Learning and Global Dynamics

Transcription

1 Learning and Global Dynamics James Bullard 10 February 2007

2 Learning and global dynamics The paper for this lecture is Liquidity Traps, Learning and Stagnation, by George Evans, Eran Guse, and Seppo Honkapohja.

3 Learning and global dynamics The paper for this lecture is Liquidity Traps, Learning and Stagnation, by George Evans, Eran Guse, and Seppo Honkapohja. This will serve as an introduction to some key ideas in the learning literature.

4 Learning and global dynamics The paper for this lecture is Liquidity Traps, Learning and Stagnation, by George Evans, Eran Guse, and Seppo Honkapohja. This will serve as an introduction to some key ideas in the learning literature. The main idea is to study stability under learning of systems analyzed by Benhabib, Schmitt-Grohe, and Uribe (2001, JET and elsewhere).

5 Learning and global dynamics The paper for this lecture is Liquidity Traps, Learning and Stagnation, by George Evans, Eran Guse, and Seppo Honkapohja. This will serve as an introduction to some key ideas in the learning literature. The main idea is to study stability under learning of systems analyzed by Benhabib, Schmitt-Grohe, and Uribe (2001, JET and elsewhere). The learning dynamics give a different perspective from the RE dynamics.

6 What Benhabib, Schmitt-Grohe, and Uribe said Monetary models normally impose a Fisher relation R = ρ + π and a zero lower bound on nominal interest rates R > 0.

7 What Benhabib, Schmitt-Grohe, and Uribe said Monetary models normally impose a Fisher relation R = ρ + π and a zero lower bound on nominal interest rates R > 0. Many analyses include a continuous, active Taylor type monetary policy rule R 0 (π? ) > 1, where π? is the target inflation rate of the monetary authority.

8 What Benhabib, Schmitt-Grohe, and Uribe said Monetary models normally impose a Fisher relation R = ρ + π and a zero lower bound on nominal interest rates R > 0. Many analyses include a continuous, active Taylor type monetary policy rule R 0 (π? ) > 1, where π? is the target inflation rate of the monetary authority. Main point: This combination of assumptions always implies the existence of a second steady state inflation rate π L < π?.

9 What Benhabib, Schmitt-Grohe, and Uribe said Monetary models normally impose a Fisher relation R = ρ + π and a zero lower bound on nominal interest rates R > 0. Many analyses include a continuous, active Taylor type monetary policy rule R 0 (π? ) > 1, where π? is the target inflation rate of the monetary authority. Main point: This combination of assumptions always implies the existence of a second steady state inflation rate π L < π?. Perfect foresight equilibria may exist in which inflation begins in the neighborhood of π? but converges asymptotically to π L along an oscillatory path.

10 Benhabib, et al.: Existence of a liquidity trap

11 Benhabib, et al., Figure 3

12 Japan Japan: Gross Domestic Product % Change Year to Year SAAR, Bil.Chn.2000.Yen 6 Japan: Uncollateralized Overnight Call Rate % p.a Source: OECD /Haver

13 A learning application? Rational expectations dynamics suggested by Benhabib, et al., seem unlikely.

14 A learning application? Rational expectations dynamics suggested by Benhabib, et al., seem unlikely. Multiple steady states, which one would be attained in an actual economy?

15 A learning application? Rational expectations dynamics suggested by Benhabib, et al., seem unlikely. Multiple steady states, which one would be attained in an actual economy? Would it be possible under learning to switch from a neighborhood of one steady state to a neighborhood of the other?

16 A learning application? Rational expectations dynamics suggested by Benhabib, et al., seem unlikely. Multiple steady states, which one would be attained in an actual economy? Would it be possible under learning to switch from a neighborhood of one steady state to a neighborhood of the other? How do policy choices influence these dynamics?

17 Main ideas Sticky price, stochastic, discrete time version of Benhabib, et al. (2001).

18 Main ideas Sticky price, stochastic, discrete time version of Benhabib, et al. (2001). Replace rational expectations with recursive learning.

19 Main ideas Sticky price, stochastic, discrete time version of Benhabib, et al. (2001). Replace rational expectations with recursive learning. Under normal policy, economy will converge to targeted steady state, and agents will behave as if they have rational expectations.

20 Main ideas Sticky price, stochastic, discrete time version of Benhabib, et al. (2001). Replace rational expectations with recursive learning. Under normal policy, economy will converge to targeted steady state, and agents will behave as if they have rational expectations. Large, pessimistic shocks can send the economy on a path toward the low inflation steady state.

21 Main ideas Sticky price, stochastic, discrete time version of Benhabib, et al. (2001). Replace rational expectations with recursive learning. Under normal policy, economy will converge to targeted steady state, and agents will behave as if they have rational expectations. Large, pessimistic shocks can send the economy on a path toward the low inflation steady state. Alternative policies may eliminate this possibility.

22 Background Idea is to get a model like the ones in this literature but also amenable to recursive learning analysis.

23 Background Idea is to get a model like the ones in this literature but also amenable to recursive learning analysis. Continuum of household-firms each produce a differentiated consumption good under monopolistic competition.

24 Background Idea is to get a model like the ones in this literature but also amenable to recursive learning analysis. Continuum of household-firms each produce a differentiated consumption good under monopolistic competition. There is no capital, production is simply where h t,j is the labor input. y t,j = h α t,j (1)

25 Background Idea is to get a model like the ones in this literature but also amenable to recursive learning analysis. Continuum of household-firms each produce a differentiated consumption good under monopolistic competition. There is no capital, production is simply where h t,j is the labor input. The labor market is competitive. y t,j = h α t,j (1)

26 More background Firms face downward sloping demand P t,j = yt,j Y t 1/ν P t. (2)

27 More background Firms face downward sloping demand P t,j = yt,j Y t 1/ν P t. (2) P t,j is the profit-maximizing price set by firm j.

28 More background Firms face downward sloping demand P t,j = yt,j Y t 1/ν P t. (2) P t,j is the profit-maximizing price set by firm j. Elasticity of substitution between goods is given by ν > 1.

29 More background Firms face downward sloping demand P t,j = yt,j Y t 1/ν P t. (2) P t,j is the profit-maximizing price set by firm j. Elasticity of substitution between goods is given by ν > 1. Price adjustment costs are of the Rotemberg type.

30 Households maximize E 0 β t U t,j c t,j, M t 1,j P t,j, h t,j, 1 P t=0 t P t 1,j subject to c t,j + m t,j + b t,j + τ t,j = m t 1,j π 1 t + R t 1 π 1 t b t 1,j + P t,j P t y t,j. where (3) (4) U t,j = c1 σ 1 t,j + χ 1 σ 1 1 σ 2 Mt 1,j P t 1 σ2 h 1+ɛ t,j 1 + ɛ 2 γ Pt,j 1. 2 P t 1,j (5)

31 Households maximize E 0 β t U t,j c t,j, M t 1,j P t,j, h t,j, 1 P t=0 t P t 1,j subject to c t,j + m t,j + b t,j + τ t,j = m t 1,j π 1 t + R t 1 π 1 t b t 1,j + P t,j P t y t,j. where U t,j = c1 σ 1 t,j + χ 1 σ 1 1 σ 2 Mt 1,j P t 1 σ2 h 1+ɛ t,j 1 + ɛ (3) (4) 2 γ Pt,j 1. 2 P t 1,j (5) Notation standard; last term in utility is Rotemberg cost of price adjustment.

32 Fiscal policy The government budget constraint is b t + m t + τ t = g t + m t 1 π t 1 + R t 1 π 1 t b t 1 (6) where τ t is a lump-sum tax and g t is government consumption.

33 Fiscal policy The government budget constraint is b t + m t + τ t = g t + m t 1 π t 1 + R t 1 π 1 t b t 1 (6) where τ t is a lump-sum tax and g t is government consumption. Assume government consumption is stochastic where u t is white noise. g t = g + u t (7)

34 Fiscal policy The government budget constraint is b t + m t + τ t = g t + m t 1 π t 1 + R t 1 π 1 t b t 1 (6) where τ t is a lump-sum tax and g t is government consumption. Assume government consumption is stochastic where u t is white noise. Assume fiscal policy follows g t = g + u t (7) τ t = κ 0 + κb t 1 + ψ t + η t (8) a linear tax rule as in Leeper (1991).

35 Monetary policy Monetary policy follows a global interest rate rule R t 1 = θ t f (π t ), (9) where f (π) is non-negative and non-decreasing.

36 Monetary policy Monetary policy follows a global interest rate rule R t 1 = θ t f (π t ), (9) where f (π) is non-negative and non-decreasing. θ t is an exogenous iid positive random shock with mean 1.

37 Monetary policy Monetary policy follows a global interest rate rule R t 1 = θ t f (π t ), (9) where f (π) is non-negative and non-decreasing. θ t is an exogenous iid positive random shock with mean 1. The monetary authority has an inflation target π? where R? = β 1 π? and f (π? ) = R? 1.

38 Monetary policy Monetary policy follows a global interest rate rule R t 1 = θ t f (π t ), (9) where f (π) is non-negative and non-decreasing. θ t is an exogenous iid positive random shock with mean 1. The monetary authority has an inflation target π? where R? = β 1 π? and f (π? ) = R? 1. For some purposes π AR f (π) = (R?? /(R? 1) 1) π? (10) where f 0 (π? ) = AR? is assumed larger than β 1.

39 Normal policy Normal policy consists of...

40 Normal policy Normal policy consists of the government budget constraint,

41 Normal policy Normal policy consists of the government budget constraint, the fiscal policy rule,

42 Normal policy Normal policy consists of the government budget constraint, the fiscal policy rule, and the monetary policy rule.

43 Normal policy Normal policy consists of the government budget constraint, the fiscal policy rule, and the monetary policy rule. The baseline analysis is under normal policy.

44 Equilibrium Private sector optimization yields three equations given in the text.

45 Equilibrium Private sector optimization yields three equations given in the text. Combine these three with the government budget constraint, the fiscal policy rule, the monetary policy rule, and market clearing.

46 Benhabib et al If f (π) is continuous, differentiable, and has a steady state π? in which f 0 (π? ) > β 1, a second steady state π L exists with f (π L ) < β 1.

47 Benhabib et al If f (π) is continuous, differentiable, and has a steady state π? in which f 0 (π? ) > β 1, a second steady state π L exists with f (π L ) < β 1. At both steady states, R = β 1 π.

48 Benhabib et al If f (π) is continuous, differentiable, and has a steady state π? in which f 0 (π? ) > β 1, a second steady state π L exists with f (π L ) < β 1. At both steady states, R = β 1 π. Unique values c > 0 and h > 0 are associated with positive steady state inflation rates.

49 Benhabib et al If f (π) is continuous, differentiable, and has a steady state π? in which f 0 (π? ) > β 1, a second steady state π L exists with f (π L ) < β 1. At both steady states, R = β 1 π. Unique values c > 0 and h > 0 are associated with positive steady state inflation rates. At deflationary steady states, c > 0 and h > 0 are unique provided π is close to one and g > 0.

50 Benhabib et al If f (π) is continuous, differentiable, and has a steady state π? in which f 0 (π? ) > β 1, a second steady state π L exists with f (π L ) < β 1. At both steady states, R = β 1 π. Unique values c > 0 and h > 0 are associated with positive steady state inflation rates. At deflationary steady states, c > 0 and h > 0 are unique provided π is close to one and g > 0. Corresponding stochastic steady states exist when the support of the exogenous shocks is sufficiently small.

51 Linearization Linearization produces a decoupled system of four equations in c, π, b, and m.

52 Linearization Linearization produces a decoupled system of four equations in c, π, b, and m. Equilibrium dynamics can be analyzed by considering the equations for c and π alone, provided debt dynamics are stationary.

53 Linearization Linearization produces a decoupled system of four equations in c, π, b, and m. Equilibrium dynamics can be analyzed by considering the equations for c and π alone, provided debt dynamics are stationary. The system can be written as ct π t = Bcc B πc B cπ B ππ + c e t+1 π e t+1 Gcu G cθ G πu G πθ ut θ t kc +. k π

54 Determinacy If both eigenvalues of B lie inside the unit circle, a unique nonexplosive solution exists of the form ct π t = c π Gcu + G πu G cθ G πθ ut θ t. (11)

55 Determinacy If both eigenvalues of B lie inside the unit circle, a unique nonexplosive solution exists of the form ct π t = c π Gcu + G πu G cθ G πθ ut The corresponding m t is a constant plus white noise. θ t. (11)

56 Determinacy If both eigenvalues of B lie inside the unit circle, a unique nonexplosive solution exists of the form ct π t = c π Gcu + G πu G cθ G πθ ut The corresponding m t is a constant plus white noise. θ t. (11) The remaining condition for determinacy is that fiscal policy is passive according to Leeper (1991), which means that β 1 κ < 1.

57 Proposition 1 Assume fiscal policy is passive.

58 Proposition 1 Assume fiscal policy is passive. Assume γ > 0 sufficiently small.

59 Proposition 1 Assume fiscal policy is passive. Assume γ > 0 sufficiently small. Then the steady state with inflation at target π = π? is locally determinate.

60 Proposition 1 Assume fiscal policy is passive. Assume γ > 0 sufficiently small. Then the steady state with inflation at target π = π? is locally determinate. And, the steady state with inflation π = π L is locally indeterminate.

61 Perceived law of motion Equilibria in this model are simple iid processes.

62 Perceived law of motion Equilibria in this model are simple iid processes. This implies agents can forecast by estimating mean values. Very helpful.

63 Perceived law of motion Equilibria in this model are simple iid processes. This implies agents can forecast by estimating mean values. Very helpful. The hallmark of the literature is the assignment of a perceived law of motion π e t+1 = π e t + φ t (π t 1 π e t), (12) c e t+1 = c e t + φ t (c t 1 c e t). (13)

64 Perceived law of motion Equilibria in this model are simple iid processes. This implies agents can forecast by estimating mean values. Very helpful. The hallmark of the literature is the assignment of a perceived law of motion π e t+1 = π e t + φ t (π t 1 π e t), (12) Here φ t is the gain sequence. c e t+1 = c e t + φ t (c t 1 c e t). (13)

65 Gain sequences Recursive least squares learning sets φ t = 1/t.

66 Gain sequences Recursive least squares learning sets φ t = 1/t. Asymptotic convergence to rational expectations possible

67 Gain sequences Recursive least squares learning sets φ t = 1/t. Asymptotic convergence to rational expectations possible Recursive constant gain learning sets φ t = φ > 0, a small positive constant.

68 Gain sequences Recursive least squares learning sets φ t = 1/t. Asymptotic convergence to rational expectations possible Recursive constant gain learning sets φ t = φ > 0, a small positive constant. More robust to structural change. Convergence properties weaker.

69 Gain sequences Recursive least squares learning sets φ t = 1/t. Asymptotic convergence to rational expectations possible Recursive constant gain learning sets φ t = φ > 0, a small positive constant. More robust to structural change. Convergence properties weaker. Theorems: LSL. Simulations: Constant gain.

70 Expectational stability Approximate π 1 t+1 c σ 1 t+1 by π e t+1 ce t+1 σ1 1. This changes the dynamic system slightly. The linearization is unchanged.

71 Expectational stability Approximate π 1 t+1 c σ 1 t+1 by π e t+1 ce t+1 σ1 1. This changes the dynamic system slightly. The linearization is unchanged. The system is now the two altered equations for c, π, π t = F π (π e t+1, ce t+1, u t, θ t ), (14) c t = F c (π e t+1, ce t+1, u t, θ t ). (15) the monetary policy rule, and the updating equations for expectations.

72 More on expectational stability The REE is said to be expectationally stable if the differential equation in notional time τ dπ e /dτ dc e /dτ = Tπ (π e, c e ) T c (π e, c e ) π e c e (16) is locally asymptotically stable at a steady state (π, c).

73 More on expectational stability The REE is said to be expectationally stable if the differential equation in notional time τ dπ e /dτ dc e /dτ = Tπ (π e, c e ) T c (π e, c e ) π e c e (16) is locally asymptotically stable at a steady state (π, c). Expectational stability is determined by the Jacobian matrix DT of T at the steady state, B for small noise.

74 More on expectational stability The REE is said to be expectationally stable if the differential equation in notional time τ dπ e /dτ dc e /dτ = Tπ (π e, c e ) T c (π e, c e ) π e c e (16) is locally asymptotically stable at a steady state (π, c). Expectational stability is determined by the Jacobian matrix DT of T at the steady state, B for small noise. The condition is then that both eigenvalues of B real parts less than zero. I have

75 More on expectational stability The REE is said to be expectationally stable if the differential equation in notional time τ dπ e /dτ dc e /dτ = Tπ (π e, c e ) T c (π e, c e ) π e c e (16) is locally asymptotically stable at a steady state (π, c). Expectational stability is determined by the Jacobian matrix DT of T at the steady state, B for small noise. The condition is then that both eigenvalues of B real parts less than zero. I have Proposition 2. For γ > 0 sufficiently small, the steady state at π = π? is locally stable under learning and the steady state at π L is locally unstable, taking the form of a saddle point.

76 Implications One could stop here and claim that the Benhabib et al. (2001) constructed dynamics are unlikely to be observed in actual economies.

77 Implications One could stop here and claim that the Benhabib et al. (2001) constructed dynamics are unlikely to be observed in actual economies. One could also claim that liquidity traps that come out of this model are theoretical curiosities that need not worry actual policymakers.

78 Implications One could stop here and claim that the Benhabib et al. (2001) constructed dynamics are unlikely to be observed in actual economies. One could also claim that liquidity traps that come out of this model are theoretical curiosities that need not worry actual policymakers. The authors take a different course, pointing out that certain regions of instability exist.

79 Implications One could stop here and claim that the Benhabib et al. (2001) constructed dynamics are unlikely to be observed in actual economies. One could also claim that liquidity traps that come out of this model are theoretical curiosities that need not worry actual policymakers. The authors take a different course, pointing out that certain regions of instability exist. They want to design policy to eliminate these regions of instability.

80 Implications One could stop here and claim that the Benhabib et al. (2001) constructed dynamics are unlikely to be observed in actual economies. One could also claim that liquidity traps that come out of this model are theoretical curiosities that need not worry actual policymakers. The authors take a different course, pointing out that certain regions of instability exist. They want to design policy to eliminate these regions of instability. They simulate the global dynamics with larger values of γ > 0.

81 Figure 1

82 More aggressive monetary policy Change the monetary policy rule to R t = 1 + θt f (π t ) if π t > π ˆR if π t < π and ˆR R t 1 + θ t f (π t ) if π t = π.

83 More aggressive monetary policy Change the monetary policy rule to R t = 1 + θt f (π t ) if π t > π ˆR if π t < π and ˆR R t 1 + θ t f (π t ) if π t = π. The authors choose 1 < ˆR < min 1 + f (π t ), β 1 π.

84 More aggressive monetary policy Change the monetary policy rule to R t = 1 + θt f (π t ) if π t > π ˆR if π t < π and ˆR R t 1 + θ t f (π t ) if π t = π. The authors choose 1 < ˆR < min 1 + f (π t ), β 1 π. The idea is to follow normal policy when π t π, but cut interest rates to a low level if inflation threatens to move below the threshold.

85 More aggressive monetary policy

86 Altered monetary and fiscal policy In this policy, interest rates are aggressively lowered as described above.

87 Altered monetary and fiscal policy In this policy, interest rates are aggressively lowered as described above. If this does not work, government expenditures are increased until inflation increases to the desired level.

88 Altered monetary and fiscal policy In this policy, interest rates are aggressively lowered as described above. If this does not work, government expenditures are increased until inflation increases to the desired level. See Lemma 4.

89 Altered monetary and fiscal policy In this policy, interest rates are aggressively lowered as described above. If this does not work, government expenditures are increased until inflation increases to the desired level. See Lemma 4. This can again create more than two steady states depending on the choice of π.

90 Figure 4

91 Conclusion Multiple equilibria, one of which is a liquidity trap.

92 Conclusion Multiple equilibria, one of which is a liquidity trap. How serious a problem is this?

93 Conclusion Multiple equilibria, one of which is a liquidity trap. How serious a problem is this? The Japanese data are alarming.

94 Conclusion Multiple equilibria, one of which is a liquidity trap. How serious a problem is this? The Japanese data are alarming. This paper suggests the Benhabib et al., 2001, dynamics are not robust to small changes in expectational assumptions.

95 Conclusion Multiple equilibria, one of which is a liquidity trap. How serious a problem is this? The Japanese data are alarming. This paper suggests the Benhabib et al., 2001, dynamics are not robust to small changes in expectational assumptions. The targeted, high inflation steady state would be locally stable in the learning dynamics.

96 Conclusion Multiple equilibria, one of which is a liquidity trap. How serious a problem is this? The Japanese data are alarming. This paper suggests the Benhabib et al., 2001, dynamics are not robust to small changes in expectational assumptions. The targeted, high inflation steady state would be locally stable in the learning dynamics. The possibility of deflationary spirals would still exist however, unless policy is chosen carefully.