How to read this lecture...

Code should execute sequentially if run in a Jupyter notebook

The Permanent Income Model: II


This lecture continues our analysis of the linear-quadratic (LQ) permanent income model of savings and consumption

As we saw in our previous lecture on this topic, Robert Hall [Hal78] used the LQ permanent income model to restrict and interpret intertemporal comovements of aggregate nondurable consumption and nonfinancial income

For example, we saw how the model asserts that

  • consumption is a random walk
  • financial wealth has a unit root and is cointegrated with consumption

It’s easy to find other applications that show the usefulness of the same LQ framework

For example, a model isomorphic to the LQ permanent income model has been used by Robert Barro [Bar79] to interpret intertemporal comovements of a government tax collections and its expenditures net of debt service

This isomorphism means that in analyzing the LQ permanent income model, we are in effect also analyzing properties of the Barro tax smoothing model

In this lecture, we’ll

  • show how the solution to the LQ permanent income model can be obtained using LQ control methods
  • represent the model as a linear state space system, and
  • apply QuantEcon’s LinearStateSpace class to characterize statistical features of the consumer’s optimal consumption and borrowing plans

We’ll then use these characterizations to construct a simple model of cross-section wealth and consumption dynamics in the spirit of Truman Bewley [Bew86]

(Later we’ll study other Bewley models—see this lecture)

The model will prove useful for illustrating concepts such as

  • stationarity
  • ergodicity
  • ensemble moments and cross section observations

Accompanying this lecture is a Jupyter notebook containing code and figures

This lecture is coauthored with Chase Coleman


Let’s recall the basic features of the model as discussed in the previous permanent income model lecture

Consumer preferences are given by

(1)\[E_0 \sum_{t=0}^\infty \beta^t u(c_t),\]

where \(u(c) = -(c - \gamma)^2\)

The consumer maximizes (1) by choosing a consumption, borrowing plan \(\{c_t, b_{t+1}\}_{t=0}^\infty\) subject to the sequence of budget constraints

(2)\[c_t + b_t = \frac{1}{1 + r} b_{t+1} + y_t, \quad t \geq 0\]

and the no-Ponzi condition

(3)\[E_0 \sum_{t=0}^\infty \beta^t b_t^2 < \infty\]

The interpretation of all variables and parameters are the same as in the previous lecture

We continue to assume that \((1 + r) \beta = 1\)

The dynamics of \(\{y_t\}\) again follow the linear state space model

(4)\[\begin{split}\begin{aligned} z_{t+1} & = A z_t + C w_{t+1} \\ y_t & = U z_t \end{aligned}\end{split}\]

The restrictions on the shock process and parameters are the same

Digression on a useful isomorphism

The LQ permanent income model of consumption is mathematically isomorphic with a version of Barro’s [Bar79] model of tax smoothing.

In the LQ permanent income model

  • the household faces an exogenous process of nonfinancial income
  • the household wants to smooth consumption across states and time

In the Barro model

  • a government faces an exogenous sequence of government purchases (excluding interest payments on its debt)
  • a government wants to smooth tax collections across states and time

If we set

  • \(T_t\), total tax collections in Barro’s model to consumption \(c_t\) in the LQ permanent income model
  • \(G_t\), exogenous government expenditures in Barro’s model to nonfinancial incomce \(y_t\) in the permanent income model
  • \(B_t\), government risk-free one-period debt in Barro’s model to risk-free one period debt in the LQ permanent income model
  • \(R\), the gross rate of return on risk-free one-period government debt in Barro’s model to the gross rate of return on financial assets in the permanent income model of consumption

then the two models are mathematically equivalent

All characterizations of a \(\{c_t, y_t, b_t\}\) in the LQ permanent income model automatically apply to a \(\{T_t, G_t, B_t\}\) process in the Barro model of tax smoothing

A specification of the state process

For the purposes of this lecture, let’s assume \(\{y_t\}\) is a second-order univariate autoregressive process:

\[y_{t+1} = \alpha + \rho_1 y_t + \rho_2 y_{t-1} + \sigma w_{t+1}\]

We can map this into the linear state space framework in (4), as discussed in our lecture on linear models

To do so we take

\[\begin{split}z_t = \begin{bmatrix} 1 \\ y_t \\ y_{t-1} \end{bmatrix}, \quad A = \begin{bmatrix} 1 & 0 & 0 \\ \alpha & \rho_1 & \rho_2 \\ 0 & 1 & 0 \end{bmatrix}, \quad C= \begin{bmatrix} 0 \\ \sigma \\ 0 \end{bmatrix}, \quad \text{and} \quad U = \begin{bmatrix} 0 & 1 & 0 \end{bmatrix}\end{split}\]

The LQ Approach

Previously we solved the permanent income model by solving a system of linear expectational difference equations subject to two boundary conditions

Here we solve the same model using LQ methods based on dynamic programming

After confirming that answers produced by the two methods agree, we apply QuantEcon’s LinearStateSpace class to illustrate features of the model

Why solve a model in two distinct ways?

Our answer is that by doing so we gather insights about the structure of the model

Our earlier approach based on solving a system of expectational difference equations brought to the fore the role of the consumer’s expectations about future nonfinancial income

On the other hand, formulating the model in terms of an LQ dynamic programming problem reminds us that

  • finding the state (of a dynamic programming problem) is an art, and
  • iterations on a Bellman equation implicitly jointly solve a forecasting problem and a control problem

The LQ Problem

Recall from our lecture on LQ theory that the optimal linear regulator problem is to choose a decision rule for \(u_t\) to minimize

\[\mathbb E \sum_{t=0}^\infty \beta^t \{x'_t R x_t+ u'_t Q u_t\},\]

subject to \(x_0\) given, and the law of motion

(5)\[x_{t+1} = A_0 x_t+ B_0 u_t+ C_0 w_{t+1}, \qquad t\geq 0,\]

where \(w_{t+1}\) is iid with mean vector zero and \(\mathbb E w_t w'_t= I\)

The subscripts in \(A_0, B_0, C_0\) are to avoid clashing with notation in (4)

The value function for this problem is \(v(x) = - x'Px - d\), where

The optimal policy is \(u_t = -Fx_t\), where \(F := \beta (Q+\beta B'PB)^{-1} B'PA\)

Under an optimal decision rule \(F\), the state vector \(x_t\) evolves according to \(x_{t+1} = (A-BF) x_t + C w_{t+1}\)

Mapping into the LQ framework

To map into the LQ framework, we’ll use

\[\begin{split}x_t := \begin{bmatrix} z_t \\ b_t \end{bmatrix} = \begin{bmatrix} 1 \\ y_t \\ y_{t-1} \\ b_t \end{bmatrix}\end{split}\]

as the state vector and \(u_t := c_t - \gamma\) as the control

With this notation and \(U_\gamma := \begin{bmatrix} \gamma & 0 & 0 \end{bmatrix}\), we can write the state dynamics as in (5) when

\[\begin{split}A_0 := \begin{bmatrix} A & 0 \\ (1 + r)(U_\gamma - U) & 1 + r \end{bmatrix} \quad B_0 := \begin{bmatrix} 0 \\ 1 + r \end{bmatrix} \quad \text{and} \quad C_0 := \begin{bmatrix} C \\ 0 \end{bmatrix} w_{t+1}\end{split}\]

Please confirm for yourself that, with these definitions, the LQ dynamics (5) match the dynamics of \(z_t\) and \(b_t\) described above

To map utility into the quadratic form \(x_t' R x_t + u_t'Q u_t\) we can set

  • \(Q := 1\) (remember that we are minimizing) and
  • \(R :=\) a \(4 \times 4\) matrix of zeros

However, there is one problem remaining

We have no direct way to capture the non-recursive restriction (3) on the debt sequence \(\{b_t\}\) from within the LQ framework

To try to enforce it, we’re going to use a trick: put a small penalty on \(b_t^2\) in the criterion function

In the present setting, this means adding a small entry \(\epsilon > 0\) in the \((4,4)\) position of \(R\)

That will induce a (hopefully) small approximation error in the decision rule

We’ll check whether it really is small numerically soon


Let’s write some code to solve the model

The code can be executed sequential in Jupyter notebook cells or in a script

We assume the following imports

using QuantEcon
using PyPlot

One comment before we start is that the bliss level of consumption \(\gamma\) in the utility function has no effect on the optimal decision rule

We saw this in the previous premanent income lecture

The reason is that it drops out of the Euler equation for consumption

In what follows we set it equal to unity

The exogeneous income process

First we create the objects for the optimal linear regulator

# Set parameters
alpha, beta, rho1, rho2, sigma = 10.0, 0.95, 0.9, 0.0, 1.0

R = 1/beta
A = [1.0 0.0 0.0;
    alpha rho1 rho2;
    0.0 1.0 0.0]
C = [0.0; sigma; 0.0]''
G = [0.0 1.0 0.0]

# Form LinearStateSpace system and pull off steady state moments
mu_z0 = [1.0, 0.0, 0.0]
sig_z0 = zeros(3, 3)
Lz = LSS(A, C, G, mu_0=mu_z0, Sigma_0=sig_z0)
muz, muy, Sigz, Sigy = stationary_distributions(Lz)

# Mean vector of state for the savings problem
mxo = [muz; 0.0]

# Create stationary covariance matrix of x -- start everyone off at b=0
a1 = zeros(3, 1)
aa = hcat(Sigz, a1)
bb = zeros(1, 4)
sxo = vcat(aa, bb)

# These choices will initialize the state vector of an individual at zero debt
# and the ergodic distribution of the endowment process. Use these to create
# the Bewley economy.
mxbewley = mxo
sxbewley = sxo

The next step is to create the matrices for the LQ system

A12 = zeros(3,1)
ALQ_l = hcat(A, A12)
ALQ_r = [0 -R 0 R]
ALQ = vcat(ALQ_l, ALQ_r)

RLQ = [0.0 0.0 0.0 0.0;
        0.0 0.0 0.0 0.0;
        0.0 0.0 0.0 0.0;
        0.0 0.0 0.0 1e-9]

QLQ = 1.0
BLQ = [0.0; 0.0; 0.0; R]
CLQ = [0.0; sigma; 0.0; 0.0]
betaLQ = beta

Let’s print these out and have a look at them

println("A = $ALQ")
println("B = $BLQ")
println("R = $RLQ")
println("Q = $QLQ")
A = [1.0 0.0 0.0 0.0; 10.0 0.9 0.0 0.0; 0.0 1.0 0.0 0.0; 0.0 -1.05263 0.0 1.05263]
B = [0.0,0.0,0.0,1.05263]
R = [0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 0.0; 0.0 0.0 0.0 1.0e-9]
Q = 1.0

Now create the appropriate instance of an LQ model


We’ll save the implied optimal policy function soon compare them with what we get by employing an alternative solution method

P,F,d=stationary_values(LQPI) # Compute value function and decision rule
ABF = ALQ - BLQ*F  #  Form closed loop system

Comparison with the difference equation approach

In our first lecture on the infinite horizon permanent income problem we used a different solution method

The method was based around

  • deducing the Euler equations that are the first-order conditions with respect to consumption and savings,
  • using the budget constraints and boundary condition complete a system of expectational linear difference equations
  • and then solving those equations to obtain the solution

Expressed in state space notation, the solution took the form

\[\begin{split}\begin{aligned} z_{t+1} & = A z_t + C w_{t+1} \\ b_{t+1} & = b_t + U [ (I -\beta A)^{-1} (A - I) ] z_t \\ y_t & = U z_t \\ c_t & = (1-\beta) [ U (I-\beta A)^{-1} z_t - b_t ] \end{aligned}\end{split}\]

Now we’ll apply the formulas in this system

# Use the above formulas to create the optimal policies for b_{t+1} and c_t
b_pol = G*(inv(eye(3, 3) - beta*A))*(A - eye(3, 3))
c_pol = (1 - beta)*(G*inv(eye(3, 3) - beta*A))

#Create the A matrix for a LinearStateSpace instance
A_LSS1 = vcat(A, b_pol)
A_LSS2 = [0, 0, 0, 1]
A_LSS = hcat(A_LSS1, A_LSS2)

# Create the C matrix for LSS methods
C_LSS = vcat(C, 0)

# Create the G matrix for LSS methods
G_LSS1 = vcat(G, c_pol)
G_LSS2 = vcat(0, -(1 - beta))
G_LSS = hcat(G_LSS1, G_LSS2)

# Use the following values to start everyone off at b=0, initial incomes zero
mu_0 = [1.0, 0.0, 0.0, 0.0]
sigma_0 = zeros(4, 4)

A_LSS calculated as we have here should equal ABF calculated above using the LQ model

4×4 Array{Float64,2}:
  0.0         0.0         0.0   0.0
  0.0         0.0         0.0   0.0
  0.0         0.0         0.0   0.0
 -9.51248e-6  9.51248e-8  0.0  -2.0e-8

Now compare pertinent elements of c_pol and F

[65.5172 0.344828 0.0][65.5172 0.344828 -0.0 -0.05]

We have verified that the two methods give the same solution

Now let’s create instances of the LinearStateSpace class and use it to do some interesting experiments

To do this, we’ll use the outcomes from our second method

Two Example Economies

In the spirit of Bewley models [Bew86], we’ll generate panels of consumers

We’ll study two examples

  • In each of these examples, there is a continuum of ex ante identical consumers

The examples differ only in the initial states with which we endow the consumers

All other parameter values are kept the same in the two examples

  • In the first example, all consumers begin with zero nonfinancial income and zero debt

    • The consumers are thus ex ante identical
  • In the second example, while all begin with zero debt, we draw their initial income levels from the invariant distribution of financial income

    • Consumers are ex ante heterogeneous

In the first example, consumers’ nonfinancial income paths display prounounced transients early in the sample that will affect outcomes in striking ways

Those transient effects will not be present in the second example

We use methods that the LinearStateSpace class contains to simulate the model with our first set of intitial conditions

  • 25 paths of the exogenous non-financial income process and the associated consumption and debt paths.

In the first set of graphs, the darker lines depict one particular sample path, while the lighter lines indicate the other 24 paths

  • A second graph that plots a collection of simulations against the population distribution that we extract from the LinearStateSpace instance LSS

Comparing sample paths with population distributions at each date \(t\) is a useful exercise—see our discussion of the laws of large numbers

lss = LSS(A_LSS, C_LSS, G_LSS, mu_0=mu_0, Sigma_0=sigma_0)

Population and sample panels

In the code below, we use the LinearStateSpace class to

  • compute and plot population quantiles of the distributions of consumption and debt for a population of consumers
  • simulate a group of 25 consumers and plot sample paths on the same graph as the population distribution
This function takes initial conditions (m0, s0) and uses the Linear State Space
class from QuantEcon to  simulate an economy `npaths` times for `T` periods.
It then uses that information to generate some graphs related to the discussion
function income_consumption_debt_series(A, C, G, m0, s0, T=150, npaths=25)

    lss = LSS(A, C, G, mu_0=m0, Sigma_0=s0)

    # Simulation/Moment Parameters
    moment_generator = moment_sequence(lss)

    # Simulate various paths
    bsim = Array{Float64}(npaths, T)
    csim = Array{Float64}(npaths, T)
    ysim = Array{Float64}(npaths, T)

    for i = 1:npaths
        sims = simulate(lss,T)
        bsim[i, :] = sims[1][end, :]
        csim[i, :] = sims[2][2, :]
        ysim[i, :] = sims[2][1, :]

    # Get the moments
    cons_mean = Array{Float64}(T)
    cons_var = Array{Float64}(T)
    debt_mean = Array{Float64}(T)
    debt_var = Array{Float64}(T)
    state = start(moment_generator)
    for t = 1:T
        (mu_x, mu_y, sig_x, sig_y),state=
             next(moment_generator, state)
        cons_mean[t], cons_var[t] = mu_y[2], sig_y[2, 2]
        debt_mean[t], debt_var[t] = mu_x[4], sig_x[4, 4]
    return bsim, csim, ysim, cons_mean, cons_var, debt_mean, debt_var

function consumption_income_debt_figure(bsim, csim, ysim)

    # Get T
    T =  size(bsim,2)

    # Create first figure
    fig, ax = subplots(2, 1, figsize=(10, 8))
    xvals = 1:T

    # Plot consumption and income
    ax[1][:plot](csim[1, :], label="c", color="b")
    ax[1][:plot](ysim[1, :], label="y", color="g")
    ax[1][:plot](csim', alpha=.1, color="b")
    ax[1][:plot](ysim', alpha=.1, color="g")
    ax[1][:set_ylabel]("y and c")

    # Plot debt
    ax[2][:plot](bsim[1, :], label="b", color="r")
    ax[2][:plot](bsim', alpha=.1, color="r")

    fig[:suptitle]("Nonfinancial Income, Consumption, and Debt")


function consumption_debt_fanchart(csim, cons_mean, cons_var,
                              bsim, debt_mean, debt_var)
    # Get T
    T =  size(bsim,2)

    # Create Percentiles of cross-section distributions
    cmean = mean(cons_mean)
    c90 = 1.65*sqrt.(cons_var)
    c95 = 1.96*sqrt.(cons_var)
    c_perc_95p, c_perc_95m = cons_mean + c95, cons_mean - c95
    c_perc_90p, c_perc_90m = cons_mean + c90, cons_mean - c90

    # Create Percentiles of cross-section distributions
    dmean = mean(debt_mean)
    d90 = 1.65*sqrt.(debt_var)
    d95 = 1.96*sqrt.(debt_var)
    d_perc_95p, d_perc_95m = debt_mean + d95, debt_mean - d95
    d_perc_90p, d_perc_90m = debt_mean + d90, debt_mean - d90

    # Create second figure
    fig2, ax2 = subplots(2, 1, figsize=(10, 8))
    xvals = 1:T

    fig2[:suptitle]("Consumption/Debt over time")

    # Consumption fan
    ax2[1][:plot](xvals, cons_mean, color="k")
    ax2[1][:plot](xvals, csim', color="k", alpha=.25)
    ax2[1][:fill_between](xvals, c_perc_95m, c_perc_95p, alpha=.25, color="b")
    ax2[1][:fill_between](xvals, c_perc_90m, c_perc_90p, alpha=.25, color="r")
    ax2[1][:set_ylim]((cmean-15, cmean+15))

    # Debt fan
    ax2[2][:plot](xvals, debt_mean, color="k")
    ax2[2][:plot](xvals, bsim', color="k", alpha=.25)
    ax2[2][:fill_between](xvals, d_perc_95m, d_perc_95p, alpha=.25, color="b")
    ax2[2][:fill_between](xvals, d_perc_90m, d_perc_90p, alpha=.25, color="r")
#     ax2[1].set_ylim()

Now let’s create figures with initial conditions of zero for \(y_0\) and \(b_0\)

out = income_consumption_debt_series(A_LSS, C_LSS, G_LSS, mu_0, sigma_0)
bsim0, csim0, ysim0 = out[1:3]
cons_mean0, cons_var0, debt_mean0, debt_var0 = out[4:end]

fig_0 = consumption_income_debt_figure(bsim0, csim0, ysim0)
fig_02 = consumption_debt_fanchart(csim0, cons_mean0, cons_var0,
                                   bsim0, debt_mean0, debt_var0)

Here are the two figures


Here is what is going on in the above graphs

For our simulation, we have set initial conditions \(b_0 = y_{-1} = y_{-2} = 0\)

Because \(y_{-1} = y_{-2} = 0\), nonfinancial income \(y_t\) starts far below its stationary mean \(\mu_{y, \infty}\) and rises early in each simulation

Recall that we can represent the optimal decision rule for consumption in terms of the co-integrating relationship

(6)\[(1-\beta) b_t + c_t = (1-\beta) E_t \sum_{j=0}^\infty \beta^j y_{t+j}\]

So at time \(0\) we have

\[c_0 = (1-\beta) E_0 \sum_{t=0}^\infty \beta^j y_{t}\]

This tells us that consumption starts at the value of an annuity from the expected discounted value of nonfinancial income

To support that level of consumption, the consumer borrows a lot early and consequently builds up substantial debt

In fact, he or she incurs so much debt that eventually, in the stochastic steady state, he consumes less each period than his income

He uses the gap between consumption and income mostly to service the interest payments due on his debt

Thus, when we look at the panel of debt in the accompanying graph, we see that this is a group of ex ante indentical people each of whom starts with zero debt

All of them accumulate debt in anticipation of rising nonfinancial income

They expect their nonfinancial income to rise toward the invariant distribution of income, a consequence of our having started them at \(y_{-1} = y_{-2} = 0\)

Cointegration residual

The following figure plots realizations of the left side of (6), which, as discussed in our last lecture, is called the cointegrating residual

The right side can be thought of as an annuity payment on the expected present value of future income \(E_t \sum_{j=0}^\infty \beta^j y_{t+j}\)

Early along a realization, \(c_t\) is approximately constant while \((1-\beta) b_t\) and \((1-\beta) E_t \sum_{j=0}^\infty \beta^j y_{t+j}\) both rise markedly as the household’s present value of income and borrowing rise pretty much together

This example illustrates the following point: the definition of cointegration implies that the cointegrating residual is asymptotically covariance stationary, not covariance stationary

The cointegrating residual for the specification with zero income and zero debt initially has a notable transient component that dominates its behavior early in the sample.

By specifying different initial conditions, we shall remove this transient in our second example to be presented below

Plots the cointegration
function cointegration_figure(bsim, csim)
    # Create figure
    fig, ax = subplots(figsize=(10, 8))
    ax[:plot]((1-beta)*bsim[1, :] + csim[1, :], color="k")
    ax[:plot]((1-beta)*bsim' + csim', color="k", alpha=.1)

    fig[:suptitle]("Cointegration of Assets and Consumption")
cointegration_figure(bsim0, csim0)

Here’s the resulting figure


A “borrowers and lenders” closed economy

When we set \(y_{-1} = y_{-2} = 0\) and \(b_0 =0\) in the preceding exercise, we make debt “head north” early in the sample

Average debt rises and approaches asymptote

We can regard these as outcomes of a “small open economy” that borrows from abroad at the fixed gross interest rate \(r\) in anticipation of rising incomes

So with the economic primitives set as above, the economy converges to a steady state in which there is an excess aggregate supply of risk-free loans at a gross interest rate of \(R\)

This excess supply is filled by “foreigner lenders” willing to make those loans

We can use virtually the same code to rig a “poor man’s Bewley [Bew86] model” in the following way

  • as before, we start everyone at \(b_0 = 0\)
  • But instead of starting everyone at \(y_{-1} = y_{-2} = 0\), we draw \(\begin{bmatrix} y_{-1} \\ y_{-2} \end{bmatrix}\) from the invariant distribution of the \(\{y_t\}\) process

This rigs a closed economy in which people are borrowing and lending with each other at a gross risk-free interest rate of \(R = \beta^{-1}\)

Across the group of people being analyzed, risk-free loans are in zero excess supply

We have arranged primitives so that \(R = \beta^{-1}\) clears the market for risk-free loans at zero aggregate excess supply

So the risk-free loans are being made from one person to another within our closed set of agent

There is no need for foreigners to lend to our group

Let’s have a look at the corresponding figures

out = income_consumption_debt_series(A_LSS, C_LSS, G_LSS, mxbewley, sxbewley)
bsimb, csimb, ysimb = out[1:3]
cons_meanb, cons_varb, debt_meanb, debt_varb = out[4:end]

consumption_income_debt_figure(bsimb, csimb, ysimb)
consumption_debt_fanchart(csimb, cons_meanb, cons_varb,
                          bsimb, debt_meanb, debt_varb)

The graphs confirm the following outcomes:

  • As before, the consumption distribution spreads out over time

But now there is some initial dispersion because there is ex ante heterogeneity in the initial draws of \(\begin{bmatrix} y_{-1} \\ y_{-2} \end{bmatrix}\)

  • As before, the cross-section distribution of debt spreads out over time
  • Unlike before, the average level of debt stays at zero, reflecting that this is a closed borrower-and-lender economy
  • Now the cointegrating residual seems stationary, and not just asymptotically stationary

Let’s have a look at the cointegration figure

cointegration_figure(bsimb, csimb)