Solving Difference Equations
Stochastic Foundations for Finance
Fall 2024
Simple Linear Difference Equations
Introduction
- In finance and economics, we are usually interested in describing the evolution of economic quantities over time.
- For example, we might want to describe how a stock price or an interest rate evolves as time goes by.
- In these notes, time can only take discrete values.
- For example, we might assume that t = 0, 1, 2, 3, \ldots
- In some cases, we might even consider t = \ldots, -2, -1, 0, 1, 2, \ldots
- We will then index the variable of interest by t.
- For example, the stock price at time t is given by S_{t}.
A Simple Difference Equation
- Consider the evolution of y_{t} given by
y_{t} = a + b y_{t-1}
\tag{1} and assume that y_{0} is given.
- We can iterate to find future values of y_{t},
\begin{aligned}
y_{1} & = a + b y_{0}, \\
y_{2} & = a + b y_{1} = a + b (a + b y_{0}) = a + b a + b^{2} y_{0}, \\
y_{3} & = a + b y_{2} = a + b (a + b a + b^{2} y_{0}) = a + b a + b^{2} a + b^{3} y_{0}.
\end{aligned}
- We can see that there is a pattern such that
y_{t} = a (b^{0} + b^{1} + b^{2} + \ldots + b^{t-1}) + b^{t} y_{0}.
A Useful Result
- Consider the sum
S = 1 + b + b^{2} + \ldots + b^{t - 1},
and assume that b \neq 1.
- Multiplying S by b yields
b S = b + b^{2} + b^{3} \ldots + b^{t}.
- Subtracting the second from the first expression allows us to solve for b
S = \frac{1 - b^{t}}{1 - b}.
- If b = 1, it’s easy to see that S = t, and we can show that \lim_{b \rightarrow 1} S = t.
Solving the Difference Equation
- We can now see that
y_{t} = a \left(\frac{1 - b^{t}}{1 - b}\right) + b^{t} y_{0},
is a solution of y_{t} = a + b y_{t-1} given y_{0}.
- If \left\lvert b \right\rvert < 1, we know that \lim_{t \rightarrow \infty} b^{t} = 0.
- If 0 < b < 1 the convergence is monotonic but if -1 < b < 0 the convergence is alternating.
- In finance applications we typically assume that 0 < b < 1 but some interesting economic models use -1 < b < 0.
- Denote by \bar{y} the limit of y_{t} as t goes to infinity, so that
\bar{y} = \lim_{t \rightarrow \infty} y_{t} = \frac{a}{1 - b}.
Expressing the Solution as Deviations from Steady-State
- Note that \bar{y} solves (1) since
\bar{y} = a + b \bar{y}.
- Therefore \bar{y} is a particular solution.
- We can re-express the general solution of (1) as
y_{t} = \bar{y} + b^{t} (y_{0} - \bar{y}).
- The model defined by (1) implies that y_{t} starts at y_{0} and converges exponentially to its steady-state value \bar{y}.
Figure 1: The figure shows solutions to y_{t} = a + b y_{t-1} for different values of y_{0} and b > 0. In the figure, the dashed line represents the steady-state solution given by \bar{y} = a / (1 - b). The parameters a is chosen so that \bar{y} = 1.
Mean-Reversion
- The model defined by (1) exhibits mean-reversion.
- The values of y_{t} converge to their long-run mean \bar{y}.
- Note that we could write it as
\Delta y_{t} = (1 - b) (\bar{y} - y_{t - 1}),
\tag{2} where \Delta y_{t} = y_{t} - y_{t - 1}.
- Assume 0 < b < 1.
- In this case, when y_{t-1} is above \bar{y}, we have that \Delta y_{t} < 0, implying that y_{t} < y_{t - 1}.
- When y_{t-1} is below \bar{y}, the opposite happens and y_{t} > y_{t - 1}.
Figure 2: The figure shows the solution to y_{t} = a + b y_{t-1} when b = -0.90 and \bar{y} = a / (1 - b) = 1. The solution oscillates around its steady-state value until convergence.
Discounting Dividends
- The price of a stock at time t should be equal to the present value of its price next period plus any dividend paid during the period, i.e.,
p_{t} = \frac{p_{t+1} + d_{t+1}}{1 + r}.
\tag{3}
- We assume here that the dividend is paid at the end of the period, meaning p_{t} is the price of the stock after it goes ex-dividend.
- We can iterate forward to find
p_{t} = \frac{d_{t+1}}{1 + r} + \frac{d_{t+2}}{(1 + r)^{2}} + \ldots + \frac{d_{t+n}}{(1 + r)^{n}} + \frac{p_{t+n}}{(1 + r)^{n}}.
\tag{4}
A Note on Bubbles
- Taking limits in (5) we get
p_{t} = \sum_{n = 1}^{\infty} \frac{d_{t+n}}{(1 + r)^{n}} + \lim_{n \rightarrow \infty} \frac{p_{t+n}}{(1 + r)^{n}}.
\tag{5}
- The first component in (5) is the fundamental value of the asset which is the present value of future dividends.
- The second component is a bubble!
- If this term is not zero, the price of the asset can be anything.
- To pin down a unique price for the stock, we assume
\lim_{n \rightarrow \infty} \frac{p_{t+n}}{(1 + r)^{n}} = 0.
- In economics, this is known as the transversality condition.
The Dividend-Discount Model (DDM)
- Imposing the transversality condition, we must have that
p_{t} = \sum_{n = 1}^{\infty} \frac{d_{t+n}}{(1 + r)^{n}}.
\tag{6}
- Equation (6) determines the fundamental value of the stock.
- The fundamental value of the asset at any time is the present value of future dividends.
- In practice, it is quite hard to forecast dividends, so academics and practitioners add assumptions to model how dividends evolve over time.
Constant Dividend Growth
- The Gordon growth model assumes that dividends grow at a constant rate g < r, i.e.,
d_{t+1} = (1 + g) d_{t}.
- According to (6), the price of the stock next period is
p_{t+1} = \sum_{n = 1}^{\infty} \frac{d_{t+1+n}}{(1 + r)^{n}} = (1 + g) \sum_{n = 1}^{\infty} \frac{d_{t+n}}{(1 + r)^{n}} = (1 + g) p_{t}.
- Thus, we can re-write (3) as
p_{t} = \frac{(1 + g) p_{t} + d_{t+1}}{1 + r}.
\tag{7}
The Gordon-Growth Model
- Equation (7) implies
p_{t} = \frac{d_{t+1}}{r - g} = \frac{1 + g}{r - g} d_{t}.
\tag{8}
- Equation (8) ties the price of the asset to fundamentals.
- It is widely used by practitioners in corporate finance, investments and real estate.
- Thus, constant dividend growth implies that price-dividend ratios are constant over time, i.e.,
\frac{p_{t}}{d_{t}} = \frac{1 + g}{r - g}.
Present Value of a Growing Annuity
- We can use (4) and (8) to derive the present value of a growing annuity, i.e.,
PV = \frac{d_{t+1}}{1 + r} + \frac{d_{t+1}(1 + g)}{(1 + r)^{2}} + \ldots + \frac{d_{t+1}(1 + g)^{n-1}}{(1 + r)^{n}}.
- Note that
\begin{aligned}
PV & = p_{t} - \frac{p_{t+n}}{(1 + r)^{n}} = p_{t} - \frac{(1 + g)^{n} p_{t}}{(1 + r)^{n}} \\
& = p_{t} \left(1 - \frac{(1 + g)^{n}}{(1 + r)^{n}}\right) = \frac{d_{t+1}}{r - g} \left(1 - \frac{(1 + g)^{n}}{(1 + r)^{n}}\right).
\end{aligned}
\tag{9}
Application: Two-Stage DDM
Amco last year paid a dividend of $0.20 per share. This dividend is expected to grow at 15% per year for the next three years, then at 10% per year until year 6, after which it is expected to grow at 5% in perpetuity. What is the stock’s value if your required rate of return is 10%?
Dividend |
0.2000 |
0.2300 |
0.2645 |
0.3042 |
0.3346 |
0.3681 |
0.4049 |
Terminal Value |
|
|
|
|
|
8.0971 |
|
Cash Flows |
|
0.2300 |
0.2645 |
0.3042 |
0.3346 |
8.4652 |
|
- The price of the stock is the present value of the cash flows which is $6.14.
Where Does Growth Come From?
- Dividends are paid from earnings, so assume that a fraction 1 - \varphi of earnings is paid as dividends and the rest reinvested in projects generating a return on equity k > r.
- Earnings next period are thus equal to current earnings plus reinvested earnings times the return of equity, i.e.,
e_{t+1} = e_{t} + \varphi e_{t} k = (1 + \varphi k) e_{t},
implying that g = \varphi k.
- In this simple model, the dividend growth rate is equal to reinvested earnings times the return of equity.
Positive NPV Projects
- In which projects should the company invest?
- The price-to-future-earnings ratio is given by
\frac{p_{t}}{e_{t+1}} = \frac{1 - \varphi}{r - \varphi k}.
- We can show that
\frac{d}{d\varphi} \frac{p_{t}}{e_{t+1}} > 0 \Longleftrightarrow k > r.
- The stock price increases if the company reinvests in projects with positive NPV, i.e., projects in which the IRR is greater than the cost of equity.
- Of course, the company cannot reinvest everything forever, otherwise it would be a bubble.
Systems of Linear Difference Equations
Systems of Linear Difference Equations
- The previous model is an example of a system of difference equations.
- For example, we could model the evolution of x_{t} and y_{t} as
\begin{pmatrix}
x_{t} \\ y_{t}
\end{pmatrix}
=
\begin{pmatrix}
a_{1} \\ a_{2}
\end{pmatrix}
+
\begin{pmatrix}
b_{11} & b_{12} \\
b_{21} & b_{22}
\end{pmatrix}
\begin{pmatrix}
x_{t-1} \\ y_{t-1}
\end{pmatrix}.
\tag{10}
- Systems such as (10) can explode, mean-revert or oscillate depending on the values of b_{ij}.
- Nevertheless, it is always possible to iterate forward to solve for future values of x_{t} and y_{t}
Solving the System
- It is convenient to write down (10) in more compact form as
\mathbf{z}_{t} = \mathbf{a} + \mathbf{B} z_{t-1},
\tag{11} where \mathbf{z}_{t} = \begin{pmatrix} x_{t} \\ y_{t} \end{pmatrix}, \mathbf{a} = \begin{pmatrix} a_{1} \\ a_{2} \end{pmatrix}, and \mathbf{B} = \begin{pmatrix} b_{11} & b_{12} \\ b_{21} & b_{22} \end{pmatrix}.
- We can iterate again to find
\mathbf{z}_{t+n} = \mathbf{a} (\mathbf{I} + \mathbf{B} + \mathbf{B}^{2} + \ldots + \mathbf{B}^{n-1}) + \mathbf{B}^{n} z_{0},
\tag{12} where \mathbf{I} = \begin{pmatrix} 1 & 0 \\ 0 & 1 \end{pmatrix} is the identity matrix.
- We can write (12) more compactly as
\mathbf{z}_{t+n} = \mathbf{a} (\mathbf{I} - \mathbf{B}^{n})(\mathbf{I} - \mathbf{B})^{-1} + \mathbf{B}^{n} z_{0}.
\tag{13}
Note on Solutions
- With modern computers it is usually easier to solve the system directly by iterating equation (11).
- Equation (13) will converge to a steady-state solution if
\lim_{n \rightarrow \infty} \mathbf{B}^{n} = 0,
in which case
\bar{\mathbf{z}} = \mathbf{a} (\mathbf{I - \mathbf{B}})^{-1}.
- This will happen whenever the absolute value of the eigenvalues of \mathbf{B} are strictly less than one.
Figure 3: The figure shows the solutions to \begin{pmatrix} x_{t} \\ y_{t} \end{pmatrix} = \begin{pmatrix} 0.5 \\ 0.25 \end{pmatrix} + \begin{pmatrix} 0.85 & -0.3 \\ 0.4 & 0.9 \end{pmatrix} \begin{pmatrix} x_{t-1} \\ y_{t-1} \end{pmatrix}, given x_{0} = y_{0} = 0.
Second Order Difference Equations
- A second order difference equation is defined as
x_{t} = a + b_{1} x_{t-1} + b_{2} x_{t-2},
\tag{14} where x_{0} and x_{1} are given.
- This means you can only start computing x_{t} from t > 1 onwards.
- If we define y_{t} = x_{t-1}, we can write (14) as
\begin{pmatrix}
x_{t} \\ y_{t}
\end{pmatrix}
=
\begin{pmatrix}
a \\ 0
\end{pmatrix}
+
\begin{pmatrix}
b_{1} & b_{2} \\
1 & 0
\end{pmatrix}
\begin{pmatrix}
x_{t-1} \\ y_{t-1}
\end{pmatrix},
\tag{15} where y_{1} = x_{0} and x_{1} is given.
- The eigenvalues of \begin{pmatrix} b_{1} & b_{2} \\ 1 & 0 \end{pmatrix} can tell you whether x_{t} will converge to a steady-state value, oscillate, or diverge to infinity as t \rightarrow \infty.
Figure 4: The figure shows the solutions to \begin{pmatrix} x_{t} \\ y_{t} \end{pmatrix} = \begin{pmatrix} 0.5 \\ 0 \end{pmatrix} + \begin{pmatrix} 0.5 & -0.8 \\ 1 & 0 \end{pmatrix} \begin{pmatrix} x_{t-1} \\ y_{t-1} \end{pmatrix}, given x_{0} = y_{0} = 0.
Adding a Forcing Term
- We can add a non-autonomous component to a first-order difference equation
y_{t} = b y_{t-1} + w_{t}.
\tag{16}
- We will now start at t = -1.
- To solve it we proceed as before,
\begin{aligned}
y_{0} & = b y_{-1} + w_{0}, \\
y_{1} & = b y_{0} + w_{1} = b (b y_{-1} + w_{0}) + w_{1} = b^{2} y_{-1} + b w_{0} + w_{1}, \\
y_{2} & = b y_{1} + w_{2} = b (b^{2} y_{-1} + b w_{0} + w_{1}) + w_{2} = b^{3} y_{-1} + b^{2} w_{0} + b w_{1} + w_{2}, \\
y_{t} & = b^{t} y_{-1} + (b^{t} w_{0} + b^{t-1} w_{1} + \ldots + b^{t-t} w_{t}) = b^{t} y_{-1} + \sum_{i = 0}^{t} b^{t-i} w_{i}.
\end{aligned}
Interpretation
- In the next chapter, w_{t} will represent random shocks to the system transforming the process for y_{t} into a stochastic process.
- For the moment, though, we assume that w_{t} is deterministic and represents exogenous inputs to the model.
Impulse Response
- The impulse response represents the solution to (16) if
w_{t} = \begin{cases} 1 \quad \text{if $t = 0$}, \\ 0 \quad \text{if $t > 0$}, \end{cases}
assuming that the system is in steady-state, i.e., y_{-1} = 0.
- In this scenario, the solution to (16) is
y_{t} = b^{t},
which is the impulse response of the system.
A Second Take on the Gordon Growth Model
- The Gordon growth model can be written as
\begin{aligned}
p_{t+1} & = (1 + r) p_{t} - (1 + g) d_{t}, \\
d_{t+1} & = (1 + g) d_{t}.
\end{aligned}
- In matrix form
\begin{pmatrix}
p_{t+1} \\ d_{t+1}
\end{pmatrix}
=
\begin{pmatrix}
1 + r & -(1 + g) \\
0 & 1 + g
\end{pmatrix}
\begin{pmatrix}
p_{t} \\ d_{t}
\end{pmatrix}.
- We can iterate again to find
\begin{pmatrix}
p_{t+n} \\ d_{t+n}
\end{pmatrix}
=
\begin{pmatrix}
1 + r & -(1 + g) \\
0 & 1 + g
\end{pmatrix}^{n}
\begin{pmatrix}
p_{t} \\ d_{t}
\end{pmatrix}.
\tag{17}
Explicit Solution
- It is possible to get an explicit solution for the matrix exponential in (17),
\begin{pmatrix}
p_{t+n} \\ d_{t+n}
\end{pmatrix}
=
\begin{pmatrix}
(1 + r)^{n} & \dfrac{1 + g}{r - g}((1 + g)^{n} - (1 + r)^{n}) \\
0 & (1 + g)^{n}
\end{pmatrix}
\begin{pmatrix}
p_{t} \\ d_{t}
\end{pmatrix},
implying
\begin{aligned}
p_{t+n} & = (1 + r)^{n} p_{t} + \dfrac{1 + g}{r - g}((1 + g)^{n} - (1 + r)^{n}) d_{t}, \\
d_{t+n} & = (1 + g)^{n} d_{t}.
\end{aligned}
\tag{18}
- The second equation above is not a surprise, but the first equation tells how the future stock price relates to the current price.
No Bubbles
- Equation (18) implies
\frac{p_{t+n}}{(1 + r)^{n}} = p_{t} - \dfrac{1 + g}{r - g}\left(1 - \frac{(1 + g)^{n}}{(1 + r)^{n}}\right) d_{t}. \\
- If we apply the transversality condition to the expression above we get,
0 = \lim_{n \rightarrow \infty} \frac{p_{t+n}}{(1 + r)^{n}} = p_{t} - \dfrac{1 + g}{r - g} d_{t},
as long as g < r.
- The transversality condition is necessary to relate the stock price to fundamentals.
- Otherwise the stock price can be anything regardless of the dividends paid by the stock.