize the HJ97 statistic over parameters {7 , 77 }. When 8 = 0.8 (the maximum value we use), m = 14, implying that the habit stock incorporates three-and-a-half years of lagged consumptions. Before describing our general results, we illustrate the behavior of a particular model. In figure 10 we consider how well the Constantinides model performs for two-year returns when 8 = 0.6. We consider values of 7 from zero to 40 (horizontal axis), and for each value of 7 we find the value of 7 that minimizes the HJ97 statistic. Note that both 7 and 7 are curvature 7 parameters. One might think that there is a trade-off between these two parameters: a high value of 7 acting similarly to a high value of 77. In the bottom panel of figure 10 we see that this is indeed the case: The optimal choice of 7 declines as 7 increases. It does not then follow, however, that 7 and 7 are not separately identified. As one can see in the top panel of figure 7 10, there is a unique {7 , 77} combination that minimizes the HJ97 criterion. This pattern holds 31 This exercise is analogous to that performed by Campbell (1998) for time-separable preferences. Using a variety of data sets from different countries, he computes the risk aversion parameter needed to match the mean equity premium. His test of the model is whether the requisite level of risk aversion is plausible from an economic standpoint. 33 Figure 1 : Optimal Choice of Curvature Parameters for Constantinides Model, T w o 0 Year Horizon For a grid of 7 *s ranging from zero to 40, the Hansen-Jagannathan (1997) statistic (”HJ97”) is min imized by choice of parameter t j . The Constantinides model of preferences is used with consumption measured by ND&S. Other model parameter are set as follows: /? = 0.99; 8 = 0.6, m = 6 . The top panel of this figure plots the minimized HJ97 statistic for each value of 7 . The bottom panel gives the minimizing choice of t) for each value of 7 . The time period is 1947:1-1997:4. 34 generally: In all model variants and all return horizons, for each < there is a uniquely-identified 5 optimal {7 , 77} combination. Table 2 gives the best-performing configuration of {5,7 , 77} for each model and each hori zon. (When multiple values of 5 achieve a zero value for the HJ97 statistic, we display the parameterization with the lowest value of 7 .) The results broadly confirm the patterns dis cussed in the previous section. First, for no model can we find a parameter combination that performs acceptably at the quarterly horizon. In the Abel model, the minimum HJ97 statis tic for quarterly returns is 0.214; for the Constantinides model, this minimum HJ97 value is 0.115. Varying the number of lags in the habit stock does not improve the performance of these models, nor does using alternative measures of consumption. 32 We conclude that these models are inconsistent with the unconditional moments of short-horizon asset returns. Second, for each model and each measure of consumption there exist parameter combina tions {7 , 77, <} that set the HJ97 statistic to zero for the two-year horizon. For the Constan 5 tinides model, this good fit to the data does not require an excessive value of the curvature parameter 7 . In particular, when < = 0.8, the Constantinides model requires 7 of only 7.4. 5 The Abel model requires somewhat higher values of 7 to set the HJ97 statistic to zero. The best fit for the Abel model at the two-year horizon occurs when 6 = 0.6. For this value of S, the optimal 7 equals 17.9. In that sense, one can argue that the Abel model has more difficulty than the Constantinides model in fitting the unconditional moments of the asset-return data. Furthermore, in the Abel model the coefficient of relative risk aversion is time-varying, but always exceeds 7 , often by a good deal. (See equation (11).) For example, the mean coefficient of relative risk aversion implied by the parameters that achieve a zero HJ97 statistic for the two-year horizon (5 = 0.6,7 = 17.9,7 = 0.85) is 114. Thus, the best-fitting parameterizations / of this model do imply extremely high (perhaps implausible) risk aversion. In contrast, it has been noted by Constantinides (1990), Ferson and Constantinides (1991), and Boldrin, Christiano, and Fisher (1995) that habit formation in the Constantinides model does not accentuate risk aversion to the extent that it does in the Abel model. While a closed-form expression for 32The minimal HJ97 statistics obtained for the quarterly horizon using nondurable consumption only are 0.232 and 0.136 for the Abel and Constantinides models, respectively. When housing services are omitted from ND&S, the corresponding statistics are 0.215 and 0.117. 35 the coefficient of relative risk aversion is not available for the Constantinides model, numerical analyses of simpler versions of the model by these authors suggest that the mean coefficient of relative risk aversion is only slightly higher than 7 .We conclude that the Constantinides model can match the unconditional properties of these asset returns at the two-year horizon without imposing extreme levels of risk aversion. Third, in both models we can find parameter combinations {7 , 7 ,5} that set the HJ97 cri terion to zero for the one-year horizon. This generally holds for other measures of consumption as well. 33 However, the requisite value of 7 needed at this one-year horizon is always sub stantially higher than that needed for the two-year horizon. Specifically, the best-performing parameter configuration for the Abel model at the one-year horizon sets 7 equal to 72. As with the two-year horizon, the Constantinides model requires a substantially lower value of 17.2. Still, this value of 7 exceeds that required to fit the data at the two-year horizon. Fourth, the Constantinides model does not set the HJ97 statistic to zero at the three-year horizon for any value of 8. The Abel model can achieve a zero value for the HJ97 statistic at the three-year horizon (when 8 equals 0.7 or 0.8), if a high value of 7 (in excess of 68) is assumed. Finally, the results do seem sensitive to the number of lagged consumptions (”m”) used to construct the habit stock. While the Constantinides model performs fairly well at the two-year horizon when 6 = 0.6 or 0.8, it does rather poorly when 8 = 0.7. The problem is not with the value of < p e r 5 m but with the value of m = 9 implied by 8 = 0.7. When we fix 8 = 6, the model behaves much like the case in Table 2 with 8 (but 5 se, 8 = = 0.7 but set 0.6. Similarly, when m = 10 remains fixed at 0.7), the model resembles the case in Table 2 with 8 = 0.8. Conclusions In this paper, we ask whether consumption-based pricing models work better at longer hori zons than at the quarterly horizon. Our motivation is that if frictions, transactions costs, or durability affect the comovements of consumption growth and asset returns, they should primarily affect the higher frequency components. 3 3 An exception is the Abel model using nondurable consumption, where the minimal value of the HJ97 statistic is 0.015 (when 8 is set to 0.8). 36 Table 2: Minim um H J97 statistics for A b el and Constantinides Models For each model, < is fixed at 0.6,0.7, or 0.8, m is set to the minimum value such that 6m < 0.05, and 5 the HJ97 statistic is minimized over {7 , 7 These tables report the minimal value of the test statistic ?}. obtained, for horizons of one quarter and 1, 2, and 3 years. Consumption is measured as ND&S. When the minimized HJ97 statistic equals zero for more than one value of <, the case with the lower value of 5 7 is displayed. Horizon (years) Abel Model HJjntn S 0.25 1 2 3 0.214 0.015 0.000 0.000 0.8 0.8 0.6 0.6 7 23.0 72.1 17.9 68.2 Constantinides Model i) m 0.79 0.59 0.85 0.79 14 14 6 6 HJmin S 6.115 0.8 0.8 0.8 0.6 0 .0 0 0 0 .0 0 0 0.014 7 28.9 17.2 7.4 58.8 V m 0.77 0.80 0.87 0.85 14 14 14 6 We first show that consumption-growth and equity returns are virtually uncorrelated at high frequencies. However, we find that at lower frequencies, corresponding to swings longer than one and one-half years, the two series are highly correlated. We then test three models of the pricing kernel: time-separable power utility; the Abel (1990) “Catching up with the Joneses” preferences; and the Constantinides (1990) habit-formation preferences. We find while all models perform poorly at quarterly horizons, the Abel (1990) and Constantinides (1990) models perform well at longer-horizons. The Hansen and Jagannathan (1991) bounds, modified to take account of correlation the correlation between the pricing kernel and asset returns, are satisfied. A x 2 moment restriction test is not rejected, and the modified Hansen, Heaton, and Luttmer (1995) specification test suggests that our inability to reject the model is not due high volatility of the pricing kernel. Our results raise numerous questions for further research. First, it suggests that correlation puzzle may at least partially be due to frictions that disrupt the high-frequency co-movements of marginal utility growth and returns. However, it also necessary to first understand why past attempts at modeling frictions to explain the equity-premium puzzle have not been entirely successful. The Abel and Constantinides models motivate agents to save at relatively low interest rates, even though consumption grows, because the marginal utility-value of a given level of consumption also grows. There may be other ways to motivate saving, such as by looking 37 carefully at life-cycle models of saving and investment. Finally, we consistently find that the models with time-nonseparabilities perform better at the two-year horizon than at the three-year horizon. (The Constantinides model does rather poorly at the three-year horizon.) This represents somewhat of a puzzle. If the only reason for the poor performance of consumption-based models at short horizons is transient, short-term frictions, then one would expect the models’ performance to improve monotonically as the horizon lengthens. One possible reason is that consumers are particularly averse to risk at business cycle frequencies. At the three year horizon, much of the business cycle behavior of consumption growth and asset returns is filtered out. However, it is unclear why consumers would be relatively insensitive to lower-frequency risk. Whether this reflects a degree of myopia or a hitherto un-modeled aspect of preferences remains to be determined. 38 References Abel, Andrew B., 1990, Asset prices under habit formation and catching up with the Joneses, A m e r ic a n E c o n o m ic R e v ie w 80, 38-42. Baxter, Marianne, 1994, Real exchange rates and real interest rate differentials: Have we missed the business cycle relationship?, J o u r n a l o f M o n e t a r y E c o n o m ic s 33, 5-37. -------- , and Robert King, 1994, Measuring business cycles: Approximate band-pass filters for economic time series, unpublished manuscript, Department of Economics, University of Virginia. Bloomfield, Peter, 1976, Sons: New York). F o u r ie r A n a ly s is o f T im e S e rie s : A n I n tro d u c tio n (John Wiley & Boldrin, Michele, Lawrence Christiano, and Jonas Fisher, 1995, Asset pricing lessons for mod eling business cycles, unpublished manuscript. Brainard, W.C., M.D. Shapiro, and J. Shoven, 1991, Market value and fundamental value, in W.C. Brainard, W.D. Nordhaus, and H.W. Watts, ed.: M o n e y , M a c ro e c o n o m ic s , a n d E c o n o m ic P o lic y : E s s a y s in H o n o r o f J a m e s T o b in (MIT Press: Cambridge). Breeden, Douglas T., 1979, An intertemporal asset pricing model with stochastic consumption and investment opportunities, J o u r n a l o f F in a n c i a l E c o n o m ic s 7, 265-96. Campbell, John, 1998, Asset prices, consumption, and the business cycle, NBER Working Paper No. 6485. Campbell, John Y., and John H. Cochrane, 1994, By force of habit: A consumption based explanation of aggregate stock market behavior, Working paper, University of Chicago. Cecchetti, Stephen G., Pok-Sang Lam, and Nelson C. Mark, 1993, The equity premium and the risk-free rate: Matching the moments, J o u r n a l o f M o n e t a r y E c o n o m ic s 31, 21-46. Chen, Nai-Fu, 1991, Financial investment opportunities and the macroeconomy, F in a n c e 46, 529-55. Jo u rn a l of Cochrane, John H., 1989, The sensitivity of tests of the intertemporal allocation of consumption to near-rational alternatives, A m e r ic a n E c o n o m ic R e v ie w 79, 319-337. -------- , and Lars Peter Hansen, 1992, Asset pricing explorations for macroeconomics, in Olivier Blanchard, and Stanley Fischer, ed.: 1992 N B E R M a c ro e c o n o m ic s A n n u a l (Cambridge: MIT Press). Constantinides, George M., 1990, Habit formation: A resolution of the equity premium puzzle, J o u r n a l o f P o lit ic a l E c o n o m y 98, 519-543. -------- , and Darrell Duffie, 1996, Asset pricing with heterogeneous consumers, P o lit ic a l E c o n o m y 104, 219-240. Jo u rn a l of Daniel, Kent D., and David Marshall, 1997, Equity premium puzzle and risk-free rate at long horizons, M a c ro e c o n o m ic D y n a m ic s 1 , 452-484. 39 Fama, Eugene F., and Kenneth R. French, 1989, Business conditions and expected returns on stocks and bonds, J o u r n a l o f F in a n c i a l E c o n o m ic s 25, 23-49. Ferson, Wayne E., and George M. Constantinides, 1991, Habit persistence and durability in aggregate consumption; empirical tests, J o u r n a l o f F i n a n c i a l E c o n o m ic s 29, 199-240. Grossman, Sanford J., and Guy Laroque, 1990, Asset pricing and optimal portfolio choice in the presence of illiquid durable consumption goods, E c o n o m e t r ic a 58, 25-51. Grossman, Sanford J., and Robert J. Shiller, 1982, Consumption correlatedness and risk mea surement in economies with non-traded assets and heterogeneous information, J o u r n a l o f F i n a n c i a l E c o n o m ic s 10, 195-210. Hansen, Lars P., 1982, Large sample properties of generalized method of moments estimators, E c o n o m e t r ic a 50, 1029-1054. -------- , John Heaton, and Erzo Luttmer, 1995, Econometric evaluation of asset pricing models, R e v ie w o f F i n a n c i a l S tu d ie s 8, 237-274. Hansen, Lars P., and Ravi Jagannathan, 1991, Implications of security market data for models of dynamic economies, J o u r n a l o f P o lit ic a l E c o n o m y 99, 225-262. -------- , 1997, Assessing specification errors in stochastic discount factor models, F in a n c e 52, 557-590. Jo u rn a l of He, Hua, and David M. Modest, 1995, Market frictions and consumption-based asset pricing, J o u r n a l o f P o lit ic a l E c o n o m y 103, 94-117. Heaton, John C., and Deborah J. Lucas, 1995, The importance of investor heterogeneity and fi nancial market imperfections for the behavior of asset prices, C a m e g ie - R o c h e s t e r C o n fe r e n c e S e r ie s o n P u b lic P o lic y 42, 1-32. Lucas, Deborah J., 1994, Asset pricing with undiversifiable income risk and short sales con straints: Deepening the equity premium puzzle, J o u r n a l o f M o n e t a r y E c o n o m ic s 34, 325341. Lucas, Jr., Robert E., 1978, Asset prices in an exchange economy, E c o n o m e t r ic a Luttmer, Erzo. G. J., 1996, Asset pricing in economies with frictions, 1467. 46, 1426-14. E c o n o m e t r ic a 64, 1439- Marshall, David, 1994, Asset return volatility with extremely small costs of consumption ad justment, Federal Reserve Bamk of Chicago Working Paper No. WP-94-23. Marshall, David A., and Nayam G. Parekh, 1998, Can costs of consumption adjustment explain asset pricing puzzles?, J o u r n a l o f F in a n c e , forthcoming. Weil, Philippe, 1989, The equity premium puzzle and the risk-free rate puzzle, M o n e t a r y E c o n o m ic s 24, 401-421. 40 Jo u rn a l of Appendices A C o n s t r u c t i o n of t h e D a t a The total quarterly real non-durable, durable, services, and total consumption series, the deflators for each of the three components series, and the population series (GPOP) were extracted from CITIBASE for the 1947:1-1997:4 period. Monthly VW index returns were obtained from CRSP, and were cumulated to obtain quar terly returns. One month T-Bill returns were taken from the CRSP RISKFREE file. One, two, and three year nominally risk-free rates were computed as the returns to one-, two-, and three-year zero-coupon bonds, computed from the Fama-Bliss data in CRSP. B Constr u c t i o n of the C o h e r e n c y a n d P h a s e Estimates a n C o n f i d e n c e Intervals To construct the coherency and phase estimates plotted in Figures 1 and 2, the two series x and y (log consumption growth and returns) and are first each subdivided into N non-overlapping subsamples of length n (here, n = 16 quarters). Each of these subsamples is then detrended and windowed using a Hamming window. The detrended, windowed subsamples are then fast-fourier transformed to generate J x (u>) and J y (w), which are equal to, for x, n- 1 Jx{u ) = n- 1 ^ 2 xte*™ 1. t= o where each of the subsamples is indexed from t = 0, ...,n —1 . Note that J x and J y will be complex. The x and y power-spectral densities and the cross-spectral densities are then defined as: k=1 a ,M = ^ where J^(w) denotes the Fourier transform of the fc’th subsample, and * denotes the complex conjugate. Although Pxx and p yy are real valued, in general the cross- spectral density will not be. The coherency between the two series is defined as: S .J , M = ia » i [PXX(w)Pyy (u>)]1//2 41 and the phase is defined as: < x,y(u ) j> = a rc ta n ( Irn (P xy(u )) \R e{P xy(u > )) y where I m (-) and i?e(-) denote the imaginary and real components, respectively. With these definitions, PXy(w ) = (Pxx((j)P yy(u >)) ^ SXty(l^)ei^x'v^ °\ Finally, confidence intervals for the coherency and phase were calculated using the method described in Bloomfield (1976, Section 9.5). The upper and lower bounds of the 95% confidence intervals are therefore: ta n k ^a rc ta n h {P xy(u;)) ± where g 2 — (2/3)(n/T), is a constant based on the Hamming window (see p. 224 of Bloomfield (1976).) The 95% confidence intervals for the phase are: 42