1 Overview Let’s …rst look at standard preferences –additive expected utility and de…ne some notation. Time is discrete, t = 0; 1; :::T: At each t 0, an event zt is drawn from a set t : A history, denoted z t is a collection of events up to an including zt , i.e., e.g., fz0 ; z1 ; :::; zt g : The set of possible histories at t is t : Let c (z t ) denote the vector of consumption goods (possibly including service ‡ows and leisure) consumed in period t if z t is realized:Then, the T standard assumption is that for any t; preferences over fc (z t ) jz t 2 t gt=0 with associated probabilities p (z t ) can be represented by the function U t t c z jz 2 t T t=0 = T X t t=0 X zt 2 t p z u c z t =E t T X t u c zt t=0 where u is a smooth increasing concave function, called the per-period utility or felicity function. Several important features of these preferences will be relaxed during the course. 1. Curvature of u determines elasticity of substitution both between states within a period (inverse of risk aversion) and between periods, intertemporal elasticity of substitution. Therefore, these intrinsically di¤erent concepts cannot be disentangled theoretically. 2. No preference for time revelation –it does not matter when we get to now our faith (unless, of course, if we can condition our actions on the information). 3. No history e¤ect; the utility E T X t c zt t=s is independent of consumption taking place before s: (However, history of course matters by a¤ecting the budget, i.e., the consumption possibility set, usually simply captured by a state variable like wealth. 4. Time consistency; if U c1 ; c z t then U c zt T t=2 T t=2 U c1 ; c0 z t U c0 z t T T t=2 T t=2 ; T i.e., if the sequence of consumption fc (z t )gt=2 is preferred over fc0 (z t )gt=2 at t = 1; it is preferred also at t = 2: 1 5. Risk matters only through second order e¤ects. E.g., suppose in period t; there is a lottery with payo¤ x~ and E x~ > 0. Then all individuals will prefer to hold a strictly positive amount of this lottery over having a certain consumption level c. I.e., there is a k > 0 such that for all k 2 0; k E (u (c + k~ x)) > u (c) : To see this, let’s take the derivative of the left hand side w.r.t. k and evaluate at k = 0: @ (E (u (c + k~ x))) jk=0 = u0 (c) E x~ @k which is positive if E x~ > 0; regardless of riskaversion. Empirical …ndings, introspection and lab experiments have all shown that these implications are often invalidated. 1.1 Two important macro-puzzles. 1. Equity premium puzzle 2. Too little insurance puzzle. Most well-known example. Mehra-Prescott, "equity premium puzzle". Consider the following example (almost) in there original formulation and we will return to it later. Suppose consumption grows fast with p = 1=2 and slowly otherwise. Starting from a consumption level ct , history at t + 1 is either z t+1 = zh in which case ct+1 = ct (1 + g + ) or z t+1 = zl and ct+1 = ct (1 + g ) : We can now using the standard Euler condition for maximizing utility to price a bond that gives 1 unit of consumption in both cases and a share that gives 1+g + or 1+g . The share is thought to be a claim to the process that provides the consumption possibilities. Individuals eat only apples. Apples are grown on a tree (or a number of identical trees( which has exogenous but stochastic output that grows with a rate g + or g . A share is ownership of the tree. The Euler equation is u0 (ct ) = Eu0 (ct+1 ) Rt+1 : Well known intuition, give up one unit of consumption "costs" u0 (ct ) ; invest it and increasing consumption by Rt+1 next period gives an increase in utility given by Eu0 (ct+1 ) Rt+1 : We do this by de…ning state-contingent prices. In period t there is one lottery that gives 1 unit of consumption in if growth is high and zero otherwise. There is also a lottery that gives one unit of consumption in the low 2 growth state. The price of these lotteries, ph and pl are and the returns and p(zl ) pl p(zh ) ph respectively. Therefore, u0 (c (zh )) p (zh ) u0 (ct ) u0 (c (zl )) p (zl ) u0 (ct ) ph = pl = Using the standard CRRA felicity function u (c) = 1 c 1; recalling that the coe¢ cient of RRA is c u00 (c) = u0 (c) Setting p (zh ) = p (zl ) = 1 2 c ( 1) c c 2 1 =1 : we have ph = pl = (1 + g + ) 2 (1 + g ) 2 1 ; 1 : A portfolio consisting of one each of the lotteries mimics perfectly the safe one-period bond. The price of a bond is thus pb = ph + pl with a return, rb = 1 : ph + pl Let us now compute the return on a claim to next periods dividends –a one-period share. A portfolio consisting of (1 + g + ) of the h lottery and (1 + g ) of the l lottery exactly such a one period share. The price of this risky portfolio is pr = ph ((1 + g + )) + pl (1 + g 3 ) and its expected return is therefore rr = = = 1+g ph ((1 + g + )) + pl (1 + g 1+g (1+g+ ) 2 1 (1 + g + ) + 2 (1 + g) ((1 + g + ) + (1 + g ) (1+g ) 1 2 (1 + g ) ) ) In this simple economy, the return on a normal share, i.e., a one that gives rights to all future dividends is the same. Why? To see this, we recall that the price of the share with CARA utility will be proportional to current income/consumption. The price of the share will therefore be Pr ct . and the return (1 + Pr ) ct+1 Pr ct+1 + ct+1 = : Pr c t Pr c t From the Euler equation, Pr = E u0 (ct+1 ) (1 + Pr ) ct+1 u0 (ct ) ct 1 + Pr Pr 1 (ct (1 + g + )) + (ct (1+g2 = (1 + Pr ) ct 1 ct (1 + g + ) + (1 + g ) = (1 + Pr ) : 2 2 = ((1 + g + ) + (1 + g ) ) (ct (1+g+ )) 2 )) 1 (ct (1 + g Finally, calculate the expected return on the share: 1 + Pr ct+1 E Pr ct 1 + pr = (1 + g) pr 2 (1 + g) = ((1 + g + ) + (1 + g ) ) ; as with the one-period share. Using (US) data g is around 1.8% per year and is around 3.6%. Stock market returns have averaged around 8% per year and the risk-free rate 4 )) around 1% over the last 100 years or so. Therefore (1 + 0:018 + 0:036) 1 2 (1 + 0:018 0:036) 1 = 2 1 = ph + pl 1 + 0:018 = ph (1 + 0:018 + 0:036) + pl (1 + 0:018 ph = pl rb rr Can we …nd 0.94 and 0.96 0:036) to generate the observed values of ra and rb ? 0.98 1 1.02 1.04 1.06 1.08 1.1 0 -2 -4 -6 -8 -10 Combinations of and such that ra = 0:08 (red) and rb = 0:01 (black) If, for example, we set = 0:98; riskaversion. 1 should be 3.5, to motivate 8% stock return. But then the bond return should be 7.6%, leaving a mere 0.4% risk premium. In fact, it is di¢ cult to get the right risk premium. Let’s look closer at the risk premium. Let’s express it as the ratio of the price of the bond to the ratio of the price of the risky asset pb ph + pl = pr ph (1 + g + ) + pl (1 + g (1 + g + ) 1 + (1 + g = (1 + g + ) + (1 + g 5 ) ) ) 1 In reality, this ratio is 1 1:01 1+g 1:08 1:05 However, by plotting " (1 + g + ) 1 + (1 + g (1 + g + ) + (1 + g against RRA = 1 ) ) 1 # g=0:018; =0:036 ; 1.015 1.01 1.005 1 0.995 0.99 0.985 0 20 40 R 60 80 100 we see that it is very di¢ cult to get the right risk premium. In fact, in the realistic case where > g; it is easy to bound the risk premium, " # " # 1 1 1 (1 + g + ) + (1 + g ) (1 + g ) = lim lim ! 1 ! 1 (1 + g + ) + (1 + g ) (1 + g ) g< = 1 1+g 1:0183: The …st line comes from the fact that (1 + g ) goes to in…nity as approach 1; while (1 + g + ) approach zero. Of course, with lower growth and higher risk, the risk premium can get larger, but we are stuck with data. 2. Too little risk-sharing In a complete markets equilibrium where individuals have homothetic preferences, e.g., CRRA, there should be full risk sharing. Consumption 6 growth should be perfectly correlated between individuals and everyone should hold a share in a global portfolio of assets. This is not the case, obviously frictions and asymmetric information may be one explanation. But sometimes these explanations don’t seem to su¢ ce. An example is the home bias puzzle. All around the world local investors hold unbalanced portfolios with to much domestic assets. It is shown in the literature that expected returns could increase a lot, without increasing risk by having more balanced portfolios, containing more foreign assets. The explanation cannot be that information is superior. Then, domestic holders should sometimes have more negative information than foreign investors, in which case they should sell moving to foreign ones, this we don’t see. Conversely, they should sometimes go short abroad, having an investment share above unity at home, which we don’t see either. 1.2 Lab puzzles Ambiguity aversion –Ellsberg Paradox. Consider following lottery. There are two urns, each with 100 balls. In urn 1, there are 50 red and 50 black. In urn 2, there are only red and black balls but the proportions are unknown. The subject is given a color and can pick one ball. If a ball with the given color comes up, the gain is 50$, if not the gain is zero. The subject is asked to rank lotteries. Typically the following response comes up. 1. Red from urn 1 Black from urn 1. 2. Red from urn 2 Black from urn 2. 3. Red from urn 1 Red from urn 2. 4. Black from urn 1 Black from urn 2. This contradicts expected utility since from 2, we expect subjective probability to imply that they believe p (red) = 1=2: Then, 3 and 4 should be with indi¤erence. Time inconsistency. In lab experiments, preference reversal occurs. Example. Suppose you can choose between 10CD’s 1 year from now or 11 CD’s 1 year and a week. Often the latter is preferred. However, after a year, 10 CD’s today is preferred over 11 CD’s in a week. This is inconsistent with standard time-additive utility with geometric discounting. 7 Other examples, people sometimes seems to pay to commit. They tend to over-consume during the year, and, for example, ask their employer to keep money for tax-payments at then en of the year or, say for big holidays. First-order risk-aversion With smooth preferences, people should as we have seen not care much about small gambles. Big ones, on the other hand, are detrimental. In fact, with CRRA coe¢ cient bigger than unity, su¢ ciently big losses can never be compensated since U (ct ) is bounded from above but not from below. For example, consider a lottery that gives a relative loss of x, forcing a consumption loss of xc with p = 21 and otherwise gives consumption (1 + k) c. For di¤erent values of x; how large must k be to compensate for so that 1 U (c) = U ((1 2 1 x) c) + U ((1 + k)c) 2 Here, I plot this k as a function of x for = 3; 4 and 6. 1.8 1.6 1.4 1.2 k1 0.8 0.6 0.4 0.2 0 0.02 0.04 0.06 0.08 0.1x 0.12 0.14 0.16 0.18 0.2 :To get people to behave like they do for small gambles, has to be so large as to give unreasonable predictions for large gambles. In fact, sometimes no upside can compensate for a su¢ ciently large but …nite downside. This fact is due to that when CRRA coe¢ cient larger than 1, i.e., when < 0; utility is bounded, since c < 0; 8c; < 0: 8 This means that we solving c 1 ((1 x) c) 2 1 (1 x) 1 = 2 1 x = 1 2 = gives the largest possible downside that could be compensated by any upside. In the graph, we see the maximum loss occuring with 50% chance that could be compensated by any gain as a function of the level of riskaversion. 0.8 0.6 0.4 0.2 2 4 6 8 10R 12 14 16 18 20 For = 11; x is as low as 6:1%: Would you refuse a 50/50 bet of loosing 25% of your lifetime income vs. getting the fortune of Bill Gates? If, not, you cannot have absolute riskaversion above 3.4. 9 2 Non-additive recursive preferences 2.1 Aggregation over time Now, disregard risk. In general, preferences can be described as a function that associates a particular level of overall utility to any sequence of consumption levels U (c1 ; c2 :; ; ; cT ) U fct gT0 MRS is de…ned @U (c1 ;c2 :;;;cT ) @ct+1 @U (c1 ;c2 :;;;cT ) @ct M RSt;t+1 We de…ne time preference as MRS along a path of constant consumption c (c)t;t+1 = @U (c1 ;c2 :;;;cT ) @ct+1 j @U (c1 ;c2 :;;;cT ) ct =c8t @ct noting that this may depend onf c: For the time additive utility with constant discounting, however, we have U= T X t u (ct ) t=s with (c)t;t+1 = 8c: Koopman’s time aggregator Assume preferences at all dates are represented by a time zero utility function, so preferences are time consistent. First notation, fct ; ct+1 ; ct+2 ; ::::ct+1 g tc Utility at time zero is U (0 c) = U (c0; 1 c) Assume history independence, here marginal rate of substitution M RSt;t+1 does not depend on consumption prior to t (is this innocuous?) and if ct is a vector, also the intra-temporal MRS between goods in t; is independent of prior consumption. Then, but not otherwise, we can write U (0 c) = V~ [c0 ; U1 (1 c)] 10 for an aggregator function V:and a function that gives the continuation utility U1 (1 c) Choices over 1 c1 in particular, what maximizes U1 in some choice set, does not depend on c0 : But the choice set can, of course, be a¤ected. Now also assume future independence preferences over ct does not depend on t+1 c: (Is this innocuous? Yes, clearly if c0 is a scalar, then more is just better, but if c0 is a vector this is a restriction..One could prefer chicken over …sh if one plans to eat a lot of …sh in the future. However, future independence seems like a less strong assumption than history independence). Now, we can write utility as U (0 c) = V [u (c0 ) ; U1 (1 c)] V aggregates utility coming from current consumption, and future consumption. It is not restricted to simply add them like standard preferences. Finally, assume stationarity, then for all t, U (t c) = V [u (ct ) ; U (t+1 c)] ; and recursivity is implied U (t c) = V [u (ct ) ; V [u (ct+1 ) ; U (t+2 c)]] MRSt;t+1 is @U (c1 ;c2 :;;;cT ) @ct+1 @U (c1 ;c2 :;;;cT ) @ct = V2 [u (ct ) ; U (t+1 c)] V1 [u (ct+1 ) ; U (t+2 c)] u0 (ct+1 ) V1 [u (ct ) ; U (t+1 c)] u0 (ct ) As we know, time preference is MRS evaluated at a constant consumption path, where by stationarity, also u (ct ) and U (t c) is constant at u (c) and U (c) (excuse the notation, I am here letting c denote a path of constant levels of consumption). Then, (u (c)) = V2 [u (c) ; U (c)] which can depend on c unless V is a linear aggregator (standard). The Uzawa simpli…cation is a particular example of the Koopmans aggregator. U (t c) = u (ct ) + (u (ct )) U (t+1 c) First order condition: @u (ct ) (1 + @ct;i 0 (u (ct )) U (t+1 c)) + (u (ct )) 11 @U (t+1 c) =0 @ct;i A bit more general, by not imposing future independence. U (t c) = u (ct ) + (ct ) U (t+1 c) Note that here, preference over elements in ct may depend on U (t+1 c), which matters if ct is a vector. First order condition: @u (ct ) @ (ct ) @U (t+1 c) + U (t+1 c) + (ct ) =0 @ct;i @ct;i @ct;i Examples: Growth and …scal policy (Dolmas and Wynne (1998)). Using Usawa max U (t c) = u (ct ) + (u (ct )) U (t+1 c) s:t:ct = f (kt ) kt+1 gt We can derive a Bellman equation: J (k) = max u (f (kt ) kt+1 kt+1 gt ) + (u (f (kt ) kt+1 gt )) J (kt+1 ) : Only non-standard is endogenous discounting. FOC: u0 (ct ) (1 + 0 (u (ct )) J (kt+1 )) = (u (ct )) J 0 (kt+1 ) Envelope: J 0 (kt ) = u0 (ct ) f 0 (kt ) + 0 (u (ct )) u0 (ct ) f 0 (kt ) J (kt+1 ) = u0 (ct ) f 0 (kt ) (1 + 0 (u (ct )) J (kt+1 )) : Giving J 0 (kt ) = (u (ct )) J 0 (kt+1 ) f 0 (kt ) In a steady state 1= (u (ct )) f 0 (kt ) Compare this to the standard case 1 = f 0 (kt ) being independent of …scal policy. In particular, an increase in g; must reduce c one-for-one, since css = f (kss ) kss gt . 12 Now changes in g can a¤ect the steady state. To see this, consider 1= (u (f (kss ) kss g)) f 0 (kss ) An increase in g reduces u, suppose this makes people more patient, i.e., < 0: Then, the increase in g makes (u) f 0 > 1. This will lead to more saving and a growing capital stock. Crowding out of consumption more than one-for-one. In this case, since both 0 and f 00 are negative, there is a unique steady state. See left panel of the …gure. 0 If, instead > 0; there may be multiple solutions to 0 1= (u (f (kss ) g)) f 0 (kss ) : kss In this case, some steady states are unstable. In the right panel of the …gure, the left steady state is unstable. A higher level of k increases u; and more than the fall in f 0 : Therefore, > f 0 and individuals accumulates capital. Here a reduction in g; could move the economy out of the unstable equilibrium. β (u ( f (k ss )− k ss − g ')) −1 β (u ( f (k ss )− k ss − g )) −1 β (u ( f (k ss )− k ss − g ')) −1 f ' (k ss ) β (u ( f (k ss )− k ss − g )) −1 f ' (k ss ) k ss 0 k ss 0 < 0; g increases. > 0, g decreases. Other examples is small open economies with a …xed interest rate, r The steady state is 1 = (u (f (kss ) kss g)) r With standard preferences no steady state exists generically. With a unique one exists. 13 0 <0 2.2 Aggregation over states As in the intro, consider states (within a period) to be z 2 associated probability measure p (z) : Utility is now ; with an U fc (z)gz2 We can solve for the certainty equivalent consumption level from U fc (z)gz2 = U (f g) of fc (z)gz2 Standard theory says X U (fc (z)g) = p (z) u ((z)) z2 and since u( ) = X p (z) u (c (z)) z2 we have (fc (z)g) = u X 1 ! p (z) u (c (z)) : z2 For example: Suppose preferences are CRRA, then u (c) = 1 c So 1 = X 1 p (z) c (z) z2 = X 1 p (z) c (z) z2 !1 Note the linear homogeneity in this case. For a constant k > 0: (k fc (z)g) = k (fc (z)g) : Chew and Dekel generalizes this by allowing the certainty equivalent of fc (z)g to be a more general function while maintaining …rst order conditions that are linear in probabilities by implicitly de…ning X (fc (z)g) = p (z) M [c (z) ; ] : z2 14 As we see, this generalizes standard utility by implying that the marginal value of consumption in state z depends on consumption in other states through their e¤ect on : An example of this is that people might care more (or less) about consumption in states that provide less consumption than the certainty equivalence (disappointment aversion). Notice the relative comparison here. With concave utility, marginal utility in states with low consumption is high, but independent of consumption in other states. This is not necessarily the case here since consumption in state z; relative to depends on consumption in all other states since they a¤ect : We assume that M ( ; ) = (Why?), M1 > 0; M11 < 0 (…rst order stochastic dominance and riskaversion). Often we want to maintain the linear homogeneity of preferences like in CRRA. M (kc; k ) = kM (c; ) Examples: To show that the Chew-Dekel generalizes and includes e.g., CRRA preferences: Note that if we set M (c; ) = 1 c (1) + we get = X p (z) 1 c (z) + z2 0 = X z2 = = 1 X X p (z) c (z) z2 p (z) (c (z) ) z2 = 1 p (z) c (z) X p (z) c (z) z2 !1 which is the CRRA certainty equivalence. Examples: "Weighted expected utility" Let c 1 c M= 15 + Compare to (1). Now, we have = X p (z) c (z) c (z) 1 c (z) c (z) 1 c (z) + z2 0 = X X p (z) z2 p (z) c (z) = z2 = X 1 P Here, we can interpret c (z) p (z) c (z) c (z) z2 z2 p (z) c (z) c (z) z2 p (z) c (z) P p (z) c (z) z2 p (z) c (z) P p^ (z) as a weighted probability. If < 0; these weights decrease in c; i.e., bad outcomes are weighted higher than otherwise. Suppose we have two outcomes c1 and c2 with equal probabilities 12 each. The weighted probabilities are then p^ (z1 ) = p^ (z2 ) = 1 c 2 1 1 c 2 1 + 1 c 2 1 1 c 2 2 1 c 2 1 + 1 c 2 1 = c1 c1 + c1 = c2 c1 + c1 Consequently, = c1 c2 c1 + c c1 + c1 c1 + c1 2 For any constant k; indi¤erence curves then satisfy k= c + c1 + + 2 c1 + c2 c1 + c2 Plotting one together with standard utility we …nd these preferences produce a higher level of riskaversion. 16 3 2.8 2.6 2.4 2.2 2 1.8 1.6 1.4 1.2 1 0.8 0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 3 We can also have Faruk Gul’s disappointment aversion. Then ( 1 (1 + ) c + if c (z) < M (c; ) = 1 c + else This means that all outcomes worse than the certainty equivalence get scaled up. = X X p (z) (1 + I [c (z) < ]) z2 p (z) (1 + I [c (z) < ]) = z2 = 1 P The weighted probability here is X p (z) (1 + I [c (z) < ]) c (z) z2 z2 p (z) (1 + I [c (z) < ]) c (z) z2 p (z) (1 + I [c (z) < ]) P p (z) (1 + I [c (z) < ]) z2 p (z) ((1 + I [c (z) < ])) P c (z) 17 1 + For any constant k: indi¤erence curves can be written as 8 1 1 < (1+ )12 c11 + 2 c21 1 if c1 < c2 (1+ ) 2 + 2 (1+ ) 2 + 2 k= 1 c (1+ ) 12 c2 1 2 : if c1 > c2 1 1 + (1+ ) + (1+ ) 1 + 1 2 2 2 2 2 1.8 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 Here, it is important to notice the kink around unity. This is …rst order risk-aversion. Individuals will be more averse to small risks than under standard preferences. Risk premia under Chew-Dekel Preferences. De…ne the risk-premium R from R = Ec (z) (fc (z)g) Let us consider two state case above. In the two cases of CRRA E utility and weighted utility. Suppose c = c1 = 1 with prob 21 and c = c2 = 1 + with prob. 12 : The shock is thus and is the standard deviation of c: 18 1.8 2 Rs tan d (1 ) + (1 + ) = 1 2 1 Rs (0) + Rs0 (0) + Rs00 (0) 2 2 1 (1 ) +(1+ ) 2 6@ 1 = 6 4 @ 2 (1 @2 1 16 + 6 24 1 (1 = 2 ) 1 ) +(1+ ) 2 Rs ( ) 2 3 7 7 5 =0 1 7 7 5 2 @ 3 2 2 =0 In the weighted expected utility, we have Rwheigt = (1 1 ) (1 + + (1 + ) ) + (1 + ) 1 Rw (0) + Rw0 (0) + Rw0 (0) 2 2 + (1 ) +(1+ ) + (1 ) +(1+ ) 6@ 1 = 6 4 @ 2 16 + 6 24 1 = (1 2 @ 2 (1 1 ) (1 @ 2 ) 2 + ! Rw ( ) 2 +(1+ ) ) +(1+ ) 2 1 + 1 + 3 7 7 5 =0 1 3 7 7 5 2 =0 In the disappointment aversion case, the …rst derivative is not going to 19 be zero, people are not locally riskneutral. Therefore, Rdisapp = (1 + ) (1 1 1 Rd (0) + Rd0 (0) + Rd00 (0) 2 2 (1+ )(1 ) +(1+ ) 2+ 6@ 1 = 6 4 @ 2 = 16 + 6 24 2+ @ 2 1 (1+ )(1 @ + 2 (1 1 ) + (1 + ) 2+ Rd ( ) 2 1 ) +(1+ ) 2+ 2 1+ ) (2 + )2 ! 2 3 7 7 5 =0 1 3 7 7 5 2 =0 The key here is the linear term. 2.3 Time and Risk Let us now combine the aggregation over time and over states so we can de…ne preferences over both states and time, allowing for rates of substitution to di¤er depending on whether we are talking about states or time periods. Generalize the notation so that tc denotes a stochastic consumption stream starting from period t; i.e., ct ; c (zt+1 ) ; :::; and denote U (t c) Ut The aggregator can be written Ut = V (ct ; (Ut+1 )) or by imposing future independence Ut = V (u (ct ) ; (Ut+1 )) (2) Note that we are now aggregating the direct utility of ct and the certainty equivalent of the stochastic continuation payo¤ from t + 1 and onwards. 20 Thus, since at t; Ut+1 is stochastic (depending on the realization of zt+1 ), we construct the period t + 1 certainty equivalent of this stochastic utility, i.e., (Ut+1 ). If we use a standard expected utility certainty equivalent in (2), we get what is typically called Kreps-Porteus utility (some times also Epstein-Zin, although they often are associated with more general speci…cation of state aggregation, a la Chew-Dekel). The key here is that we can now separate risk-aversion, which is determined by the properties of ; from intertemporal substitution, determined by V: Let us use the CRRA speci…cation for ; so (U ) = [EU ] 1 and a constant elasticity aggregator V (u (ct ) ; (Ut+1 )) = [(1 1 )u + (U ) ] : If we let u be linear, u = c; so we let all curvature be taken care of by and V; we can now interpret 1 a s the degree of risk-aversion and 1 1 as the elasticity of intertemporal substitution. To see this simply, look at a two period example where consumption is stochastic in the second period. In the second period, stochastic utility is U2 = c (z2 ) Suppose …rst second period consumption is non-stochastic at c2 : Then, 1 (U2 ) = [EU ] = c2 Then, U1 = [(1 ) c1 + c 2 ] 1 MRS is now, @U1 @U1 = @c2 @c1 M RSc1 ;c2 1 = = 1 1 ((1 ((1 ) c1 + c 2 ) ) c1 + c 2 ) c2 c1 21 1 1 1 1 1 (1 c2 1 ) c1 1 as usual and c1 c2 @M RSc1 ;c2 c1 c2 @ M RS 1 = IES c1 c2 =1 : Note that we have imposed that U is linearly homogeneous by having the power 1 ; which is with no consequence. Consider U~1 = (1 ) c1 + c 2 and calculate @ U~1 @ U~1 = = @c2 @c1 1 c2 (1 ) c1 1 = c2 c1 1 1 It should be clear from the aggregator (c2 (z2 )) = [Ec (z2 ) ] 1 that we can think of 1 as measuring risk aversion. A key characteristic of Kreps-Porteus preferences is that they give rise to preference for when risk is revealed. A basic intuition is that early revelation implies that some risk is converted into intertemporal substitution. To see an example of this, consider the example in the introduction, where second period consumption is high or low (z2 = zh or zl ): Suppose that this is reveled in the …rst period. Denote U~1 (z) …rst period utility given a realized value of z: Then, U~1 (zl ) = [(1 U~1 (zh ) = [(1 ) c1 + c l ] ) c1 + c h ] 1 1 and U1early = (U1 ) = 1 ((1 2 = 1 ((1 2 ) c1 + c h ) 1 1 + ((1 2 ) c1 + c h ) Compare this to the late resolution case. Then, (U2 ) = 22 ) c1 + c h ) 1 1 ) c1 + ch ) + ((1 2 ch + cl 2 1 1 1 and U1late = ch + cl 2 ) c1 + (1 !1 Comparing these, we see U1early U1late 2 1 1 ((1 ) c1 + ch ) + ((1 = 4 2 2 1 ) c1 + c l ) (1 ) c1 + 0.0004 0.0002 0.2 0.4 0.6 0.8 1 -0.0002 -0.0004 -0.0006 Here I plotted the di¤erence for = 12 and for < ; we see that early resolution is preferred. Note that when < , risk aversion (1 ) is larger than the inverse of intertemporal elasticity of substitution, so in a sense, it is less costly to substitute between time periods than between states of nature. Is this important? With Kreps-Porteus preferences we can clearly get low interest rates if individuals are facing little risk and high if they are facing large. But can we get what we want when we have a representative household that prices both risky and non-risky assets while having to bear reasonable amounts of risk himself. An example. Look at the case in the introduction; given c1 c2 = ch = c1 (1 + g + ) with prob cl = c1 (1 + g ) with prob 23 5 c1 =1 0.0006 0 ch + cl 2 !1 3 1 2 1 2 First, we note that Arrow-Debreu prices now have the property that the price of an asset that pays out in state z = zh ; depends on consumption also in the other state. This is clear from the expression of utility: U1 = (1 !1 ch + cl 2 ) c1 + where we see that the marginal contribution to utility U1 of an asset that pays out in one of the states, say in z = zh depends on consumption also in the other state. The simple calculation of Arrow-Debreu securities is lost. What is now the price of a safe bond pb and a risky asset (the apple tree) pr that pays dividends (1 + g + ) or (1 + g ) per share depending on which state occurs? The prices have to be such that if a consumer buys one marginal unit of the assets, utility is unchanged. Take the safe bond, it has to satisfy 0 = pb (1 0 = pb (1 2 6 pb = 4 = 2 (1 @U1 @U1 + E @c1 @c2 0 ch +cl 2 @c1 B@ ) + @ @c1 @ch ) ch +cl 2 ch 2 (1 ) c1 1 + cl 1 C A @cl 3 7 5 ch =c1 (1+g+ );cl =c1 (1+g ) (1 + g + ) + (1 + g 2 With the risky asset we have 0 @c B@ 0 = pr (1 ) 1+ @ @c1 ) pr = 1 @ + 1 ch +cl 2 2 (1 ) ch +cl 2 ) (1 + g + ) (1 + g + ) @ + @ch (1 + g + ) + (1 + g 2 24 ch +cl 2 1 + (1 + g (1 + g @cl ) ) ) ) 1 1 1 )C A ((1 + g + ) + (1 + g Note that the ratio pb (1 + g + ) 1 + (1 + g = pr (1 + g + ) + (1 + g ) ) ) is determined by risk aversion and exactly the same as in the standard case. Therefore it seems as we have got little help from Kreps and Porteus. Not entirely right. We have learnt that it is di¢ cult to get the risk premia by assuming that stock market returns are fully determined by consumption growth. Maybe this is not the way to go. Suppose there is some other more stable income so that the share dividend is more volatile than consumption. Say that the standard deviation of consumption remains at while the standard deviation of dividends is x : Then (1 + g + ) 1 + (1 + g pb = pr (1 + g + ) 1 (1 + g + x ) + (1 + g Setting g = 0:018; ) ) 1 1 (1 + g x) = 0:036 and x = 4:5; we get the following graph. 1.14 1.12 1.1 1.08 1.06 1.04 1.02 1 -50 -40 Targeting yields = h -30 (1+g+ ) 1 +(1+g (1+g+ ) 1 (1+g+x )+(1+g ) ) -20 1 1 (1+g 11:1; i.e., a CRRA coe¢ cient of 10. 25 i 0.98 0 -10 x) g=0:018; =0:036;x=4:5 = 1:05 To also get the riskfree rate right, we need pb = 1 1:01 with a reasonable value of : Setting = :99 ) 1 and using " 2 (1 ) = = 0:497 49 11:1; we solve (1 + g + ) + (1 + g 2 ) (1 + g + ) 1 + (1 + g ) 1 # g=0:018; =0:036; = 1 = we …nd = 0:34; so the intertemporal elasticity of substitution is 1 0:34 1:515: This is quite far away from the EU case where we would have IES= 1 ( 111:1) = 0:08: Producing an interest rate of 0" # (1 + g + ) + (1 + g ) 1 1 @ (1 + g + ) + (1 + g ) 2 (1 ) 2 which equals 13:8%: Recent paper by Bansal and Yaron use Kreps-Porteus speci…cation with CRRA=10 and IES 1.5, i.e., a large deviation from standard EU. They use data to show that dividends are more volatile than consumption, 4.5 times higher. As you can see, these assumptions seems able to account for the risk-premia. 2.4 Closed form value function in partial equilibrium Consider an individual with no labor income who has access to a market ~ His budget for investments with an exogeneous stochastic i.i.d. return R: constraint is thus ~ t+1 At+1 = (At ct ) R We can now write a Bellman equation W (At ) = max V u (ct ) ; ct W (At where as above u (c) = c (W ) = [EW ] 26 1 ~ t+1 ct ) R g=0:018; =0:036; and V (u; ) = [(1 1 )u + (3) ] : Conjecture that W (A) = 1A and c= for yet undetermined coe¢ cients Using this, we …nd that 2A and 1 2: ~ t+1 = ct ) R W (At+1 ) = W (At 1 ~ t+1 = ct ) R (At 1 ~ (1 2 ) At Rt+1 1 (W (At+1 )) = E 1 ~ 2 ) At Rt+1 (1 1 = 1 (1 ~ t+1 E R 2 ) At Using this in the Bellman equation; 1 1 1 At = (1 )( 2 At ) + 1 (1 2 ) At ~ t+1 E R 1 = (1 )( 2) + (1 1 2) 1 ~ t+1 E R At which is satis…ed, provided 1 1 = (1 )( 2) + 1 (1 2) ~ t+1 E R 1 : The …rst-order condition for ct is the same as choosing 2 (4) optimally, 1 @ (1 )( 2) + ~ 2 ) E Rt+1 1 (1 @ =0 (5) 2 Solving these equations gives us the solution to the problem. Assuming ~ makes it possible to calculate a speci…c form for the distribution of R; 1 ~ ~ E R : If, for example, if ln R r is normal with mean m and standard deviation ; i.e., ln R s n (R; m; ) we have ~ E R = = e 1 2 2 2 m+ 1 ~ E R Z1 1 p 2 = em+ 2 2 27 ; (er ) e (r m)2 2 2 dr Note here that increasing has a direct e¤ect on the expected return, so it is not a mean preserving spread, i.e., 2 ~ = em+ 2 : E R We can de…ne a mean-preserving spread as increasing 2 m ; then 2 ~ =e E R 2 2 m + 2 2 by letting m = : 2 That is, if for any , the mean of r is m 2 ; and the standard deviation ; an increase in ; is a mean preserving spread. In this case, 1 ~ E R = em 2 2 + 2 2 2 (1 2 = em ) Clearly, we here see that an increase in ; while keeping the mean of the return constant, reduces the certainty equivalent, since 1 and more so the smaller is : 1 ~ ; the choice of 2 solves (5), implying that 2 Denoting R E R is a root of 2 1 1 2 1 1 2 (1 (1 )+ ( )+ ( 1 1 (1 (1 2 ) R) = 0: 0 2 ) R) = 0 In the simplest case of unitary intertemporal elasticity of substitution, i.e., = 0; the solution is 2 = 1 : We should then substitute our optimized value of 2 back into the Bellman equation and …nd 2 in (4). Note, however, that the formulation in (3) is not valid if = 0:We can however, look at the limit as ! 0; V (u; ) = lim ((1 !0 )u + ) 1 = u1 Using this is formalution plus ut = ct = 2 At and (W (At+1 )) = in the Bellman equation gives 1 (1 2 ) At R with 2 = 1 1 At 1 = = = = = u1t t ((1 ) At )1 ( 1 At R) At ((1 ))1 ( 1 R) At exp ((1 ) ln (1 ) + ln 1 )) ( R) 1 ((1 = (1 ) ( R) 1 28 > 0: 1 R) ~ is log-normal In which case we conclude that if R W (A) = (1 2 em+ ) 1 2 A and c = (1 ) A: In this case, high riskaversion or higher risk reduces welfare by reducing the certainty equivalent return (high riskaversion means low ) but saving is una¤ected. What happens if > 0? First we note that higher risk reduces R and therefore tends to reduce 1 ; then, in (??) a lower R and lower 1 reduces (increases) the term ( 1 (1 > (<) 0. In optimum, 2 ) R) if this term must equal 1 (1 ) 2 1 2 which is downward sloping since d 2 1 1 2 d (1 ) <1 = (1 ) 1 1 + 2 2 < 0: 2 1.2 1 0.8 0.6 0.4 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Therefore, an increase in risk leads to a higher (lower) 2 if > (<) 0. I.e., consumption increases and savings decreases if the certainty equivalent return decreases i¤ the intertemporal elasticity of substitution is higher than unity. 29 3 Ambiguity Aversion to unknown odds, as is demonstrated in labs, e.g., the Ellsberg paradox, is given axiomatic foundations by Gilboa and Schmeidler. They show that reasonable axioms capturing such ambiguity averse behavior can be represented by a sort of max min preferences. Suppose it is known that there is a set of possible realizations of the state, z 2 with associated consumption levels c (z) : Then, assume that the individual does not know the probability distribution over these states, but he knows that this probability distribution belongs to a set : Call the elements of this set p being particular possible probability distributions. An element p is thus a vector of probabilities p (z) : Then, preferences are given by X U (fc (z)g) = min p (z) c (z) = min Ep c (z) : p2 p2 z2 As with standard preferences, it is not necessary to take this literally in the sense the individuals actually maximize these minimized preferences. It may, for example, not be possible to ask individuals to tell us directly about : A simple example, consider again the Ellsberg paradox. There are two urns each with 100 balls. In urn 1, there are 50 red and 50 black. In urn 2, there are only red and black balls but the proportions are unknown. The subject is given a color and can pick one ball. If a ball with the given color comes up, the gain is 50$, if not the gain is zero. The subject is asked to rank lotteries. Typically the following response comes up. 1. Red from urn 1 Black from urn 1. 2. Red from urn 2 Black from urn 2. 3. Red from urn 1 Red from urn 2. 4. Black from urn 1 Black from urn 2. Suppose we now ask individuals about the "ambiguity premium", i.e., we …nd value r; such that if we let urn 1 contain 50 r red balls and 50 + r black ones. Red from urn 1 Red from urn 2. and hopefully if we let urn 1 contain 50 + r red balls and 50 ones Black from urn 1 Black from urn 2. 30 r black In this case, the relevant p0 s are the shares of red balls in urn 2. The : value of r; pins down ;being the interval 50100r ; 50+r 100 Given this, the value of urn 2 in a bet on red is min r 50+r p2[ 50 ; 100 ] 100 = pU (50) + (1 p) U (0) 50 + r 50 r U (50) + U (0) 100 100 with the same value in a bet on black since in this case, we have min r 50+r p2[ 50 ; 100 ] 100 pU (0) + (1 p) U (50) 50 + r 50 r U (0) + U (0): 100 100 Suppose now that there are two individuals, one owns an asset that gives 50 dollars with a probability p he knows is 50% but that he cannot verify this knowledge to the other person. Otherwise it pays zero. The other person is in the same position, he owns an asset that gives 50 dollars with probability of p = :5 that he can not credibly verify. Suppose individuals consume 1+the payo¤ from the asset and have log utility. Under standard assumption including that both individuals assign a 50% probability to the other’s urn. Individuals should share the risk and get a payo¤ of 1 1 1 ln (51) + ln (26) + ln 1 = 2:612 4 2 4 If instead individuals have ambiguity aversion the probability of winning in the "foreign" asset is p0 2 : How big share ! should he choose to invest in the other persons asset. Given a p 2 ; the four states of the world, fhigh; highg ; flow; highg ; fhigh; lowg ; flow; lowg happen with probabilities, 12 p; 21 p; 12 (1 p) ;and 12 (1 p) : Given the symmetric nature of the game, we focus on the case when the price of the "foreign" asset in terms of the "domestic" is unity. Consumption, given ! is then 51; :1 + !50; 1 + (1 !) 50; 1: The utility of the optimal portfolio is therefore min max p2P ! 1 1 1 p ln (1 + 50) + p ln (1 + !50) + (1 2 2 2 p) ln (1 + (1 !) 50) + The …rst-order condition for the maximization problem is d 1 p ln (1 2 + 51) + 12 p ln (1 + !50) + 21 (1 d! 31 p) ln (1 + (1 !) 50) 1 (1 2 p) ln 1 52p 1 50 Of course, this is increasing in p: Note also that for p = 1=52; ! = 0 and for p < 1=52; ! < 0: For example, if p = 0:01; ! 0:01;i.e., a short position in the foreign asset. The short position implies that if the foreign asset pays 50, a payment from hom to abroad takes place. Now, we have to pick p:To do this, we consider the maximized value, i.e., != 1 1 1 p ln (1 + 51) + p ln (1 + !50) + (1 2 2 2 max ! = 1 (p ln (51) + ln (52) + p ln (p) + (1 2 p) ln (1 p) ln (1 + (1 !) 50) p)) Let’s plot this, what we see at the this is increasing in p for p close to 0.5. But, it is not monotone. 2.6 2.5 2.4 2.3 2.2 2.1 2 0 0.1 0.2 0.3 p 32 0.4 0.5 2.01 2 1.99 1.98 1.97 0 0.02 0.04 p 0.06 0.08 0.1 . In fact, arg min p 1 (p ln (51) + ln (52) + p ln (p) + (1 2 p) ln (1 p)) = 1 52 which corresponds to a share ! = 0: Why is this? The answer is that if the individual would be short in the foreign asset, what is a bad realization has ‡ipped. It is now the state when the foreign asset pays 50, happening with probability p: Therefore, in the range where ! < 0; a higher p means lower utility. In general, since trade can only increase utility, the worst p; is the one that implies autarky. The situation would be quite di¤erent under asymmetric information, in which case we would expect to sometimes see agents with more information (domestic) go short in home securities. Let’s take another example, suppose production in each country is 2 or 1, so the state space is f1; 2; 3; 4g ; implying output fy1 ; y2 g is f2; 2g ; f2; 1g ; f1; 2g or f1; 1g : Utility is X Ui = min p (z) ln ci (z) p2 Suppose also individual the knows that the probability of there own production being high is 0.5, while the set of possible probabilities for the other countries production being high is 0:5 2 ; for 2 [ a; a] and that these events are independent. The agent decides how much to invest abroad !: Due to symmetry, the relative price of the two assets, foreign and domestic production should be 33 one. The budget constraint is therefore for agent 1. c1 (1) = 2; c1 (2) = (1 !) 2 + !; c1 (3) = (1 !) + 2!; c1 (4) = 1 Given ; the maximization problem is X max p (z) ln ci (z) ! 1 1 1 1 2 ln 2 + + 2 ln ((1 !) 2 + !) ! 2 2 2 2 1 1 1 1 + 2 ln ((1 !) + 2!) + + 2 ln 1 2 2 2 2 = max The …rst order condition is d 1 2 1 2 +2 ln ((1 1 2 !) 2 + !) + d! 1 2 2 ln ((1 !) + 2!) = 0 1 2 ! = Substituting this into the utility function, yields X max p (z) ln c 6 ! = = 1 2 1 2 2 ln 2 + 1 1 ln 2 + ln 3 4 2 1 2 1 +2 2 ln ((1 1 + 4 (ln 2) + !) 2 + !) + ln (1 + 4 ) + 1 2 1 2 1 4 2 ln ((1 != 12 6 ln (1 4 ): Now, this the …rst order condition for minimizing this is d 1 4 ln 2 + 12 ln 3 (ln 2) + 1 4 1 giving min = 12 : This means that in particular that ! min = 1 2 6 + d ln (1 + 4 ) + 1 4 ln (1 will never be chosen larger than min : 34 = 1 2 6 = 0: 12 !) + 2!) 4 ) 1 12 and 0.65 0.6 0.55 0.5 0.45 0.4 0.35 -0.2 -0.1 0 0.1 0.2 This is implies the important result that shortsales cannot occur. This is in line with empirical evidence and is not the prediction of models with asymmetric information. Why is this? Note …rst that if ! = 0; the probability of success has no e¤ect on utility. If ! is negative, a reduction in p actually increases welfare. Why, the stream of payment from the foreign asset is ! (2p + (1 p)) = !(1 + p); which decreases in p if ! is negative. In this case, high production abroad is the bad state! 35 4 Time-inconsistency and temptation Lab evidence discussed in the introduction shows preference reversal, quicker discounting in for close time periods than for distant. Also preference for commitment. People sometimes prefer to restrict their future behavior – force themselves to save, hide the jar of cookies, not bring to much money to the bar, and so on. Evidence that hyperbolic discount factor represents time preference better than geometric. Two approaches: Quasi geometric preferences. Preferences over sets (menus), allows modelling of temptation, cost of employing self control and welfare analysis. 4.1 Quasi-geometric preferences Between current and next period, an "extra" discount factor ; is introduced (the model). Self 0 U0 (0 c) = u (c0 ) + Self 1 Self 2 u (c1 ) + 2 u (c2 ) + 3 2 u (c3 ) + U (1 c) = u (c1 ) + u (c2 ) + u (c3 ) + U (2 c) = u (c2 ) + ( u (c3 ) + 3 u (c3 ) + ) This implies that preferences are changing over time. 2 0 u0 (c2 ) u (c2 ) = u0 (c1 ) u0 (c1 ) u0 (c2 ) u0 (c1 ) Self 0 M RS1;2 = Self 1 M RS1;2 = Self 1 cares relatively less of period 2 utility than self 1. Assumptions about behavior: The consumer cannot commit to future actions. The consumer is “sophisticated”: he realizes that his preferences will change and makes the current decision taking this into account. The decision-making process is viewed as a dynamic game, with the agent’s current and future selves as players. (Alternative: “naive” behavior. The agent thinks that he will not change preferences.) 36 Focus is on Markov equilibria, but other equilibria with trigger strategies also exists. For example, a self "behaves" and does not overconsume as long as previous selves has behaved well. Markov equilibria can be strange or non-existent. Standard existence theorems not applicable. Example: Consumption and savings problem. Suppose the agent has log period utility Ut (t c) = ln (ct ) + 1 X s ln (ct+s ) s=1 and face a constant return r. The budget constraint is at+1 = r (at ct ) The state variable is at and let us implicitly de…ne a continuation value from J (at ) = ln ct + J (r (at ct )) ; (6) for some value ct : If = 1; J is also the value function if ct is the argmax to the RHS (the whole equation is then the Bellman equation). Under quasi geometric discounting, ct is NOT arg maxc ln ct + J (r (at ct )) instead ct = arg max ln ct + ct J (r (at ct )) J (r (at ct )) and the utility of self t is W (at ) = max ln ct + ct Note that the value of giving assets to the next self is depreciated by the fact that self 0 knows that self 1 is going to overconsume in the eyes of self 0. Now, we can guess that J has the form J (at ) = A + b ln at Given this, ct is the solution to the …rst order condition 0 = = 1 ct 1 ct J 0 (r (at b r (at at ) ct = 1+ b 37 ct ) ct )) r r Substituting this in (6) gives A + b ln at = ln at + 1+ b at 1+ b A + b ln r at = (1 + b) ln at + A + b ln r b 1+ b ln (1 + b) This is satis…ed for all at iif (1 + b) = b 1 ) b= 1 implying ct = at = 1 1+ 1 1 1 (1 + ln 1 ) at and A is A= ln 1 1 r (1 ) 1 (1 ) : 1 As we see, if < 1; consumption is higher and savings are lower than in the time-consistent case, when the consumption rate is 1 : Let’s now …nd the commitment solution if self 0 determines all consumption values. In this case, we …rst calculate Jc (at ) ;which is the continuation value when everything is determined by self 0. Note, however, that future selves will agree with self zero on this. The di¤erence between the nocommitment case is that now the continuation value maximizes the standard Bellman equation without any : Thus, Jc must satisfy, Jc (at ) = max ln ct + Jc (r (at ct ct )) If we don’t remember the solution to this, we do as usual. We take the …rst order condition 1 = rJ 0 (r (at ct )) : ct Guessing Jc (at ) = Ac + bc ln at implies 1 = ct ct bc r r (at at = : 1 + bc 38 1 ct ) Substituting, 1 at + Ac + bc ln rat 1 1 + bc 1 + bc 1 = ln at + ln + Ac + bc ln at + bc ln r 1 + bc Ac + bc ln at = ln Ac + bc ln at bc 1 + bc Giving ) bc = (1 + bc ) ; bc = ct = at = (1 1 + bc Ac = ln 1 1+ = ln (1 ) Ac = 1 1 1 ) at + Ac + ln (1 )+ 1 1+ 1 1 ! ln r 1 1 ln r 1 1 ) + Ac + 1 1 ln r 1 As we see, the coe¢ cient on ln at+1 is the same in both cases, commitment and no commitment. The di¤erence between the constants under no commitment and commitment is negative, I think. The fact that the coe¢ cient on ln at+1 is the same in Jc and J; implies that the marginal value of leaving assets to self 1 is the same in both cases. Thus, consumption in period 0 is independent of whether there is commitment or not. Note that two forces here are a¤ecting the results. On the one hand, giving assets to self 1 under no commitment has a lower value since he consumers too much in the eyes of self 0. This reduces the incentive to save for self 0. On the other hand, this leaves self 2, 3,... with too little consumption and the way only way self 0 can increase consumption of self 2,3,... is to save. This increases the value of saving. Apparently, this two e¤ects cancel in the log utility case. From period 1 and onwards, savings is higher under commitment. > 1 (1 ) Commitment would of course increase welfare for self 0; it can never reduce it. What about later selves? 39 In the no commitment case, self 1 gets Wnc = ln ct+1 + J r at+1 ct+1 1 ln at+1 1 (1 ) 1 ln rat+1 + A+ 1 1 (1 ) 1 (1 ) 1 = ln at+1 + ln + 1 1 (1 ) A+ 1 1 ln r 1 With commitment, the continuation value is di¤erent since now self 0 determines everything. Self 1 gets under commitment ln (1 = ln (1 1 = ) at+1 + ) at+1 + (1 ) 1 Jc (r at+1 ) bc ln (r at+1 ) + ln at+1 + ln (1 Ac )+ 1 ln r + Ac Commitment gives extra value which is good also for self 1, but she cannot control her consumption which reduces the value. For self 1, commitment therefore be better than no commitment, also if it is done by self 1. For later individuals, it may be even better with previous commitment since asset levels are higher. 4.2 Preferences over choice sets, Gul and Pesendorfer An alternative approach. Does not assume multiple selves, no game thus no multiplicity. Also allows resistance to temptation and to model costs of resisting temptation. Two subperiods. Second subperiod preferences de…ned over ordered pairs (A; x), where A is a choice set and x 2 A is a choice (consumed). De…nition: y tempts x if (fxg; x) is preferred to (fx; yg; x). That is, individuals are better of getting x without having y in the choice set. Assumptions: 1. Eliminating temptations cannot make the consumer worse o¤. 40 (1 ) 2. If y tempts x, then x does not tempt y. 3. The utility of a …xed choice is a¤ected by the choice set only through its most tempting element. Second-period preferences induce …rst-period preferences over choice sets themselves: A B if and only if there is an x 2 A such that (A; x) is preferred to (B; y) for all y 2 B. The above assumptions imply what is labelled set betweenness: A B)A A[B B: Choice sets cannot be compared simply by looking at their "best" or chosen elements. Instead, the utility of a …xed choice depends on the choice set (through its most “tempting”element). Note that this violates one of the axioms in standard theory. Removing a non-choosen element from a choice set cannot change utility or behavior (independence of irrelevant alternatives). Set betweenness allows three possibilities: 1. Standard decision maker: A A[B B. 2. Preference for commitment and self-control: A A [ B B. Interpretation: there is an element in B that tempts me. Nevertheless, I choose the same element in A and A [ B; but if faced with only A; I don’t have to take the e¤ort of controlling myself. Thus A A [ B: Furthermore, A [ B B since the choice in A [ B provides higher utility,than the tempting choice. 3. Preference for commitment and succumbing to temptation: A A [ B B. Interpretation: there is an element in B that tempts me. A A [ B since it provides higher utility. Faced with the tempting choice, however. I cannot resist. I choose the same element in A [ B and B and there is no cost of controlling myself:Thus, A[B B 4.3 The representation theorem The assumptions implies that preference over sets in the …rst period can be written W (A) = maxx2A fU (x) + V (x)g maxx~2A V (~ x) 41 Second period, preference are represented by W (A; x) = fU (x) + V (x)g maxx~2A V (~ x) Interpretation: U determines the commitment ranking (i.e., the utility of singleton sets, no temptation). V determines the temptation ranking (i.e., V gives higher values to more tempting elements). arg maxx~2A V (~ x) is the most tempting element in A. The second-period choice chooses x by solving maxfU (x) + V (x)g x maxx~2A V (~ x) If a person is given x 2 A, without anything else to choose from, there is no cost of self control. The utility is U (x) + V (x) V (x) = U (x) : If a person chooses x 2 A; the disutility of self-control is V (x) maxx~2A V (~ x) 0;so utility is U (x) + V (x) maxx~2A V (~ x): If a person chooses x~ = arg maxx~2A V (~ x), he gives in to temptation, there is no cost of self-control, and the utility is U (~ x) + maxx~2A V (x) 4.4 maxx~2A V (~ x) = U (~ x): A 2-period consumption-savings model Consumption today and tomorrow. Neoclassical production. Standard budget set (borrowing and lending at r). General equilibrium. 42 With U (c1 ; c2 ) playing the role of U and V (c1 ; c2 ) the role of V , let the temptation function V have a stronger preference for present consumption. For example, let U (c1 ; c2 ) = u(c1 ) + u(c2 ) and V (c1 ; c2 ) = with ; (u(c1 ) + u(c2 )) ; < 1. Aggregate resource constraint given by c1 + k2 = f (k1 ) c2 = f (k2 ) Strength of temptation determined by : Standard model when As ! 1; Laibson model. = 0: In equilibrium choices are made to maximize U (c1 ; c2 ) + V (c1 ; c2 ) In competitive general equilibrium individuals take prices (here interest rate) as given, provides a linear budget set for the individual. 43 c2 Competitive equilibrium U(c1,c2) U(c1,c2)+V(c1,c2) V(c1,c2) c1 Temptation Best without temptation Actual choice Policy implications. Command optimum can achieve arg max U , without any temptation cost. Nothing is better than this. A subsidy to investments (tax on …rst period consumption) can improve upon laissez faire. Does so by reducing temptation. For example, let u (x) = ln (x). Then, choices are governed by solving max (ln c1 ) + ln(c2 )) + (ln(c1 ) + s.t. c1 + c2 1+r ln(c2 )) = w c1 = arg max ((1 + ) ln c1 + (1 + 1+ = w 1 + + (1 + ) ) ln((1 + r) (w which increases in : The maximum temptation is ct1 = arg max (ln(c1 ) + w = 1+ 44 ln((1 + r) (w c1 ))) c1 ))) Interesting implication Compare autarky, i.e., each individual runs his own machine. Then the interest rate is not exogenous. Autarky c2 U(c1,c2)+V(c1,c2) V(c1,c2) Actual choice Temptation in comp. eq. Temptation under autarky c1 Result: Autarky delivers the same allocation, but at higher welfare. Why? The choice sets shrinks, the temptation to overconsume is reduced and the cost of resisting temptation falls. 4.5 Macroeconomic applications Krusell, Kurus˧çu, Smith "Temptation and Taxation" Consider long horizons: the limit of the …nite-horizon problems. Study competitive equilibrium under two kinds of parametric restrictions: 1. Logarithmic utility, Cobb-Douglas production, and full depreciation: full analytical solution of recursive competitive equilibria. 2. Iso-elastic utility and no restrictions on technology: analytical characterization of steady state and computational analysis of dynamics. 45 Analysis: Vary and interest rate constant. Results: , while adjusting to keep the steady-state Almost observationally equivalence (like in Barro "Laibson meets Ramsey", when the utility is log, the speed of adjustment to the steady state does not depend on (as in Barro). With more (less) curvature in utility, the speed of adjustment is decreasing (increasing) in . The e¤ects of on the speed of adjustment is quantitatively small: observational equivalence found in Barro “almost”carries over. Savings should be subsidized. But not much, and the welfare gains are small (little reduction in temptation costs). 46 5 Habits 6 Topic 4. Habits We have previously assumed history independence, meant to mean that the marginal rate of substitution, evaluated at t between: goods consumed at t and t + 1; and di¤erent goods consumed in the same time is independent of the consumption history prior to t: This is not necessarily a good assumption. There are, in particular two, cases in which we might want to relax history independence. 1. The case when previous consumption leads to higher aspiration. To live on a small budget may be easier if you are used to it than if your are used to the good life. 2. When relative consumption matters for utility. "Catching up with the Joneses or "Poverty is more easily accepted if it is shared by everyone" Ernst Wigfors, Social Democratic Finance Minister 1932-49. 3. The relative taste for some goods is a¤ected by previous consumption of them, e.g., food, culture goods and drugs. In the literature, the …rst case have been used to try to explain asset market puzzles, i.e., why standard models have great problems explaining the co-movements of prices and consumption. The third case is used to explain, for example, cultural diversity. To simplify, in particular in order to be able to specify recursive preferences, utility is assumed to be Ut = 1 X j u (ct+j ; t+j ) j=0 where t = v (~ ct 1 ; c~t 2 ; :::~ ct n) ; is denoted the habit. A couple of things to note, 1. We maintain time additivity here, although this should be straightforward to generalize to the constant elasticity aggregator. 47 2. c~t s can denote the own previous consumption of the household, the consumption of some reference group or some combination. If c~t s = ct s (own consumption), we have what is called "internal habits" while if it is the consumption of some reference group, it is called "external" habit. Abel uses a geometric average t 1 cD t 1 Ct D 1 ; where Ct 1 is aggregate consumption. Here = 0 gives the standard model, 6= 0; and D = 1;gives the internal habit case and D = 0 the external case. 3. History matters only though the habit function, i.e., through v (~ ct 1 ; c~t 2 ; :::~ ct n) 4. We assume either a …nite value of n or at least that @v (~ ct 1 ; c~t 2 ; :::~ ct n ) lim = 0: n!1 @~ ct n so that we can hope to …nd stationary decision rules. 6.1 Optimal consumption under external vs. internal habits. Suppose the representative agent solves max 1 X t u (ct ; t) t=0 s.t. At+1 = (At ct ) r 1 D vt = cD t 1 Ct 1 and a no-Ponzi condition. Let’s look at the Bellman equation. Note that now, vt is a state variable. Therefore V (At ; vt ) = max u (ct ; ct s.t. At+1 = (At t) + V (At+1 ; vt+1 ) 1 ct ) r; vt = cD t 1 Ct D 1 The …rst order condition is u1 (ct ; vt ) = V1 (At+1 ; vt+1 ) r = V1 (At+1 ; vt+1 ) r 48 V2 (At+1 ; vt+1 ) D @vt+1 @ct vt+1 V2 (At+1 ; vt+1 ) ct Clearly, the …rst order condition is not a¤ected if D = 0: Suppose instead D > 0: Then an increase in today’s consumption increases the habit. Suppose this reduces utility, then this implies that there is a negative dynamic e¤ect of consumption which will show up in a negative V2 : Therefore, the V2 (At+1 ; vt+1 ) is positive and marginal utility of consumpterm D vt+1 ct tion should be set higher in period t: Consider the case of external habits, D = 0 . We have seen that in this case, the FOC is the same as under no habits. Suppose …rst for simplicity that u (ct ; t ) = ln ct ln vt Recall the solution strategy in the case when we expect that there can be an analytical solution. 1. Write Bellman equation. 2. Guess a functional form of the value function with unknown parameters. From the Bellman equation we see that it has to be of the same functional form as the per-period utility function. 3. Solve for the choice variable that maximizes the RHS of the Bellman equation given our guess on the value function. 4. Substitute your optimal choice variable into the RHS of the Bellman equation to express the maximized RHS. 5. Verify that the Bellman equation is satis…ed for all values of the state variables by …nding the unknown parameters. If step 5 fails you have made an incorrect guess and must start with another. However, most problems do not admit closed form solutions for the value function in which case this approach is useless. Now, we guess that the value function is V (At ; vt ) = B1 ln At + B2 ln vt + B3 for the unknown coe¢ cients B1 ; B2 and B3 : 49 The FOC is 1 ct 1 ct = = B1 r At+1 B1 At c t ) ct = 1 1 + B1 = (At ct ) r B1 = rAt 1 + B1 At+1 At ; Which we recognize well. We also have 1 ct+1 = At+1 1 + B1 1 B1 = 1 + B1 1 + B1 1 B1 = 1 + B1 1 + B1 B1 = rct 1 + B1 rAt r (1 + B1 ) ct Substituting this into our guess gives At 1 + B1 At = ln 1 + B1 B1 ln At + B2 ln vt + B3 = ln ln vt + (B1 ln At+1 + B2 ln vt+1 + B3 ) ln vt + B1 ln B1 r At + B2 ln vt+1 + B3 1 + B1 This cannot work unless we get rid of vt+1 in the RHS. To do this we note that vt+1 = Ct : and in general equilibrium Ct = Ct = vt+1 = B1 rCt 1 1 + B1 B1 r Ct 1 1 + B1 B1 r vt 1 + B1 50 Therefore, B1 ln At +B2 ln vt +B3 = ln At ln vt + 1 + B1 B1 ln B1 r At + B2 ln 1 + B1 r B1 vt + B3 1 + B1 We solve this by equalizing the coe¢ cients on the di¤erent terms B1 = (1 + B1 ) B2 = 1 + B2 B3 = (B1 + B2 ) ln B1 r (1 + B1 + B2 ) ln (1 + B1 ) + B3 1 B1 = 1 1 B2 = 1 Again using the FOC, we get ct = At = (1 1 + 11 ) At As we see, consumption not a¤ected by the habit. This is due to the log utility. Marginal utility of consumption is c1t regardless of vt since u is separable in c and v: Let us therefore consider a generalization. Instead of solving the full problem, we can at least characterize consumption dynamics. Suppose 1 u (ct ; vt ) = ct vt 1 1 u1 (ct ; vt ) = ct ct vt 1 1 The Euler condition under purely external habits is the usual u1 (ct+1 ; vt+1 ) r u1 (ct ; vt ) 1= In general equilibrium vt = ct 1 , giving 1 1= ct+1 1 ct ct+1 vt+1 ct vt 1 1 1 r= ct+1 1 ct 1 ct+1 ct 1 ct ct 1 51 r = ct+1 ct ( 1)+ ct (1 1 ) r Taking logs (ln ct+1 0 = ln r ln ct ) = ln r In the case constant at ln ct+1 + ( (1 ) (ln ct (1 )) ln ct + (1 ln ct 1 ) ) ln ct = 0;the standard case, the growth rate of consumption is ln ct+1 ln ct = ln r ; as we should now from standard models. We have also seen that with = 1; ln ct+1 ln ct = ln r Dynamics becomes interesting now under ln ct+1 ln ct 6= 1 and > 0: De…ne gt+1 Then, gt+1 = ln r (1 ) (7) gt When riskaversion is low (IES high), that is a > 0; we get oscillations! A low growth rate is followed by a high and vice versa. The oscillations may even be unstable if (1 ) > 1: If instead < 0;we get a monotone path, (1 ) that is stable if < 1: Note that we have assumed a constant interest rate r; this is quite easy to relax. With a varying interest rate, for example if we include capital accumulation, we still have gt+1 = 6.2 ln rt+1 (1 ) gt : Adding stochastics With stochastics, we have Et 1 ct+1 Et ct+1 ct Et 1 ct+1 ct rt+1 = ( 1) ct+1 ct gt+1 Et (e ct 1 ( 1) 1 (1 ct ct gt rt+1 = (e ) 52 1 ct rt+1 = ct ct rt+1 = ) 1 ct 1 (1 ) ) : 1 : : : This can be analyzed by linearization. Suppose rt+1 = ezt+1 r where r = 1; giving a steady state of the economy. Approximating around g = 0; z = 0, (egt+1 ) ezt+1 r (egt ) (1 r ) gt+1 r + rzt+1 1 + (1 ) gt Giving Et ( r gt+1 r + rzt+1 ) = 1 + (1 (1 Et gt+1 = ) gt ) Et zt+1 gt + ; which can provide interesting dynamics. To understand the reason for the oscillatory behavior, it may be of help to note that individuals with habits might prefer variations over time. Clearly, when = 0; individuals are risk averse for > 0 and also averse to variations over time. Consider instead the case when = 1: Assume that (individual and aggregate) consumption is c (1 + ") if t odd c (1 ") else. ct = In this case, utility in even periods (multiplied by (1 is ut = 1+" 1 " and in odd ut+1 = 1 " 1+" ) for convenience) 1 1 we have ut + ut+1 = 2 = 1 " 1 1+" 1 " 1 1+" 2 6d 1 + "6 4 = 1 + 2 (1 + 1+" 1 " 2 + 1+" 1 " 2 1 ( 11+"" ) 1 1 1 +( 11+"" ) 2 d" 3 7 7 5 "=0 2 ) " 2 53 2 1 2 d "2 6 6 + 4 2 ( 11+"" ) 1 +( 11+"" ) 2 d"2 3 7 7 5 "=0 which is increasing in "2 : 6.3 Asset market implications. Abel has shown that habits also may have an ability to explain asset market puzzles. Let us …nally go over this. Using 1 u (ct ; vt ) = 1 1 ct 1 u1 (ct ; vt ) = ct vt ct 1 vt utility in period t can be written Ut = s.t. At+1 At+2 1 ct 1 1 ct+1 1 + vt 1 vt+1 D ct ) rt+1 ; vt = ct 1 Ct1 1D ct+1 ) rt+2 1 = (At = (At+1 + V (At+2 ; vt+2 ) Therefore @Ut 1 = @ct ct ct 1 = ct ct 1 = ct 1 ct+1 D t t+1 1 D t ct ! 1 ct+1 t ct t+1 1 ct+1 ct 1 1 1 D t t Now, de…ne gross output growth yt+1 yt xt+1 and since the economy is closed xt+1 = and thus t+1 vt = ct+1 Ct+1 = ct Ct 1 cD t Ct D 1 cD t 1 Ct D 1 54 ! 1 ct = xt ; 1 t t+1 ! implying @Ut 1 = @ct ct 1 ct 1 D t 1 = ct t = ct t 1 1 ct+1 ct xt+1 xt D 1 1 1 t t+1 ! ! (8) Ht+1 where 1 xt+1 xt Clearly the marginal utility of consumption in t is increasing in Ht+1 :The latter, in turn, is high when xt+1 is low relative to xt if > 0, and vice-versa otherwise. So if growth is expected to be low between t and t + 1; and IES is large ( > 0), this boosts the marginal utility of consumption. In other words, an expectation of low growth has a negative e¤ect on savings. Note that this goes against the standard smoothing results that if you expect low income in the future, this strengthens the savings motive. Furthermore, in asset market equilibrium, this e¤ect tends to reduce the price of assets, i.e., increasing the expected return. Let us now …rst consider the case when D = 0: We will see that we can get some interesting results for bond and asset returns. t = ct t 1 :The Euler equation implies When D = 0; Ht = 18t and @U @ct as usual @Ut @Ut+1 = Et rt+1 @ct @ct+1 1 c 1 = Et rt+1 t+1 t+11 ct t Ht+1 1 D Now consider the endowment economy, where ct = yt and we de…ne yt+1 xt+1 : yt Then, we have c 1 = Et rt+1 t+1 ct 1 t+1 1 t = Et rt+1 xt+1 xt ( 1) Now, consider a risky share that pays yt as dividend (the apple trees) with price pr;t : The price of this asset must satisfy rr;t+1 = pr;t+1 + yt+1 : pr;t 55 Denoting the price-dividend ratio pr;t yt wt we can write wt+1 yt+1 + yt+1 ; wt yt 1 + wt+1 xt+1 : = wt rr;t+1 = Now, using this last expression for the return on stocks into the Euler equation yields ( 1) 1 = Et rt+1 xt+1 xt 1 + wt+1 ( 1) = Et xt+1 xt+1 xt wt ( 1) wt = xt Et (1 + wt+1 ) x1t+1 ; since wt is known in t: When growth rates are i.i.d., this can be calculated quite easily. In particular, we will show that the expression 1 (1 + wt+1 ) xt+1 Et is constant at some value A, so that we can write wt = Axt ( 1) for some A and verify that this satis…es the pricing equation (11). Using our "guess", we have Axt ( 1) = xt ( A = Et 1) Et ( 1 + Axt+1 ( 1 + Axt+1 1) 1) 1 xt+1 ; x1t+1 ( = Et x1t+1 + Et Axt+1 = 1 Et xt+1 + Et Axt+1 (1 1)+(1 )(1 ) ) Clearly, the RHS is a constant if growth rates are i.i.d. and we have established that ( 1) wt = Axt 56 where Ex1 A = A = 1 + EAx(1 Ex1 Ex(1 )(1 ) )(1 ) To calculate the expected stock market return, we use rr;t+1 = = 1 + wt+1 xt+1 wt ( 1) 1 + Axt+1 Axt ( 1) xt+1 and Et rr;t+1 = = 1 + AEx ( Axt 1 + AEx Axt ( ( 1) Ex 1) ( 1) 1) Ex As we see, the expected return is time dependent, despite the i.i.d. assumption, provided 6= 0: Why? Furthermore, if > 0; and < 1; the denominator decreases in xt and yt the expected return is thus higher when xt is high. If > 1; the yt 1 opposite is true. We can easily calculate the unconditional return, i.e., the average return over time 1 + AEx ( 1) Ex : Err = E Ax ( 1) In a similar fashion, the unconditional return on bonds is Erb = Ex ( (Ex 1) ) Consider the special case when output growth x is lognormal, with mean g and standard deviation : Recall that then 1 2 Ex = eg+ 2 : Furthermore, if ln x is normal, ln x = ln x is normal with mean g and standard deviation :Thus, Ex = e 57 g+ 2 2 2 : Now, let ln x be normally distributed with mean g and standard deviation . Then, Ex1 A = + AEx(1 Ex1 Ex(1 1 ( ) ) e(1 = e(1 1 1) 1 + AEx AEx ( ( ln Err = ln Ex = ln e )(1 )(1 1)g+( ( 1)) 2 ( )2 2 2 (1 ((1 )g+ )(1 2 ))2 2 1) 1) 2 2 )(1 )g+ eg+ Ex 2 2 + Ae(1+ ( 1))g+(1+ ( 1))2 A 2 2 ! and ln ERB Ex ( (Ex rB = ln = ln e ( 1)g+ e = ( ( 1) ) ( ( g+ 1))2 2 2 2 2 2 1)) g + ! ( ( 1))2 2 2 2 ln Setting, = 0; = 0:036; g = 0:018; = :99 and plotting against rs and rB against 1 ; we have the no habit case. 58 0.14 0.12 0.1 0.08 0.06 0.04 0.02 0 2 4 Relative Riskaversion 6 8 10 12 Average stockmarket return rS (solid line) and safe return rB against risk aversion ( ): Standard utility ( = 0): Keeping the other parameters, but now introducing external habits by setting = 1; the returns are given in the second …gure. 59 0.2 0.18 0.16 0.14 0.12 0.1 0.08 0.06 0.04 0.02 0 2 4 Relative riskaversion 6 8 10 12 Average stockmarket return rS (solid line) and safe return rB against risk aversion ( ): Extrenal Habit ( = 1): As we see, the stock market return quickly becomes very high as we reduce : 6.4 Appendix: The case when D > 0: The stochastic Euler equation Et Note that is ator. @Ut @ct @Ut @Ut+1 = Et+1 rt+1 @ct @ct+1 is actually not realized at t; thus the expectations oper- Et @Ut = @ct 1 = Et rt+1 @Ut+1 @ct+1 t+1 Et rt+1 @U @ct+1 t Et @U @ct 60 : (9) Now, it is convenient to …nd an expression for @Ut+1 @ct+1 Et @Ut @ct Shifting (8) forward, yields @Ut+1 = Ht+2 @ct+1 Using, t+1 t = xt and @Ut+1 @ct+1 Et @Ut @ct ct+1 ct 1 t+1 ct+1 : = xt+1 ;yields Ht+2 t+11 ct+1 = Et Ht+1 t 1 ct Ht+2 = Et Ht+1 Ht+2 ( = xt Et Ht+1 1 t+1 t 1) ct+1 ct xt+1 : Now, consider a risky share that pays yt as dividend (the apple trees) with price pr;t : The price of this asset must satisfy rr;t+1 = pr;t+1 + yt+1 : pr;t Denoting the price-dividend ratio pr;t yt wt we can write wt+1 yt+1 + yt+1 ; wt yt 1 + wt+1 xt+1 : = wt rr;t+1 = Now, using this last expression for the return on stocks into the Euler equation yields 1 = wt = t+1 Et rr;t+1 @U @ct+1 t Et @U @ct ; t+1 Et (1 + wt+1 ) xt+1 @U @ct+1 t Et @U @ct 61 ; (10) since wt is known in t: From (10), we have t+1 Et (1 + wt+1 ) xt+1 @U @ct+1 wt = Et @Ut @ct t+1 Et (1 + wt+1 ) xt+1 @U @ct+1 = Et @Ut @ct Et (1 + wt+1 ) xt+1 ct+1 t+11 Ht+2 Et ct t 1 Ht+1 = Et (1 + wt+1 ) xt+1 = 1 ct+1 ct t+1 t Ht+2 Et (Ht+1 ) 1 Et (1 + wt+1 ) xt+1 xt+1 (xt ) = Ht+2 Et (Ht+1 ) Et (1 + wt+1 ) xt = ( 1) Ht+2 x1t+1 Et Ht+1 xt = ( 1) 1 Et (1 + wt+1 ) Ht+2 xt+1 : Et Ht+1 Using the law of iterated expectations 1 Et 1 + wt+1 )Ht+2 x1t+1 = Et (1 + wt+1 ) xt+1 Et+1 (Ht+2 ) De…ne Jt Et (Ht+1 ) = 1 Dxt ( 1) 1 Et xt+1 : Then we have wt = xt ( 1) 1 Et (1 + wt+1 ) Jt+1 xt+1 : Jt (11) We now need to …nd wt as a function of state variables (which are they?) that satis…es (11). When growth rates are i.i.d., this can be calculated quite easily. In particular, we will show that the expression Et (1 + wt+1 ) Jt+1 x1t+1 62 is constant at some value A, so that we can write ( 1) Axt wt = Jt for some A and verify that this satis…es the pricing equation (11). Using our "guess", we have ! ! ( 1) ( 1) ( 1) Ax Axt xt t+1 = Et 1+ Jt+1 x1t+1 Jt Jt Jt+1 ! ! ( 1) Axt+1 1 Jt+1 xt+1 A = Et 1+ Jt+1 ( = 1 Et Jt+1 xt+1 + Et Axt+1 = Et Jt+1 x1t+1 + Et Axt+1 (1 1)+(1 )(1 ) ) So A 1 (1 Et xt+1 )(1 ) = Et Jt+1 x1t+1 (12) Now, under the assumption of i.i.d. output (consumption) shocks, Et x1t+1 = Et x1t+2 = Ex1 we have Jt = Et Ht+1 = 1 Dxt ( 1) Et x1t+1 = 1 Dxt ( 1) Ex1 ( 1) so Jt+1 = 1 Jt+1 x1t+1 1 = xt+1 Dxt+1 1 Ex1 ( Dxt+1 1) Ex1 Take conditional expectation at t Et Jt+1 x1t+1 = Et x1t+1 1 Dxt+1 Ex1 = Et x1t+1 DEx1 = Ex1 DEx1 = Ex1 1 63 ( 1 Et xt+1 xt+1 E x(1 DE x(1 )(1 )(1 ) ) 1) Finally, use this in (12), and use the i.i.d. assumption to replace conditional expectations A 1 (1 Et xt+1 )(1 ) = A = Ex1 1 Ex1 1 (1 DE x(1 )(1 ) DE x(1 )(1 Ex(1 )(1 ) ) ) which is clearly a constant under the i.i.d. assumption. To calculate the expected stock market return, we use pSt+1 + yt+1 pSt+ (1 + wt+1 ) xt+1 = wt S Rt+1 = and ( Et S Et Rt+1 1+ = = Axt+1 Jt+1 1) xt+1 wt Ex + Et 1+ ( Axt+1 Jt+1 ( Axt Jt Which we at least can simulate. 64 1) 1) ; 7 Loss-Aversion Substantial amounts of lab-evidence suggests that individuals behave like if the formed reference levels for consumption. Preferences over actual consumption then depend in a particular way of consumption relative to this reference level. Speci…cally, preferences are consistent with utility maximization if 1. utility is concave in consumption if consumption is above the reference level, 2. utility is convex in consumption if consumption is below the reference level, and 3. marginal utility is discretely larger below the reference level than what it is above. The utility function can then be depicted as follows: Two important implications of this is that 1. Individuals have a strict distaste also for abitrarily small gambles (because of the discontinuity). 2. Individuals are risk-lovers for losses. This means that they may prefer a 50/50 bet of loosing x or nothing over a sure loss of x=2: Kahneman & Tversky proposes u (c (c r) if c r ( (c r)) else r) = 65 and in lab-experiments …nds that following graph1 = = 0:88; and = 2:25as in the 1 0.5 0 -0.5 -1 -1.5 -2 -0.8 -0.6 -0.4 -0.2 0x 0.2 0.4 0.6 0.8 1 In the paper by Bowman et al., this implication is expanded into a dynamic setting. One implication is then that consumption may respond asymmetrically to positive and negative news about future income. To get the intuition, consider a two period setting and suppose that income at the outset is expected to be w in both periods. Suppose also that the reference point for consumption is r = w: For simplicty suppose that the interest rate equals the subjective discount rate. Clearly, the optimal consumption is now c = w in both periods. Consider now a positive but uncertain signal about period 2 income. Say that income is w + 2x with probability 1=2 and w with probability 1=2: Expected lifetime income is then 2w + x and unless there is some precautionary savings, consumption in period 1 will be w + 12 x: This is the permanent income hypothesis. In any case, consumption will certainly increase when this positive signal comes. Behavior is "standard" for gains. Consider now instead a negative signal saying that second period income is w 2x with probability 12 and w otherwise. Now, it may very well pay for the household to continue to consume w in period 1 and then with probability 1=2 consume w also in the …nal period and with probability 21 consume w 2x, rather than consuming the permanent income w 12 x in the …rst period, in which case second period consumption is either w + 12 x or w 32 x: Why? 1 A better formulation, allowing ; < 0; is u (c 66 r) = ( (c r) if c ( (c r)) r else Let r denote the reference level for consumption. Then, the expected utility of the …rst strategy is 1 1 u (w r) + u (w r) + u (w 2x r) 2 2 1 0+0 (2x) 2 and for the second strategy is it u w r 1 x 2 = 1 1 1 x + u w+ x 2 2 2 + 1 2 1 x 2 1 r + u w 2 3 x 2 3 x 2 The di¤erence is 1 x 2 1 (2x) 2 = 1 (2) + 2 1 2 r 3 2 + 1 1 + x 2 2 ! x 3 x 2 ! 1+ 1 2 x Under the assumption a = ; this is positive if ! 1+ 3 1 1 1 + > 0 (2) + 2 2 2 2 > 2 2 +2 1 1+ 2 1 + 2 2 3 2 2 1 1 ; 4 2 Due to the convexity of utility in losses, its better to take a chance that consumption might not need to be reduced below the reference point. In other words, there is a tendency that consumption does not fall "unless it is clear that it has to". Bowman et al documents such an asymmetry in U.S. consumption data. In a dynamic setting, a key issue is how reference points are formed. Unfortunately, not much empirical evidence is collected regarding this issue. Bowman et al assume r1 ; the reference point for period 1, is given and that r2 = (1 ) r1 + c 1 If = 0; we have the case discussed above –static reference points. For = 1; next periods reference points is completely determined by the previous periods consumption. Here, we could think both of the case when the agent internalize the e¤ect her consumption has on the reference point and the case when she doesn’t (due to naivite or external reference points). 67 7.1 Loss-aversion as commitment (Hassler&Rodriguez Mora) Two types of rational and non-altruistic individuals, (poor) workers and entrepreneurs, living in a two period OLG-setup. The workers make no private choices, having a …xed wage normalized to zero, consuming in the second period of life, having high marginal utility since they are poor. i2 Young entrepreneurs in t choose investments it at a utility cost 2t , returning it (1 t+1 ) in second period of life when consumption takes place and the capital fully depreciates. Old workers get a transfer Gt , …nanced by taxes on installed capital. Taxes, t 2 [0; 1], are determined without commitment by probabilistic voting with equal weight on all living individuals. Alternative interpretation, a benevolent planner that cares equally of all living individuals. Without commitment, the only Markov equilibrium is one with 100 percent taxation since temptation to tax installed capital is too high. As Köszegi and Rabin, we consider the case when reference points for consumption are forward-looking. We can refrase reference points for cosumption in terms of the corresponding tax-level, r : We require r t+1 to be in the set of equilibrium tax rates for t + 1. We allow politicians to a¤ect reference points by making "promises" about the future. But remember that the promise is empty – the politician does not remain in o¢ ce nor runs again and he has no formal commitment power. The promise can a¤ect the future if it is believed, in which case it becomes the the reference point. It is believed if it is done by the winning candidate and is in the set of equilibria for next period. If the promise is not an equilibrium, rt+1 is some element of the set of equilibrium tax rates. As a variation, we consider the opposite case of history dependence. Reference tax-levels are backward-looking, rt+1 = t 68 7.1.1 Results Under both backward and forward-looking reference points, there is a Markov equilibrium with limited amounts of taxation. Dynamics di¤er between the two cases. The level of taxation in equilibrium depends inversely on on the degree of loss-aversion. Intuition: If people have reference points, implying that they feel "entitled" to some return on their investments – if becomes politically costly to go against this. If the entitlements are not too large, they will be satis…ed in equilibrium. 69