# The Permanent Income Model¶

Contents

## Overview¶

This lecture describes a rational expectations version of the famous permanent income model of Friedman [Fri56]

Hall cast Friedman’s model within a linear-quadratic setting [Hal78]

Like Hall, we formulate an infinite-horizon linear-quadratic savings problem

We use the model as a vehicle for illustrating

- alternative formulations of the
*state*of a dynamic system - the idea of
*cointegration* - impulse response functions
- the idea that changes in consumption are useful as predictors of movements in income

Background readings on the linear-quadratic-Gaussian permanent income model are Robert Hall’s [Hal78] and chapter 2 of [LS12]

## The Savings Problem¶

In this section we state and solve the savings and consumption problem faced by the consumer

### Preliminaries¶

The discussion below requires a casual familiarity with martingales

A discrete time martingale is a stochastic process (i.e., a sequence of random variables) \(\{X_t\}\) with finite mean and satisfying

Here \(\mathbb{E}_t := \mathbb{E}[ \cdot \,|\, \fF_t]\) is a mathematical expectation conditional on the time \(t\)
*information set* \(\fF_t\)

The latter is just a collection of random variables that the modeler declares to be visible at \(t\)

- When not explicitly defined, it is usually understood that \(\fF_t = \{X_t, X_{t-1}, \ldots, X_0\}\)

Martingales have the feature that the history of past outcomes provides no predictive power for changes between current and future outcomes

For example, the current wealth of a gambler engaged in a “fair game” has this property

One common class of martingales is the family of *random walks*

A *random walk* is a stochastic process \(\{X_t\}\) that satisfies

for some iid zero mean *innovation* sequence \(\{w_t\}\)

Evidently \(X_t\) can also be expressed as

Not every martingale arises as a random walk (see, for example, Wald’s martingale)

### The Decision Problem¶

A consumer has preferences over consumption streams that are ordered by the utility functional

where

- \(\mathbb{E}_t\) is the mathematical expectation conditioned on the consumer’s time \(t\) information
- \(c_t\) is time \(t\) consumption
- \(u\) is a strictly concave one-period utility function
- \(\beta \in (0,1)\) is a discount factor

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

Here

- \(y_t\) is an exogenous endowment process
- \(r > 0\) is the risk-free interest rate
- \(b_t\) is one-period risk-free debt maturing at \(t\)

The consumer also faces initial conditions \(b_0\) and \(y_0\), which can be fixed or random

### Assumptions¶

For the remainder of this lecture, we follow Friedman and Hall in assuming that \((1 + r)^{-1} = \beta\)

Regarding the endowment process, we assume it has the *state-space representation*

where

- \(\{w_t\}\) is an iid vector process with \(\mathbb{E} w_t = 0\) and \(\mathbb{E} w_t w_t' = I\)
- the spectral radius of \(A\) satisfies \(\rho(A) < 1/\beta\)
- \(U\) is a selection vector that pins down \(y_t\) as a particular linear combination of the elements of \(z_t\).

The restriction on \(\rho(A)\) prevents income from growing so fast that some discounted geometric sums of some infinite sequences below become infinite

Regarding preferences, we assume the quadratic utility function

where \(\gamma\) is a bliss level of consumption

Note

Along with this quadratic utility specification, we allow consumption to be negative. However, by choosing parameters appropriately, we can make the probability that the model generates negative consumption paths as low as desired.

Finally, we impose the *no Ponzi scheme* condition

This condition rules out an always-borrow scheme that would allow the household to enjoy bliss consumption forever

### First Order Conditions¶

First-order conditions for maximizing (1) subject to (2) are

These equations are also known as the *Euler equations* for the model

If you’re not sure where they come from, you can find a proof sketch in the appendix

With our quadratic preference specification, (5) has the striking implication that consumption follows a martingale:

(In fact quadratic preferences are *necessary* for this conclusion [1])

One way to interpret (6) is that consumption will only change when “new information” about permanent income is revealed

These ideas will be clarified below

### The Optimal Decision Rule¶

Now let’s deduce the optimal decision rule [2]

Note

One way to solve the consumer’s problem is to apply *dynamic programming*
as in *this lecture*. We do this later. But first we use
an alternative approach that is revealing and shows the work that dynamic
programming does for us automatically

In doing so, we need to combine

- the optimality condition (6)
- the period-by-period budget constraint (2), and
- the boundary condition (4)

To accomplish this, observe first that (4) implies \(\lim_{t \to \infty} \beta^t b_{t+1}= 0\)

Using this restriction on the debt path and solving (2) forward yields

Take conditional expectations on both sides of (7) and use the martingale property of consumption and the *law of iterated expectations* to deduce

Expressed in terms of \(c_t\) we get

where the last equality uses \((1 + r) \beta = 1\)

These last two equations assert that consumption equals *economic income*

*financial wealth*equals \(-b_t\)*non-financial*wealth equals \(\sum_{j=0}^\infty \beta^j \mathbb{E}_t [y_{t+j}]\)- A
*marginal propensity to consume out of wealth*equals the interest factor \(\frac{r}{1+r}\) *economic income*equals- a constant marginal propensity to consume times the sum of nonfinancial wealth and financial wealth
- the amount the household can consume while leaving its wealth intact

#### Reacting to the state¶

The *state* vector confronting the household at \(t\) is \(\begin{bmatrix} b_t & z_t \end{bmatrix}\)

Here

- \(z_t\) is an
*exogenous*component, unaffected by household behavior - \(b_t\) is an
*endogenous*component (since it depends on the decision rule)

Note that \(z_t\) contains all variables useful for forecasting the household’s future endowment

It seems likely that current decisions \(c_t\) and \(b_{t+1}\) should be expressible as functions of \(z_t\) and \(b_t\)

This is indeed the case

In fact, from this discussion we see that

Combining this with (9) gives

Using this equality to eliminate \(c_t\) in the budget constraint (2) gives

To get from the second last to the last expression in this chain of equalities is not trivial

Try using the fact that \((1 + r) \beta = 1\) and \((I - \beta A)^{-1} = \sum_{j=0}^{\infty} \beta^j A^j\)

We’ve now successfully written \(c_t\) and \(b_{t+1}\) as functions of \(b_t\) and \(z_t\)

#### A State-Space Representation¶

We can summarize our dynamics in the form of a linear state-space system governing consumption, debt and income:

To write this more succinctly, let

and

Then we can express equation (11) as

We can use the following formulas from *state-space representation* to compute population mean \(\mu_t = \mathbb{E} x_t\) and covariance \(\Sigma_t := \mathbb{E} [ (x_t - \mu_t) (x_t - \mu_t)']\)

We can then compute the mean and covariance of \(\tilde y_t\) from

#### A Simple Example with iid Income¶

To gain some preliminary intuition on the implications of (11), let’s look at a highly stylized example where income is just iid

(Later examples will investigate more realistic income streams)

In particular, let \(\{w_t\}_{t = 1}^{\infty}\) be iid and scalar standard normal, and let

Finally, let \(b_0 = z^1_0 = 0\)

Under these assumptions we have \(y_t = \mu + \sigma w_t \sim N(\mu, \sigma^2)\)

Further, if you work through the state space representation, you will see that

Thus income is iid and debt and consumption are both Gaussian random walks

Defining assets as \(-b_t\), we see that assets are just the cumulative sum of unanticipated income prior to the present date

The next figure shows a typical realization with \(r = 0.05\), \(\mu = 1\) and \(\sigma = 0.15\)

Observe that consumption is considerably smoother than income

The figure below shows the consumption paths of 250 consumers with independent income streams

The code for these figures can be found in perm_inc_figs.jl

## Alternative Representations¶

In this section we shed more light on the evolution of savings, debt and consumption by representing their dynamics in several different ways

### Hall’s Representation¶

Hall [Hal78] suggests a sharp way to summarize the implications of LQ permanent income theory

First, to represent the solution for \(b_t\), shift (9) forward one period and eliminate \(b_{t+1}\) by using (2) to obtain

If we add and subtract \(\beta^{-1} (1-\beta) \sum_{j=0}^\infty \beta^j \mathbb{E}_t y_{t+j}\) from the right side of the preceding equation and rearrange, we obtain

The right side is the time \(t+1\) *innovation to the expected present value* of the endowment process \(\{y_t\}\)

We can represent the optimal decision rule for \(c_t, b_{t+1}\) in the form of (16) and (8), which is repeated here:

Equation (17) asserts that the household’s debt due at \(t\) equals the expected present value of its endowment minus the expected present value of its consumption stream

A high debt thus indicates a large expected present value of surpluses \(y_t - c_t\)

Recalling again our discussion on forecasting geometric sums, we have

Using these formulas together with (3) and substituting into (16) and (17) gives the following representation for the consumer’s optimum decision rule:

Representation (18) makes clear that

The state can be taken as \((c_t, z_t)\)

- The endogenous part is \(c_t\) and the exogenous part is \(z_t\)
- Debt \(b_t\) has disappeared as a component of the state because it is encoded in \(c_t\)

Consumption is a random walk with innovation \((1-\beta) U (I-\beta A)^{-1} C w_{t+1}\)

- This is a more explicit representation of the martingale result in (6)

### Cointegration¶

Representation (18) reveals that the joint process \(\{c_t, b_t\}\) possesses the property that Engle and Granger [EG87] called cointegration

Cointegration is a tool that allows us to apply powerful results from the theory of stationary processes to (certain transformations of) nonstationary models

To clarify cointegration in the present context, suppose that \(z_t\) is asymptotically stationary [4]

Despite this, both \(c_t\) and \(b_t\) will be non-stationary because they have unit roots (see (11) for \(b_t\))

Nevertheless, there is a linear combination of \(c_t, b_t\) that *is* asymptotically stationary

In particular, from the second equality in (18) we have

Hence the linear combination \((1-\beta) b_t + c_t\) is asymptotically stationary

Accordingly, Granger and Engle would call \(\begin{bmatrix} (1-\beta) & 1 \end{bmatrix}\) a *cointegrating vector* for the state

When applied to the nonstationary vector process \(\begin{bmatrix} b_t & c_t \end{bmatrix}'\), it yields a process that is asymptotically stationary

Equation (19) can be arranged to take the form

Equation (20) asserts that the *cointegrating residual* on the left side equals the conditional expectation of the geometric sum of future incomes on the right [6]

### Cross-Sectional Implications¶

Consider again (18), this time in light of our discussion of
distribution dynamics in the *lecture on linear systems*

The dynamics of \(c_t\) are given by

or

The unit root affecting \(c_t\) causes the time \(t\) variance of \(c_t\) to grow linearly with \(t\)

In particular, since \(\{ \hat w_t \}\) is iid, we have

when

Assuming that \(\hat \sigma > 0\), this means that \(\{c_t\}\) has no asymptotic distribution

Let’s consider what this means for a cross-section of ex ante identical households born at time \(0\)

Let the distribution of \(c_0\) represent the cross-section of initial consumption values

Equation (22) tells us that the distribution of \(c_t\) spreads out over time at a rate proportional to \(t\)

A number of different studies have investigated this prediction (see, e.g., [DP94], [STY04])

### Impulse Response Functions¶

Impulse response functions measure the change in a dynamic system subject to a given impulse (i.e., temporary shock)

The impulse response function of \(\{c_t\}\) to the innovation \(\{w_t\}\) is a box

In particular, the response of \(c_{t+j}\) to a unit increase in the innovation \(w_{t+1}\) is \((1-\beta) U (I -\beta A)^{-1} C\) for all \(j \geq 1\)

### Moving Average Representation¶

It’s useful to express the innovation to the expected present value of the endowment process in terms of a moving average representation for income \(y_t\)

The endowment process defined by (3) has the moving average representation

where

- \(d(L) = \sum_{j=0}^\infty d_j L^j\) for some sequence \(d_j\), where \(L\) is the lag operator [3]
- at time \(t\), the household has an information set [5] \(w^t = [w_t, w_{t-1}, \ldots ]\)

Notice that

It follows that

The object \(d(\beta)\) is the *present value of the moving average coefficients* in the representation for the endowment process \(y_t\)

## Two Classic Examples¶

We illustrate some of the preceding ideas with the following two examples

In both examples, the endowment follows the process \(y_t = z_{1t} + z_{2t}\) where

Here

- \(w_{t+1}\) is an iid \(2 \times 1\) process distributed as \(N(0,I)\)
- \(z_{1t}\) is a permanent component of \(y_t\)
- \(z_{2t}\) is a purely transitory component

### Example 1¶

Assume as before that the consumer observes the state \(z_t\) at time \(t\)

In view of (18) we have

Formula (26) shows how an increment \(\sigma_1 w_{1t+1}\) to the permanent component of income \(z_{1t+1}\) leads to

- a permanent one-for-one increase in consumption and
- no increase in savings \(-b_{t+1}\)

But the purely transitory component of income \(\sigma_2 w_{2t+1}\) leads to a permanent increment in consumption by a fraction \(1-\beta\) of transitory income

The remaining fraction \(\beta\) is saved, leading to a permanent increment in \(-b_{t+1}\)

Application of the formula for debt in (11) to this example shows that

This confirms that none of \(\sigma_1 w_{1t}\) is saved, while all of \(\sigma_2 w_{2t}\) is saved

The next figure illustrates these very different reactions to transitory and permanent income shocks using impulse-response functions

The code for generating this figure is in file `perm_income/perm_inc_ir.jl`

from the applications repository, as shown below

```
#=
@author : Spencer Lyon
Victoria Gregory
@date: 07/09/2014
=#
using Plots
pyplot()
const r = 0.05
const beta = 1.0 / (1.0 + r)
const T = 20 # Time horizon
const S = 5 # Impulse date
const sigma1 = 0.15
const sigma2 = 0.15
function time_path(permanent=false)
w1 = zeros(T+1)
w2 = zeros(T+1)
b = zeros(T+1)
c = zeros(T+1)
if permanent === false
w2[S+2] = 1.0
else
w1[S+2] = 1.0
end
for t=2:T
b[t+1] = b[t] - sigma2 * w2[t]
c[t+1] = c[t] + sigma1 * w1[t+1] + (1 - beta) * sigma2 * w2[t+1]
end
return b, c
end
function main()
L = 0.175
b1, c1 = time_path(false)
b2, c2 = time_path(true)
p = plot(0:T, [c1 c2 b1 b2], layout=(2, 1),
color=[:green :green :blue :blue],
label=["consumption" "consumption" "debt" "debt"])
t = ["impulse-response, transitory income shock"
"impulse-response, permanent income shock"]
plot!(title=t', xlabel="Time", ylims=(-L, L), legend=[:topright :bottomright])
vline!([S S], color=:black, layout=(2, 1), label="")
return p
end
```

### Example 2¶

Assume now that at time \(t\) the consumer observes \(y_t\), and its history up to \(t\), but not \(z_t\)

Under this assumption, it is appropriate to use an *innovation representation* to form \(A, C, U\) in (18)

The discussion in sections 2.9.1 and 2.11.3 of [LS12] shows that the pertinent state space representation for \(y_t\) is

where

- \(K :=\) the stationary Kalman gain
- \(a_t := y_t - E [ y_t \,|\, y_{t-1}, \ldots, y_0]\)

In the same discussion in [LS12] it is shown that \(K \in [0,1]\) and that \(K\) increases as \(\sigma_1/\sigma_2\) does

In other words, as the ratio of the standard deviation of the permanent shock to that of the transitory shock increases

Applying formulas (18) implies

where the endowment process can now be represented in terms of the univariate innovation to \(y_t\) as

Equation (29) indicates that the consumer regards

- fraction \(K\) of an innovation \(a_{t+1}\) to \(y_{t+1}\) as
*permanent* - fraction \(1-K\) as purely transitory

The consumer permanently increases his consumption by the full amount of his estimate of the permanent part of \(a_{t+1}\), but by only \((1-\beta)\) times his estimate of the purely transitory part of \(a_{t+1}\)

Therefore, in total he permanently increments his consumption by a fraction \(K + (1-\beta) (1-K) = 1 - \beta (1-K)\) of \(a_{t+1}\)

He saves the remaining fraction \(\beta (1-K)\)

According to equation (29), the first difference of income is a first-order moving average

Equation (28) asserts that the first difference of consumption is iid

Application of formula to this example shows that

This indicates how the fraction \(K\) of the innovation to \(y_t\) that is regarded as permanent influences the fraction of the innovation that is saved

## Further Reading¶

The model described above significantly changed how economists think about consumption

At the same time, it’s generally recognized that Hall’s version of the permanent income hypothesis fails to capture all aspects of the consumption/savings data

For example, liquidity constraints and buffer stock savings appear to be important

Further discussion can be found in, e.g., [HM82], [Par99], [Dea91], [Car01]

## Appendix: The Euler Equation¶

Where does the first order condition (5) come from?

Here we’ll give a proof for the two period case, which is representative of the general argument

The finite horizon equivalent of the no-Ponzi condition is that the agent cannot end her life in debt, so \(b_2 = 0\)

From the budget constraint (2) we then have

Here \(b_0\) and \(y_0\) are given constants

Subsituting these constraints into our two period objective \(u(c_0) + \beta \mathbb{E}_0 [u(c_1)]\) gives

You will be able to verify that the first order condition is

Using \(\beta R = 1\) gives (5) in the two period case

The proof for the general case is similar

Footnotes

[1] | A linear marginal utility is essential for deriving (6) from (5). Suppose instead that we had imposed the following more standard assumptions on the utility function: \(u'(c) >0, u''(c)<0, u'''(c) > 0\) and required that \(c \geq 0\). The Euler equation remains (5). But the fact that \(u''' <0\) implies via Jensen’s inequality that \(\mathbb{E}_t [u'(c_{t+1})] > u'(\mathbb{E}_t [c_{t+1}])\). This inequality together with (5) implies that \(\mathbb{E}_t [c_{t+1}] > c_t\) (consumption is said to be a ‘submartingale’), so that consumption stochastically diverges to \(+\infty\). The consumer’s savings also diverge to \(+\infty\). |

[2] | An optimal decision rule is a map from current state into current actions—in this case, consumption |

[3] | Representation (3) implies that \(d(L) = U (I - A L)^{-1} C\). |

[4] | This would be the case if, for example, the spectral radius of \(A\) is strictly less than one |

[5] | A moving average representation for a process \(y_t\) is said to be fundamental if the linear space spanned by \(y^t\) is equal to the linear space spanned by \(w^t\). A time-invariant innovations representation, attained via the Kalman filter, is by construction fundamental. |

[6] | See [JYC88], [LL01], [LL04] for interesting applications of related ideas. |