COBWEB THEOREMS WITH PRODUCTION LAGS AND PRICE FORECASTING

DANIEL DUFRESNE AND FELISA J. V

AZQUEZ-ABAD
ABSTRACT. The classical cobweb theorem is extended to include production lags and price
forecasts. Price forecasting based on a longer period has a stabilizing effect on prices. Longer
production lags do not necessarily lead to unstable prices; very long lags lead to cycles of con-
stant amplitude. The classical cobweb requires elasticity of demand to be greater than that of
supply; this is not necessarily the case in a more general setting, price forcasting has a stabilizing
effect. Random shocks are also considered.
Keywords: Cobweb theorem; production lags; stable markets; price uctuations
1. INTRODUCTION
But surely the cob-web cycle is an oversimplication of reality (Samuelson [26], p.4). Many
other famous and less famous economists must have expressed the same opinion over the past
decades. We propose to bring the cobweb model at least a little closer to reality by introducing
production lags and price forecasts into it. The way equilibrium is reached in a theoretical
model should then be better understood. In particular, we ask whether production lags cause
intstability of prices, and whether the classical condition for a cobweb to lead to equilibrium
(that elasticity of demand be greater than elasticity of supply) still holds in a more general
model
1
.
There is a considerable literature on business or economic cycles. The cobweb theorem has a
long history, see Ezekiel [11]. We follow Chapter 2 of van Doorn [8] in stating its classical
form. The following assumptions are made:
(A1) supply depends only on the price forecast;
(A2) actual market price adjusts to demand, so as to eliminate excess demand instantaneously
in the trading period;
(A3) price forecast equals most recent observed price, and
(A4) there are no inventories, and neither buyers nor sellers have an incentive to speculate.
Let P
t
be the market price for a unit of commodity at time t. The quantity demanded at period
t, denoted by Q
d
t
is given by
Q
d
t
= a
0
a
1
P
t
= D(P
t
),
while the quantity supplied is
Q
s
t
= b
0
+ b
1

P
t
= S(

P
t
),
where

P
t
is the price forecast (the result of forecasting at time t 1). The conditions a
1
, b
1
> 0
ensure that quantity demanded decreases and quantity supplied increases as functions of price.
The assumptions stated above mean that
D(P
t
) = S(

P
t
) and

P
t
= P
t1
.
Making the substitutions, it is seen that the price sequence follows
P
t
=
b
1
a
1
P
t1
+
a
0
b
0
a
1
. (1)
1
We take the demand elasticity to be a positive (absolute) value
1
In the classical cobweb model the solution to (1) is then
P
t
= (P
0
P

)
_

b
1
a
1
_
t
+ P

, (2)
where P

is the equilibrium solution to (1), that is,

P

=
b
1
a
1
P

+
a
0
b
0
a
1
.
The sequence dened in (2) converges to P

if and only if b
1
< a
1
. In words, the condition for
convergence is that minus the slope of demand as a function of price be larger than the slope of
supply as a funtion of price. Usually, economic modelling assumes that the market is initially
in equilibrium at P

, and that an exogenous disturbance results in P

0
= P

. An important
question is whether such disturbances persist or die out; this can be answered by studying the
conditions for convergence of P
t
to P

.
We have in mind the prices of metals, for instance copper, which have greatly uctuated and
shown some appearance of cycles over time. An important aspect of mining is the lag between
the time the decision to increase or decrease production is made and the time the decision actu-
ally takes effect in the market. It takes several years for a planned new mine to start producing,
and this has made some believe that the lags themselves may be a main cause of price uctua-
tions. In this paper, we study models that include production lags as well as price forecasting,
the latter based on current and past prices. Stability means that prices converge to an equilib-
rium as time passes. Random disturbances will also be included; in those models stability
will mean that prices have a limit probability distribution (in other words that the price process
has a stationary limit).
Stability is not synonymous with absence of uctuations, but it is a property of markets in which
price uctuations tend to dampen over time. By varying the parameters, one can get an idea of
what may generate uctuations. Is it production lags? Is it how prices are forecast? We look at
those questions in the following sections.
Given prices P
t
, P
t1
, . . . , let

P
t+
be the price forecast used by producers to establish produc-
tion at time t + . (A more explicit notation would be

P
t,t+
, but we will use the simpler

P
t+
,
as the lag will be xed.) The classical cobweb theorem has a lag of one time unit.
The market clearing condition is now
D(P
t+
) = S(

P
t+
). (3)
This leads us to study the dynamical system
P
t+
= D
1
S(

P
t+
). (4)
Important questions are under what conditions this recursion has an equilibrium point, and if
so whether P
t
converges to it when t . Once again, such convergence does not exclude
uctuations, but it does say that perturbations are damped by the system over time.
The classical description of the cobweb theorem (such as the one we gave above) assumes
that the supply and demand functions are linear. We will assume that the demand and supply
functions are respectively
D(p) = p
d
, S(p) = p
s
, d, s > 0. (5)
One advantage of these functions is that price and production always remain positive, while
linear functions may lead to negative prices (see (2)). Tractability is achieved by using the
logarithm of prices, as will be shown below. It will be seen that a very important quantity is the
ratio of the elasticities of supply and demand, which we denote c = s/d.
2
Let {1, 2, . . . } be the production lag. We retain assumptions A1, A2 and A4, but replace
(A3) with
(A3

) the price forecast

P
t+
is a weighted geometric average of (P
tm
, . . . , P
t
) for some
m {0, 1, 2, . . . }.
(N.B. In the sequel will always stand for lag, and m for memory.)
This means
log

P
t+
=
m

j=0

j
log P
tj
,
where the weights
j
add up to 1. Letting
t
= log P
t
,
t
= log

P
t
, the log-price forecast is
given by

t+
=
m

j=0

j

tj
, where
m

j=0

j
= 1. (6)
A moving average model has all weights non-negative. Several forecasting schemes have been
identied in the literature, though always for = 1 (see [29] for a brief description). Static
expectations refers to
t+1
=
t
, or m = 0; extrapolative expectations means m = 1 and

0
> 1, as

t
+
1

t1
=
t
+ (
0
1)
t
+ (1
0
)
t1
=
t
+ (
0
1)(
t

t1
).
Adaptive expectations refers to

t
=
t1
+ (1 )
t
.
This is the limit case when m and
m
is proportional to
m
, as we explain in Section 2.7.
We call this exponential smoothing. Note that all these schemes are applied to log-price, this is
what makes the model tractable.
Van Doorn [8] attributes to Hicks the use of the logarithm of the price rather than the price itself,
in the context of a one-lag model or one with distributed lags, but the study of such systems
is not carried out mathematically in [8].
Under (5), the market clearing equation (3) reads
P
t+
= (

P
t+
)
s/d
. (7)
From (6) the sequence
t
satises

t+
= c
m

j=0

tj
, c = s/d. (8)
Since c

j

j
> 0, the unique equilibrium point is

= 0 or, equivalently, P

= 1. The
solution
t
of (8) has the general form (see [12])

t
=
+m

j=1
b
j
x
t
j
,
where (x
1
, . . . , x
+m
) are the zeroes of the characteristic polynomial
h
,m
(x) = x
+m
+ c
m

j=0

j
x
mj
. (9)
The constants b
j
; j = 1, . . . , + m may be found from the initial conditions for
t
.
Denition 1. The system (8) is said to be stable if lim
t

t
=

= 0, given any initial

conditions. Otherwise it is unstable.
3
If the characteristic polynomial (9) has complex zeroes then
t
has oscillatory components,
which means that the sequence may uctuate around

even though it eventually converges to

. This will happen frequently in our examples.The magnitude of the zeroes will determine
whether the system is stable or not; if all the zeroes of the characteristic polynomial have norm
(modulus) strictly less than one then the system is stable; if at least one zero has norm greater
than or equal to one then the system is unstable. If the zero or zeroes with largest norm have
norm precisely equal to 1 then there will be oscillations with constant amplitude, at least for
some initial conditions.
The classical form of the cobweb model has = 1, m = 0 and linear supply and demand
functions. In our setting, the cobweb model with = 1 and m = 0 becomes
P
t
= (

P
t
)
s/d
= (P
t1
)
s/d
,
which gives

t
=
_
s
d
_

t1
=
_

s
d
_
t

0
.
In the classical cobweb model the market is stable if, and only if, s < d; in other words, stability
occurs if and only if, the elasticity of supply is smaller than the elasticity of demand. When this
is the case,
t

= 0. This necessary and sufcient condition for stability, s < d, is

remarkably simple and easy to interpret. We will see that when lags and price forecasts are
introduced the conditions for stablility are no longer so simple.
Chiarella [6] studies a system where expected prices follow adaptive expectations, when the
demand curve is linear, while the supply curve is non-linear (with a single point of inexion,
convex to the left, and concave to the right, a fairly general non-linear S-shaped supply func-
tion, [6], p.383). He then shows that the system is either (1) stable, (2) unstable but cyclical,
or (3) chaotic. These are very interesting results, but we follow a different route.
The paper is organised as follows. Section 2 studies the deterministic models in some de-
tail, mostly numerically, although some simple results are proved mathematically. This section
avoids the generality of Section 3 but focuses instead on experiments that lead to interesting
patterns and related questions, which will be studied more deeply in subsequent sections.
In Section 3 we derive general results on deterministic models trying to answer the questions
raised in Section 2. The models are represented by linear difference equations; stability is
determined by the study of the roots of the characteristic polynomial (9) of those difference
equations. An important tool is Rouch es Theorem, see below. Our results seem to contradict
the view that production lags, by themselves, cause instabilities.
In Section 4 we incorporate randomness into (8), in the form of additive and multiplicative
noise. This leads to the question of stability of products of random matrices, a topic that so far
belonged more in physical chaos theory than in economics.
Notation. The set of complex numbers (or complex plane) is denoted C. The norm (or
modulus, or absolute value) of z = x + iy (x, y real) is |z| =
_
x
2
+ y
2
, its conjugate is
z = x iy. The circle with centre z and radius in C is denoted C
z,
; the open disk (or ball)
with centre z and radius is denoted B
z,
; the closed disk is denoted B
z,
.
We will use Rouch es Theorem from complex analysis: if , are analytic on and inside a
closed contour L, and |(z)| > |(z)| for z L, then and + have the same number of
zeroes inside L. Here is a rst application: whatever the averaging period m and the delay ,
there will be instability if the ratio of elasticities c = s/d is large enough.
Theorem 1. Let 1, m 0. There exists 0 < c
0
< such that for all c c
0
the system
dened by (8) is unstable.
4
Proof. Let
g(z) =
m

j=0

j
z
mj
,
and thus h
,m
(z) = z
+m
+ cg(z). We show that at least one zero of
z
+m
c
+ g(z)
is outside C
0,
, for some 1. There exists 1 such that |g(z)| > 0 for all z C
0,
.
Therefore
|g(z)| >
|z|
+m
c
on C
0,
for all c large enough. Apply Rouch es theorem with (z) = g(z) and (z) = z
+m
/c.
Then (z) + (z) has the same number of zeroes inside C
0,
as g(z), that is, at most m. That
leaves at least zeroes on or outside C
0,
, implying instability.
Remark. A supercially more general version of our model would have supply and demand
functions
D(p) = k
d
p
d
, S(p) = k
s
p
s
, d, s > 0. (10)
Consider the change of variables
p = p, =
_
k
d
k
s
_ 1
d+s
, D(p) =

D( p), S(p) =

S( p), = k
s
d+s
d
k
d
d+s
s
; (11)
this is a change in currency together with a change in units. It can be veried that

D( p) = p
d
,

S( p) = p
s
.
There is thus no greater generality in (10) than in (5).
2. DETERMINISTIC MODELS: NUMERICAL EXAMPLES
In this section we present numerical experiments that illustrate the inuence of the parameters
, c, and m on stability. The patterns observed here will motivate the more general (and
mathematical) analysis in Section 3. In all the examples we choose erratic initial conditions

t
= (10)
t
sin(1/(t + 3)), t = 0, . . . , + m.
Figure 1 shows the price
t
as a function of time t when c = 1.7, m = 5, = (.2, .2, .2, .2, .1, .1),
and the delay takes one of the values = 2, 3, 4. Notice that c = 1.7 yields instabilities with
= 1 (this is the classical cobweb theorem). However, for = 2 the system is stable, for = 3
it is nearly periodic with constant amplitude, and for = 4 it is unstable. For larger lags 5
we also observed instability. (However, there is an important comment on this example at the
end of Subsection 3.2.)
10 20 30 40 50 60 70
0.85
0.90
0.95
1.00
1.05
1.10
1.15
(a) = 2
10 20 30 40 50 60 70
1.0
1.2
1.4
1.6
(b) = 3
10 20 30 40 50 60 70
0.8
1.0
1.2
1.4
1.6
1.8
(c) = 4
FIGURE 1. As production lag increases, behaviour changes from stable to unstable.
5
Figure 2 shows the effect of the forecasting period m on the price behaviour, for c = 1.7
and = 3. The rst plot shows the logarithm of price
t
for m = 1, = (.7, .3) (the price
itself soon reaches values larger than 10
6
). The next plot is the behaviour of
t
when m = 5
( = (.2, .2, .2, .2, .1, .1)) and the last one when m = 7 ( = (.2, .1, .1, .1, .1, .1, .1, .1)). These
plots hint at a stabilising effect of increasing m, and are consistent with other experiments we
made.
10 20 30 40 50 60 70
10000
5000
5000
10000
(a) m = 1
10 20 30 40 50 60 70
1.0
1.2
1.4
1.6
(b) m = 5
10 20 30 40 50 60 70
1.0
1.1
1.2
(c) m = 7
FIGURE 2. As m increases, behaviour changes from unstable to stable.
In Figure 3 the parameter c is varied, while = 3, m = 7 and = (.2, .1, .1, .1, .1, .1, .1, .1).
There is apparent stability, except that for the largest value of c there are oscillations of more or
less constant amplitude. The last plot, where c = 3.8 shows nearly cyclical behaviour. Other
experiments (not shown) with larger values of c caused the prices to diverge. Recall that in the
classical cobweb stability occurs only if c < 1.
10 20 30 40 50 60 70
1.0
1.1
1.2
(a) c = 1.7
10 20 30 40 50 60 70
1.0
1.2
1.4
1.6
(b) c = 3.0
10 20 30 40 50 60 70
0.8
1.0
1.2
1.4
1.6
1.8
(c) c = 3.8
FIGURE 3. As c increases, behavior changes from stable to unstable.
The rest of this section presents detailed discussions of particular cases, that sometimes show
intriguing patterns which, to our knowledge, have not been noted in the context of economic
cycles.
2.1. The case m = 0. Suppose m = 0 and {1, 2, 3, . . . }. The price sequence satises

t+
= c
t
.
(Once again recall that the classical cobweb theorem has m = 0 and = 1.) The solution can
be written as

kj
= (c)
k

j
, k {1, 2, . . . }, j {0, . . . , 1}.
Here, there are price dynamics that work in parallel, i.e. they are not coupled. Each initial
condition
j
determines
j
,
2j
and so on. If 0 < c < 1 then there are damped oscillations
that tend to zero as k . If c = 1 then oscillations of same amplitude and period 2 persist
endlessly, and if c > 1 then the log-prices alternate in sign but increase geometrically in size as
time goes by. The role of the ratio of elasticities is clear in this case.
6
2.2. The case = 1, m = 1.
Theorem 2. Let = m = 1. Then the sequence
t
is stable if, and only if, 1 1/c <
0
<
1/(2c) + 1/2.
Proof. The solution
t
of (8) is stable if, and only if, both zeroes of (9) have norm less than 1.
The characteristic polynomial (9) is
h
1,1
(x) = x
2
+ c
0
x + c
1
where
1
= 1
0
.
From Theorem 4.2 of [12], stability for this second order equation is equivalent to the following
three conditions holding simultaneously:
1 + c
0
+ c(1
0
) > 0
1 c
0
+ c(1
0
) > 0
1 c(1
0
) > 0.
Given that c > 0, these are in turn equivalent to 1 1/c <
0
< 1/(2c) + 1/2.
Here the system is unstable for c 3, because the condition in Theorem 2 cannot then be
satised. When = 1, m = 0, the model is stable only when c < 1, which says that the
elasticity of demand is larger than the elasticity of supply. By contrast, when m = = 1
stability can be achieved for any c smaller than 3, which means that the elasticity of supply
only needs to be smaller than three times the elasticity of demand. It is somewhat surprising
that increasing the forecasting period m from 0 to 1 has such a signicant effect. Observe
that the only
0
that produces stability for all c < 3 is 2/3. There is no obvious reason why
it should be
0
= 2/3 that makes this region largest; this corresponds to assigning twice as
much weight to the most recent price as the previous one. Note also that if 0 < c < 1 then
1/(2c) + 1/2 > 1, meaning that for those values of c the sequence is stable in particular for

0
(1, 1/(2c) + 1/2), which corresponds to establishing the price forecast by extrapolating
the two most recent prices. For example, if c = 1/2, log P
t
= 1, log P
t1
= 0,
0
= 5/4 then
the forecast is log

P
t+1
= 5/4; the price sequence is stable in this case, even though a priori
one might think that extrapolating the most recent prices would be a destabilizing policy. In
the economics literature, the expression extrapolative expectations refers to forecasting based
on the last two prices. This idea was rst studied mathematically in a macroeconomic model
of inventories by Metzler [19]. Metzler studies a somewhat different problem, but the algebra
is similar to our case = 1, m = 1. (Metzler and others believed that extrapolation was a
cause of instability). Turnovsky [29] mentions the destabilizing effect of extrapolation (i.e.

0
> 1). Extrapolative expectations is also called a myopic forecast by other authors, for
instance Wheaton [32] claims that this is a cause of real estate oscillations.
2.3. The cases m = 1, 2,
0
= 0 or 1. When m = 1 and 2, equation (8) cannot be
solved exactly, except for the two special cases
0
= 0 and
0
= 1, which are tractable. In the
rst case, the characteristic polynomial is
h
,m
(x) = x
+1
+ c
which has zeros x
j
with norm
|x
j
| = c
1/(+1)
, j = 1, . . . , + 1.
Hence the condition c < 1 is a necessary and sufcient condition for stability when
0
= 0. In
the case
0
= 1 the zeros of the characteristic polynomial
h
,m
(x) = x
+1
+ cx
are 0 and x
j
= c
1

e
2(j1)

, j = 1, . . . , , and once again c < 1 is a necessary and sufcient

condition for stability.
7
2.4. The cases m = 1, 2,
0
arbitrary. The graphs in Figure 4 show the region of stabil-
ity for m = 1 and between 1 and 100, computed using Mathematica
R
. The stability region
is the set of (
0
, c) that lead to a stable price sequence; the curves on the graphs are the up-
per boundaries of the stability regions. We have observed numerically that the stability region
shrinks to some limit set as increases, though the shape of the upper boundary is different for
even and odd values of .
1
3
5
limit
3 2 1 0 1 2 3
0.0
0.5
1.0
1.5
2.0
(a) odd.
2
4
6
limit
3 2 1 0 1 2 3
0.0
0.5
1.0
1.5
2.0
(b) even.
FIGURE 4. m = 1. The region of stability is the area below each of the curves shown.
The stability region is largest when = 1; the upper boundaries for 5 are indistinguishable
from the limit when
0
is outside (0, 1). The limit as tends to innity of the upper boundary
of the stability region coincides (as far as we can tell numerically) with the curve
c = (|
0
| +|1
0
|)
1
(solid line). We give a partial justication in Section 3.
2.5. The case 1 < m < : equal weights
j
. In our next numerical experiments the weights
are equal, i.e.
j
= 1/(m + 1), j = 0, . . . , m. Hence,
h
,m
(x) = x
+m
+
c
m + 1
m

j=0
x
mj
. (12)
Figure 5 plots the supremum of the values of c that preserve stability, that is, the values of
c

(, m) = inf
c>0
_
max
j{1,...,m+}
|x
j
(c)| = 1
_
,
where {x
j
(c)} are the zeroes of (12). The horizontal axis shows the values of the averaging
period m, and the four dotted lines correspond to = 1, 2, 3, 4. The dotted lines visually appear
to be linear functions of m, with a slope that decreases as the lag increases. A closer look
at the actual values of c

(, m) shows that for xed = 2, 3, 4 the functions are not precisely

linear in m, but for = 1 the slope is indeed constant.
We are able to prove that if = 1 then c

(, m) = m + 1, see Section 3.3.

Figure 6 shows the zeroes of the characteristic polynomial h
,m
(x) in the complex plane for
m = 6 and various values of . Each number on a plot represents the location of a zero
of h
,m
(x). It is seen that, at least in those cases, the zeroes move towards the unit circle as
increases, and, furthermore, that the zeroes are approximately uniformly spread around the
8
2 4 6 8 10
2
4
6
8
10
FIGURE 5. Critical value of the ratio of elasticities c as a function of m. The top
line corresponds to = 1, the ones below correspond to = 2, 3, 4 (in that
order).
circle. Intuitively this means that the behaviour of prices tends to oscillations when increases.
We now provide an incomplete justication for this. The characteristic polynomial is
h
,m
(z) = z
+m
+ cg(z), g(z) =
1
m + 1
m

j=0
z
mj
.
Letting w = z
+m
, this may be rewritten as H
,m
(w) = w + cg(w
1
+m
).
1
1
1
1
1
1
1
5
5
5
5
5
5
5
5
5
5
5
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
m 6, c 0.2
1
1
1
1
1
1
1
5
5
5
5
5
5
5
5
5
5
5
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
m 6, c 0.7
1
1
1
1
1
1
1
5
5
5
5
5
5
5
5
5
5
5
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
m 6, c 2
1
1
1
1
1
1
1
5
5
5
5
5
5
5
5
5
5
5
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
10
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
m 6, c 3
FIGURE 6. Zeroes of the characteristic polynomial h
,m
(x) for different values
of c, with m = 6, in the complex plane. Each zero is indicated by the number
. The circle has centre 0 and radius 1 (C
0,1
).
We now use Rouch es Theorem with (z) = cg(z), (z) = z
+m
and C = C
0,
, for 0 < < 1.
The zeroes of are all on the unit circle, and thus |(z)| > 0 for z C
0,
. Hence, for
larger than some
0
the inequality |(z)| > |(z)| is veried for z C
0,
; this implies that the
9
zeroes of h
,m
are all outside B
0,
, for any 0 < < 1. Now w
1
+m
1 as , and we are
left with
lim

H
,m
(w) = w + c.
Finally, reverting to w = z
+m
then says that the zeroes of h
,m
are, approximately, the solutions
of
z
+m
= c.
The zeroes are then approximately equal to
c
1
+m
e
i
2j
+m
, j = 1, . . . , + m.
Although not rigorously derived, this yields good approximations of the arguments 2j/( + m),
but not always a good one for the norms of the zeroes of h
,m
. For the latter it is better to rely
on the fact that the norm of the product of the zeroes of h
,m
is |h
,m
(0)| = c/(m + 1), which
yields the improved approximation
_
c
m + 1
_ 1
+m
e
i
2j
+m
, j = 1, . . . , + m. (13)
As an example, consider the rst graph in Figure 6, with equals weights
j
= 1/(m+1), c = .2
and m = 6. For = 5 the exact zeroes are r
j
e
i
j
, with
r
1
= 0.694659,
1
= 3.14159,
r
2
= 0.688829,
2
= 2.59503, r
3
= 0.688829,
3
= 2.59503,
r
4
= 0.700281,
4
= 2.03232, r
5
= 0.700281,
5
= 2.03232,
r
6
= 0.716104,
6
= 1.49399, r
7
= 0.716104,
7
= 1.49399,
r
8
= 0.730140,
8
= 0.92760, r
9
= 0.730140,
9
= 0.92760,
r
10
= 0.804108,
10
= 0.35203, r
11
= 0.804108,
11
= 0.35203,
while the approximations are re
i
j
, where r = 0.723819 and the
j
are
3.14159, 2.57039, 1.9992, 1.428, 0.856798, 0.285599.
If (13) were the true zeroes of h
,m
then the solutions

t
=

j=1
b
j
x
t
j
would have period + m. In the mining area, many believe that that the observed price cycles
correspond to the production lag . We see that this is approximately the case in our model, but
only when m is small.
2.6. Geometric weights. Geometric weights are used in many forecasting models. A parame-
ter > 0 is chosen and the weights
j
follow the geometric progression

j
=

j
(1 )
(1
m+1
)
, j = 0, . . . , m. (14)
Figure 7 shows plots of the critical boundary value c

as a function of for different values of

, m. In all cases it appears that the stability region decreases to c

= 1 as increases. The solid

line is the function
c()
def
=
(1 + )(1
m+1
)
(1 )(1 +
m+1
)
,
which is the boundary for the region for = 1 and even values of m, as we will prove in
Section 3.
10
0.0 0.2 0.4 0.6 0.8 1.0
0.0
0.5
1.0
1.5
2.0
2.5
3.0
10
5
4
3
2
1
(a) m = 2
0.0 0.2 0.4 0.6 0.8 1.0
0
1
2
3
4
5
6
10
5
4
3
2
1
(b) m = 5
0.0 0.2 0.4 0.6 0.8 1.0
0
2
4
6
8
10
10
5
4
3
2
1
(c) m = 10
FIGURE 7. Stability regions for geometric weights are shown for different pro-
duction lags .
2.7. Exponential smoothing. If 0 < < 1 and we formally let m in (14), we get

t+
= (1 )

j=0

tj
.
Rewriting the same for
t+1
and eliminating
t1
,
t2
, . . . , we then nd that

t+
=
t+1
+ (1 )
t
.
This says that the forecast made at time t for the price at time t + is a weighted average of
last periods forecast and the most recent price, using a xed proportion . This procedure is
mentioned in [8], page 24; it is sometimes called exponential smoothing or adaptive expec-
tations. From (7),
t+
= c
t+
and thus

t+
=
t+1
c(1 )
t
(15)
The case = 1 is remarkably simple:

t+1
= [(1 + c) c]
t
.
By setting
=
c
1 + c
=
s
d + s
,
one gets
t+1
= 0, i.e. there is convergence to the equilibrium price in just one time step.
11
There is an explicit result when = 2, reminiscent of the case = 1, m = 1 (Theorem 2).
Theorem 3. If (15) holds with c > 0, then
(a) if = 1, then the sequence
t
is stable if, and only if, (c 1)/(c + 1) < < 1;
(b) if = 2, then the sequence
t
is stable if, and only if, 1 1/c < < 1.
Proof. For part (a), the condition is
1 < (1 + c) c < 1 or
c 1
c + 1
< < 1.
For part (b), the zeroes of the characteristic polynomial
x
2
x + c(1 )
have norm less than 1 if, and only if ([12], p.172),
1 + + c(1 ) > 0, 1 + c(1 ) > 0, 1 c(1 ) > 0.
These are equivalent to 1 1/c < < 1.
When > 2 the characteristic polynomial has degree three or more, and an exact analysis of
the roots is not possible. Figure 8 shows the boundary of the stability region as a function of .
The solid lines are the functions (1 +)/(1 ) and 1/(1 ), and coincide numerically with
= 1, 2 respectively, as expected.
In all the experiments we made, for any c there is an interval I(c) = ((c), 1) such that I(c)
implies stability. We have not been able to prove mathematically that this is always the case
when 3.
0.0 0.2 0.4 0.6 0.8 1.0
0
2
4
6
8
10
10
5
2
1
FIGURE 8. Stability region for exponential smoothing.
12
1
1
1
1
5
5
5
5
5
5
5
5
10
10
10
10
10
10
10
10
10
10
10
10
10
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
50
Rez
Imz
m 3, c 0.2, 0.1, 0.6, 0.2, 0.1
FIGURE 9. Zeroes of the characteristic polynomial h
,m
(x) for non-equal
weights, in the complex plane. Each zero is indicated by the number . The
circle has centre 0 and radius 1 (C
0,1
).
2.8. Arbitrary weights. Figure 9 shows the locations of the zeroes in the complex plane in
one case where the weights
j
are not equal: = (0.1, 0.6, 0.2, 0.1), c = 1. The characteristic
polynomial
h
,m
(z) = z
+3
+ 0.2(0.1z
3
+ 0.6z
2
+ 0.2z + 0.1) = z
+3
+ 0.2g(z)
has all its zeroes inside the unit circle C
0,1
for all . This is a simple consequence of the reverse
triangle inequality: if |z| 1 then
|h
,m
(z)| |z
+3
| 0.2|0.1z
3
+ 0.6z
2
+ 0.2z + 0.1| |z|
+3
0.2|z|
3
.
The last expression cannot be zero if |z| 1, for any = 1, 2, . . . . We note that there are now
+ m 2 zeroes spread in a circular fashion, getting closer to C
0,1
as increases, as there are
+m zeroes in total and two zeroes that remain near 0.160.38i. It will be shown in Section
3 that inside the unit circle the zeroes of h
,m
have limits as tends to innity, and that they are
precisely the zeroes of g(z) =

m
j=0

j
z
mj
(as dened in the proof of Theorem 1).
3. DETERMINISTIC MODELS: GENERAL RESULTS
Let us recall from (3) that for a lag and an averaging period m + 1, the market clearing
condition
P
d
t+
=

P
s
t+
,
yields (8), that is

t+
=
s
d
m

j=0

tj
.
Writing c = s/d as before, the characteristic polynomial (9) is
h
,m
= x
+m
+ c
m

j=0

j
x
mj
,
and the solution of (8) is

t
=
+m

k=1
b
k
x
t
k
, (16)
13
where {x
k
} are the zeroes of the characteristic polynomial and {b
k
} are constants. The long
term behaviour of
t
is determined by the x
j
with the maximum norm, among those j such that
b
j
= 0. Analytic expressions for the zeroes of polynomials are not available for l + m > 2, but
we will derive results that narrow down the region where the zeroes are located.
In the literature, some have suggested that production lags themselves are the cause of uctuat-
ing prices. For example, in [23], p.276, Phillips mentions that the regulation of a system can be
improved if the lengths of the time delay operating around the main control loop are reduced.
Sterman ([27], Chapter 20) writes: markets with negative feedbacks through which price seeks
to equilibrate supply and demand often involve long time delays which lead to oscillation. Our
results do support this view, but maybe not in the way one might have expected, in the sense
that we do not nd that longer lags necessarily lead to instability. What we nd is that as the
lag increases the maximum of the norms of the roots x
k
tends to one. This means oscillations
of constant amplitude, whether the system is stable or unstable for small production lags. Thus,
longer lags have a stabilizing effect on unstable systems. This is the conclusion one draws from
the general results in Subsections 3.1 and 3.2. Subsection 3.3 studies the case where the weights
{
j
} are constant or form a geometric progression.
3.1. General result on the location of the roots and stability.
Theorem 4. (a) Suppose 1. If c
m

j=0
|
j
| <

then the zeroes of h

,m
are all less than in
norm, i.e. |x
j
| < . In particular, if
c
m

j=0
|
j
| < 1 (17)
then the system is stable.
(b) Suppose 1. If c
m

j=0
|
j
| <
+m
then the zeroes of h
,m
are all less than in norm.
Proof. Let 1 and |z| . Then, from the triangle inequality,
|h
,m
(z)| |z|
+m
c
m

j=0
|
j
||z|
mj
|z|
+m
c|z|
m
m

j=0
|
j
|.
If c

m
j=0
|
j
| <

then the last expression is positive. The zeroes of h

,m
must then all be in
C
0,
.
If 1 and |z| then the result follows from
|h
,m
(z)| |z|
+m
c
m

j=0
|
j
| ||
+m
c
m

j=0
|
j
|.
Observe that (17) is sufcient for stability, but not necessary. For instance, in the case m = 1
depicted in Figure 4 it is seen that when 0 <
0
< 1 the system is stable for some c > 1 =

0
+
1
.
Part (a) of the theorem implies that for any > 1 and for greater than some
0
, the zeroes of the
the characteristic polynomial are all inside C
0,
. This means that for systems that are unstable
for some there are larger s such that the maximum norm of the zeroes of the characteristic
polynomial is close to 1; in other words, an unstable system eventually becomes less unstable
as increases.
14
As an application of Theorem 4, let us now return to the cases m = 1, 1,
0
arbitrary, that
we looked at in Section 2 (cf. Figure 4).
Suppose is odd, and x
0
1. If we set c = (|
0
| + |1
0
|)
1
= 1/(2
0
1) then
a zero of h
,1
(x) is x = 1, and the system is unstable. However, Theorem 4 says that if
c < (|
0
| +|1
0
|)
1
then the system is stable. Hence, the stability region for odd,
0
1
consists of the points (
0
, c) with c < (|
0
| + |1
0
|)
1
. The situation is similar when is
even and
0
0; then c = (|
0
| +|1
0
|)
1
= 1/(1 2
0
) leads to a zero at x = 1 again,
while Theorem 4 gives stability when c < 1/(1 2
0
). Hence, the stability region for even,

0
0 consists of the points (
0
, c) with c < (|
0
| +|1
0
|)
1
.
3.2. Limiting behaviour with increasing production lags. The next result shows that when
the production lag increases without bound, all the zeroes of the characteristic polynomial
h
,m
(x) are arbitrarily close to the unit circle, with the possible exception of up to m zeroes
inside the unit circle. This means, loosely speaking, that longer production lags lead to oscilla-
tions of constant amplitude, and not to oscillations of increasing amplitude. A system that has
oscillations of increasing amplitude will be made less unstable for production lags that are long
enough. This partly contradicts the view that long production lags in themselves cause erratic
price behaviour.
We will use the following classical result from [25], page 261.
Theorem (Hurwitz) Let G be a non-empty connected open set in the complex plane. Suppose
,
n
, n 1 are analytic functions on G, and that
n
converges uniformly on compacts to in
G. Let U be a bounded open set of G with U G such that f has no zero on U. Then there is
an index n
U
N such that for each n n
U
the functions and
n
have the same number of
zeroes in U.
Theorem 5. Dene
g(z) =
m

j=0

j
z
mj
. (18)
If g has k zeroes inside the unit circle C
0,1
label them r
1
, . . . , r
k
. Let
1
(0, 1) be such that
C
0,
1
includes r
1
, . . . , r
k
in its interior. Let
2
> 1. Then there exists
0
< such that for any

0
, h
,m
() has exactly the same number of zeroes in C
0,
1
as g
m
and no zero outside C
0,
2
.
In addition, there are sequences r
1,
, . . . , r
k,
such that
lim

r
i,
= r
i
, for each 1 i k, and
For every
0
, h
,m
(r
i,
) = g(r
i
) = 0.
Proof. First, h
,m
(z) converges uniformly on compact sets (in B
0,1
) to cg(z). Take
1
as dened
above, and apply Hurwitzs Theorem for G = C
0,
1
to obtain that h
,m
has the same number of
zeroes as g inside C
0,
1
for all
0
. To obtain the limiting results for each of the zeroes, apply
Hurwitz Theorem to a sequence of balls around each zero r
i
with decreasing radius.
Second, for any
2
> 1 there is
1
such that
c
m

j=0
|
j
|

2
,
1
,
and thus by part (a) of Theorem 4 all the zeroes of h
,m
are inside C
0,
2
when
1
.
Theorem 5 explains the behaviour illustrated in Figures 6 and 9, namely that as increases, most
or all the zeroes approach the boundary of the unit circle. It does not, however, explain why
the zeroes are placed almost uniformly around the circle in the limit. The difference between
15
Figures 6 and 9 is that in the latter the polynomial g(z) has zeroes inside the unit circle. As the
theorem says, those zeroes remain there in the limit.
This leads us to reconsider the rst example in Section 2. There it appeared that increasing
led to instability (see Figure 1). When = 4 (last plot) the largest norm among the roots of the
characteristic polynomial is 1.028, which explains the increasing amplitude of the oscillations.
However, with larger that largest norm increases a bit more but then gradually decreases
towards 1, for instance for = 10 the largest norm is 1.038, for = 20 it is 1.023, and for
= 40 it is 1.013.
3.3. Constant or geometric weights.
Theorem 6. Consider the model for the log price (8) with constant weights. If = 1 then
c

(, m) = m + 1.
Proof. Write = c/(m + 1) and dene

h(z) = (1 z)h
1,m
(z) = (1 z)z
m+1
+ (1 z
m+1
).
The zeroes of

h are precisely those of h
1,m
together with the number 1. Any zero of

h satises
z
m+1
(z 1 + ) = . (19)
First, consider the case > 1, which is the same as c > m+ 1. Then the norm of the left-hand
side of (19) is
|z
m+2
+ ( 1)z
m+1
| |z|
m+2
+ ( 1)|z|
m+1
.
If |z| < 1, then this is no larger than |z|
m+1
< , which is a contradiction. Thus, if c > m+1
then h
1,m
has no zero inside the unit circle, and the system is unstable.
Next, suppose = 1, or c = m+1. Then (19) becomes z
m+2
= 1, which has m+2 zeroes, all
of norm 1, and thus h
1,m
has all its zeroes on the unit circle; thus c

(, m) m + 1.
Finally, suppose that 0 < < 1 and that we restrict our search for zeroes to |z| = 1 (i.e. to the
unit circle). Then (19) implies
|z (1 )| = ,
which has the unique solution z = 1 on the unit circle; it is readily checked that this is not a
zero of h
1,m
and we conclude that h
1,m
has no zero z with norm equal to 1. If |z| > 1, we get
|z (1 )| =

|z|
m+1
,
which has no solution because
|z (1 )| |z| (1 ) > >

|z|
m+1
.
Hence, if 0 < < 1 then all the zeroes of h
1,m
are inside the unit circle; thus c

(, m) m+1
for any 0 < < 1.
From all the above, we conclude that c

(1, m) = m + 1 for m 0.
Theorem 7. Consider the price dynamics in (8) with geometric weights (14), 0 < < 1, and
let = 1. Then the system is stable if
0 < c < c()
def
=
(1 + )(1
m+1
)
(1 )(1 +
m+1
)
.
Hence, c() c

.
16
Proof. Dene
() =
1
m+1
1
c

=
c
()
. (20)
The characteristic polynomial is
h
,m
(z) = z
+m
+ c

m
m

j=0
(z/)
mj
= z
+m
+ c

_
z
m+1

m+1
z
_
. (21)
Let y = z/, then the zeroes of h
,m
(z) are in a one-to-one correspondence with the zeroes of
y
+m
+
c

_
1 y
m+1
1 y
_
. (22)
More specically, the system (8) will be stable if, and only if, the zeroes of (22) are in B
1/
.
Multiply (22) by

(1 y) to get

h(y)
def
=

(1 y)y
+m
+ c

(1 y
m+1
)
=

y
+m+1
+
_

y
1
c

_
y
m+1
+ c

.
Except for y = 1, the zeroes of this polynomial are those of (22). Let
(y) =

y
+m+1
(y) =
_

y
1
c

_
y
m+1
+ c

.
Let = 1 and c < c(). We will show that if |y| = 1/, then |(y)| > |(y)| (see below for
the proof). Applying Rouch es Theorem, this in turn will imply that all the zeroes of

h(y) are
in B
1/
.
We need to show that if 0 < < 1, |y| = 1/ and 0 < c < c(), then
|c

+ ( c

)y
m+1
| <
m1
.
From the triangle inequality
|c

+ ( c

)y
m+1
| c

+| c

| |y
m+1
| = c

+| c

|
m1
.
If 0 < c

< then
c

+| c

|
m1
<
m1

m+1
c

+ c

< 1 (
m+1
1)c

< 1 ,
which is true for all 0 < < 1. If c

then
c

+| c

|
m1
<
m1

m+1
c

+ c

< 1
c

<
1 +
1 +
m+1
c <
(1 + )(1
m+1
)
(1 )(1 +
m+1
)
= c().
We note that the case of equal weights corresponds to = 1. The limit as 1 of c() is
evaluated straightforwardly using lH opitals rule, and it recovers the bound m+1 of Theorem6.
Theorem 8. If is odd and m even then c

c(). If = 1 and m is even then c

= c().
Proof. We show that for c = c(), z = 1 is always a zero of h
,m
(z) when is odd and m
even.
Replacing c with c() in (21)
h
,m
(z) = z
+m
+
1 +
1 +
m+1
_
z
m+1

m+1
z
_
,
17
so that, evaluating at z = 1,
h
,m
(1) = (1)
+m
+
1 +
1 +
m+1
_
(1)
m+1

m+1
1
_
= 1 + 1 = 0.
If = 1 and m is even then the above and Theorem 7 imply that c

= c().
4. RANDOM DISTURBANCES
Pryor and Solomon [24] introduce randomness in observed prices in a cobweb model, and
then study the average length of a cycle. Samuelson [26] imagines that producers might adjust
their production according to expected price, and talks of introducing randomness in the price
process, but does not develop those ideas. Turnovsky [29] studies stochastic stability for the
cobweb model with linear supply and demand functions and = 1, for forecast prices following
either the weighted average model with m = 1, or adaptive expectations. Neither of those
authors include production lags, as we do below.
In this section we introduce additive and multiplicative random disturbances in the logprice
process; not surprisingly the additive ones are a relatively straightforward extension of the
deterministic model studied above. Disturbances to the supply function mean multiplicative
errors, which lead to a rather more involved analysis. There is a parallel with the approach
used by Chiarella [6], since we end up computing Lyapunov exponents, which also relate to
chaos. In both models it is the variability of elasticity of supply that is the origin of chaotic
behaviour; in our model elasticity s changes randomly over time, while in Chiarellas case there
a deterministic S-shaped supply curve.
The system (7) has demand and supply curves that are xed through time. We now introduce
time-varying supply curves. We leave demand xed, since in the case of copper it appears that
supply is much less predictable than demand. In the words of Dunsby [9], p.157: Much of the
short-term volatility in prices resulting from physical supply-demand imbalances (e.g., ignoring
purely nancial sources of volatility) derives from supply shocks. Demand tends to grow more
steadily. The reasons given by Dunsby include technology, investment, wars, strikes, natural
disasters, and declining yields.
Starting from (3),
D(P
t+
) = S(

P
t+
), (23)
we let D(p) = k
d
p
d
as before, but write
S
t
(p) = k
s,t
p
s
t
.
Next, we successively get
k
d
(P
t+
)
d
= k
s,t+
_
exp
m

j=0

j
log P
tj
_
s
t+
P
t+
=
_
k
s,t+
k
d
_

1
d
exp
_

s
t+
d
m

j=0

j
log P
tj
_

t+
= c
t+
m

j=0

tj
+
t+
, (24)
where
c
t+
= s
t+
/d and
t+
=
1
d
k
s,t+
k
d
.
18
In order to study the effect of varying supply, we let both c
t
and
t
be random (always assuming
that c
t
> 0). To keep matters simple we assume that {(c
t
,
t
), t 1} is a sequence of indepen-
dent and identically distributed (i.i.d.) random vectors. Our rst task is to nd the expectation of

t
; simply take expectations on both sides of (24); if both Ec
t
and E
t
exist, then the expectation
of
t
satises the recurrence
E
t+
= (Ec
t+
)
m

j=0

tj
+E
t+
. (25)
This is the same system we studied before in the deterministic case. Although this is not manda-
tory, in order to simplify the algebra we will make the same change of units we made in Section
1, replacing the constants c, k
s
, s with Ec
t
, Ek
s,t
, Es
t
in (11). We then have, for the rest of this
section,
E
t
= 0, E
t+
= (Ec
t+
)
m

j=0

tj
.
The expected value of
t
is zero, and thus the equilibrium value of E
t
is also zero.
Convergence of the expected value of
t
to zero does not imply that the sequence
t
has a
limit distribution, or a nite variance, as t tends to innity. In this model we will say that the
sequence
t
is stable if it has a limit distribution as t tends to innity, for any set of initial
condtions
0
,
1
, . . . ,
tm
(the latter are not random).
When c
t
= c is deterministic the sequence {
t
} in (24) is an autoregressive process of order
+ m, and there are well-known conditions for its stability. Observe that regarding the dis-
tribution of (c
t
,
t
) we are assuming nothing besides independence over time (c
t
and
t
may
be dependent). When c
t
is not deterministic the process {
t
} is called a random coefcient
autoregressive process.
We will study the problem of stability from two different points of view. The rst one is the
existence of the limit distribution, using results for the theory of products of random matrices.
The second one will assume that (c
t
,
t
) have nite second moments, and we will look for
conditions under which the second moment of
t
remains nite as t tends to innity; this will
also imply that the sequence has a limit distribution.
Turnovsky [29] uses a stochastic Lyapunov function to nd sufcient (though not necessary)
conditions for convergence with probability one pf P
t
to some value P

. There are signicant

differences between his approach and ours. First, Turnovsky needs the variance of the distur-
bances to tend to zero as the price approaches P

, while we let the disturbances have constant

variance; second, Turnovsky considers a more complex noise process, with correlation across
time; nally, Turnovsky was writing before the work of Kesten and others, especially Vervaat,
had become known. We nd it counterintuitive to twist the model in the way Turnovsky [29]
does to obtain convergence to a specic constant price; rather, random shocks lead either to
instability or to a limit distribution, naturally excluding convergence to a specic price P

,
4.1. Existence of limit distribution under the weakest conditions. The more technical dis-
cussion below is best introduced by describing the simpler case = 1, m = 0:

t+1
= c
t+1

t
+
t+1
.
(This is a random version of the classical cobweb model.) The existence and properties of the
limit distribution of
t
in this case was studied in great detail by Vervaat in [31]. Iterating the
equation yields

t
=
t
c
t

t1
+ c
t
c
t1

t2
+ + (1)
t
c
t
c
t1
c
1

0
.
In order to determine whether this sequence has a limit distribution, Vervaat rst reverses the
order of the subscripts, which does not alter the probability distribution, since the sequence
19
{(c
t
,
t
)} is asumed i.i.d. More specically, denoting equality in distribution by
d
=,

t
d
=
t

n=1
(1)
n1
c
1
c
2
c
n1

n
+ (1)
t
c
1
c
2
c
t

0
. (26)
He then uses the n-th root test for series:
if limsup
n
|a
n
|
1
n
< 1 then

n1
|a
n
| < .
This is applied to the time-reversed series we just described:
if limsup
t
|c
1
c
2
c
t

t
|
1
t
< 1 a.s. then

t1
|c
1
c
2
c
t

t
| < a.s..
(Here a.s. stands for almost surely, which means the same as with probability one.) Next
consider c
1
c
2
c
t
and
t
separately, recalling that c
t
> 0. Since
(c
1
c
2
c
t
)
1
t
= exp
_
1
t
t

k=1
log c
k
_
,
it is then obvious that if Elog c
1
< 0 then, by the Law of Large Numbers,
lim
t
1
t
t

k=1
log c
k
= Elog c
1
,
and thus
lim
t
(c
1
c
2
c
t
)
1
t
= lim
t
exp
_
1
t
t

k=1
log c
k
_
< 1.
If Elog |
1
| is nite, then
lim
t
1
t
t

k=1
log |
k
| Elog |
1
|,
and thus
lim
t
log |
t
|
1
t
= 0.
Finally,
limsup
t
|c
1
c
2
c
t

t
|
1
t
< 1
under the assumptions Elog c
1
< 0, Elog |
1
| < , and thus the right-hand side of (26) has
a.s. a nite limit. Note that Elog |c
1
| < 0 implies that c
1
c
2
c
t
tends to zero with probability
one as t tends to innity. The assumption regarding the distribution of
1
can be weakened by
noting that values of |
1
| smaller than 1 cannot cause divergence of the sum, and so requiring
Elog
+
|
1
| < is sufcient, if log
+
x = max(log x, 0).
There are results of the same nature as the ones above in the more general case where 1 and
m 0 are arbitrary in (24), but they are not as straighforward, even though the randomness in
the system is generated by the same pair (c
t
,
t
). The process {
t
} is in general not Markovian,
and it is useful to obtain a Markovian representation for it by dening
X
t
= (
t
, . . . ,
tm+1
)
T
, B
t
= (
t
, 0, . . . , 0)
T
A
t
=
1
..
_
_
_
_
_
_
_
0 0 0
1 0
1
0
c
t

0
c
t

m1
c
t

m
0
.
.
.
1 0
_
_
_
_
_
_
_
_
.
20
Here A
t
is ( +m) ( +m), and B
t
, X
t
are ( +m) 1. The rst line of A
t
has 1 leading
zeros, followed by c
0
, c
1
, . . . , c
m
, and a subdiagonal of 1s; the other elements of A
t
are 0. The process {X
t
} is dened recursively as
X
t
= A
t
X
t1
+ B
t
, t = 1, 2, . . . (27)
This process is Markovian, because the sequence {(A
t
, B
t
)} is i.i.d.
The adaptive expectations model with random disturbances becomes

t+
=
t
c
t+
(1 )
t
+
t+
.
To obtain a Markovian representation, set
X
t
= (
t
, . . . ,
t+1
)
T
, B
t
= (
t
, 0, . . . , 0)
T
.
Then the matrix A
t
has the form
A
t
=
1
..
_
_
_
_
_
_
_
0 0
1 0
1
0 1
c
t
(1 )
0
0
0
_
_
_
_
_
_
.
Here A
t
is , and B
t
, X
t
are 1. The process {X
t
} is dened recursively as before, by
(27).
We will use the Euclidian vector norm | |
e
and a matrix norm that is compatible with it, in
the sense that
|Mx|
e
M |x|
e
(28)
(see Chapter 5 of [14]). The notation |A| refers to the matrix of the absolute values of the
elements of A.
We now consider system (27) in some generality, with A
t
an N N random matrix (not nec-
essarily of the form specied above). Conditions for the stability of (27) cannot be obtained as
simply as in the one-dimensional case. This is essentially because the logarithm and exponen-
tial of matrices do not have the same properties as the corresponding functions of real numbers;
in particular, for matrices M
1
and M
2
it is general not the case that
e
M
1
+M
2
= e
M
1
e
M
2
.
In the one-dimensional case the condition Elog |A
1
| < 0 implies that A
1
A
n
tends to 0
geometrically; in (27) the corresponding condition is
({A
n
}) = inf{
1
n
Elog A
n
A
1
, n N} < 0. (29)
This is called the top Lyapunov exponent of the matrices {A
1
, A
2
, . . . }. Some of the results we
will use go back to Furstenberg and Kesten [15]. It is known ([17]) that if {A
n
, n 1} is a
stationary process and Elog
+
A
1
< , then ({A
n
}) [, ), and, moreover,
({A
n
}) = lim
n
1
n
log A
n
A
1
, n N.
Part (a) of the following theorem was proved in one dimension by Brandt [5] and extended to
the vector case by Bougerol and Picard [4]. We have added part (b) for clarity (it is proved in
the same way as part (a)).
Theorem 9. (a) Let {(A
n
, B
n
), n Z} be a strictly stationary ergodic process such that both
E(log
+
A
1
) and E(log
+
B
1
) are nite. Suppose that the top Lyapunov exponent dened
by (29) is strictly negative. Then, for all n Z, the series
X
n
=

k=0
A
n
A
nk+1
B
nk
21
converges a.s., and the process {X
n
, n Z} is the unique strictly stationary solution of (27).
(b) Under the same conditions the process dened by (27) for t 1 has a nite limit distribution
as t , and this limit is the same irrespective of the initial condition X
0
.
There is no general formula to compute ({A
n
}) given the distribution of {A
n
}. We will give
some properties of the top Lyapunov exponent in the next subsection, and then show numerical
examples.
4.2. Existence of limit distribution under rst and second moment conditions. Sufcient
conditions for stability will now be given in terms of the rst and second moments of (A
1
, B
1
).
These are stronger conditions than the ones in Bougerol and Picard [4] (see Theorem 9), but
they are easier to verify. We use results from Conlisk [7] that lead to sufcient conditions for
stability of (27). See also [22] for similar results about a more general model. But rst we give
some relationships between spectral radius and Lyapunov exponent. The following results are
required for our analysis; they may or may not be known, but we were unable to nd proofs for
all of them in the literature.
The direct (or Kronecker) product A B of matrices A = (a
ij
)
mn
and B = (b
k
)
pq
is the
mp nq matrix
_
_
a
11
B a
1n
B
.
.
.
.
.
.
.
.
.
a
m1
B a
mn
B
_
_
.
We also use the vec operation, which stacks the columns of a matrix one on top of the other, the
rst column at the top. The main property of that operation is
vec(ABC) = (C
T
A)vecB.
Theorem 10. (a) Suppose the i.i.d. matrices {A
n
, n 1} satisfy EA
1
< . Then
({A
n
}) ({|A
n
|}) log (E|A
1
|).
If A
1
is deterministic then the second inequality is an equality.
(b) Suppose the matrices {A
n
, n 1} are i.i.d. and have nite second moments. Then
({A
n
})
1
2
log (E(A
1
A
1
)).
When A
1
is deterministic the two sides are equal.
(c) For arbitrary A
1
, if E(A
1
A
1
) is nite then E(A
1
) is also nite, and moreover
(E(A
1
))
2
(E(A
1
A
1
)). (30)
Proof. The condition EA
1
implies that E|A
1
| is a nite matrix, because |A
1
(i, j)| |A
1
e
j
|
e

A
1
, where e
j
is the unit vector with 1 in the j-th position and zeroes in the others; it also fol-
lows that Elog
+
A
1
< .
(a) Justication for the second inequality may be found in the proof of Theorem 2 of [16],
p.378), where non-negative A
n
are considered. Turn to the rst inequality, ({A
n
}) ({|A
n
|}).
It is known that A
n
A
1
0 if, and only if, ({A
n
}) < 0 (by Lemma 3.4 in [4]). Assume
that g = ({|A
n
|}) R and dene
C
n
= e
g
A
n
, n 1,
for some > 0. Then
|C
n
C
1
| |C
n
| |C
1
|
22
and
1
n
log |C
n
| |C
1
| = g +
1
n
log |A
n
| |A
1
| < 0,
which implies that C
n
C
1
tends to 0. Thus
0 > ({C
n
}) = g + ({A
n
}),
for any > 0, and it follows that ({A
n
}) g = ({|A
n
|}). There remains the case
({|A
n
|}) = ; this is seen to be equivalent to
limsup
1
n
log |e
M
A
n
| |e
M
A
1
| < 0
for all M > 0. This plainly implies that the same holds if {|A
n
|} is replaced with {A
n
}. (The
last assertion follows from the fact that ({|A
n
|}) = is equivalent to
limsup
n
1
n
log |A
n
| |A
1
| < M
for each M > 0, which is the same as
e
Mn
|A
n
| |A
1
| 0
as n tends to innity for all M > 0; this in turn implies
e
Mn
|A
n
A
1
| 0,
which entails ({A
n
}) < M for all M > 0.)
(b) Fix any x R
N
, and let
Z
n
= A
n
A
1
x.
Then Z
n
= A
n
Z
n1
and, letting V
n
= Evec(Z
n
Z
T
n
),
Z
n
Z
T
n1
= A
n
Z
n1
Z
T
n1
= V
n
= E(A
1
A
1
)V
n1
.
If (E(A
1
A
1
)) < 1 then V
n
tends to 0 at a geometric rate, i.e. for all n large enough
there is K < such that all the elements of V
n
are smaller than or equal to K
n
1
, where
(E(A
1
A
1
) <
1
< 1. Hence, for > 0
P(|Z
n
(j)| > )
EZ
n
(j)
2

2

K
n
1

2
.
Apply the Borel-Cantelli lemma: if {E
n
} is a sequence of events, then

n
P(E
n
) <
implies that P(E
n
innitely often) = 0. Let
E
n
= {|Z
n
(j)| > }.
Then

n
PE
n
< , implying that P(limsup |Z
n
(j)| > ) = 0. We conclude that if (E(A
1

A
1
) < 1 then Z
n
tends to zero a.s.. This holds for every x R
N
, and thus (E(A
1
A
1
) < 1
implies A
n
A
1
tends to 0, and in turn ({A
n
}) < 0.
Consider an arbitrary i.i.d. sequence {A
n
} with nite second moments, and note that (E((kA
1
)
(kA
1
)) = k
2
(E(A
1
A
1
) for any k > 0. If > 0 and
k

= e

[(E(A
1
A
1
)]

1
2
,
then (E((k

A
1
) (k

A
1
)) = e
2
, and thus ({k

A
n
}) < 0, implying that
({A
n
}) < +
1
2
log (E(A
1
A
1
))
for each > 0, and the inequality in part (b) follows. If A
1
is deterministic then =
log (A
1
) =
1
2
log (E(A
1
A
1
)), since (A
1
A
1
) = (A
1
)
2
.
(c) Among the elements of the matrix E(A
1
A
1
) there is E(A
1
(i, j)
2
), and if this is nite then
EA
1
(i, j) is also nite.
23
In [7] there is a proof that if (E(A
1
A
1
)) < 1, then (E(A
1
)) < 1 as well, from which
we could derive (30). That proof is not very intuitive, however, and we propose a more direct
argument. Let x C
N
and {A
n
} an i.i.d. sequence of N N matrices. For any complex Y it
is elementary that E|Y |
2
|EY |
2
. If M = A
n
A
1
and Y = x
T
Mx then
E|Y |
2
= E(x
T
Mxx
T
M
T
x)
= E[(x
T
x
T
)vec(Mxx
T
M
T
)]
= (x
T
x
T
)E(M M)vec(xx
T
).
Now, since (N
1
N
2
) (N
1
N
2
) = (N
1
N
1
)(N
2
N
2
),
E(M M) = E[(A
n
A
1
) (A
n
A
1
)] = E
n

j=1
(A
nj+1
A
nj+1
) = [E(A
1
A
1
)]
n
,
and thus
E|Y |
2
= (x
T
x
T
)[E(A
1
A
1
)]
n
vec(xx
T
) |EY |
2
= |Ex
T
Mx|
2
= |x
T
(EA
1
)
n
x|
2
.
If x is a non-zero eigenvector of E(A
1
) and the corresponding eigenvalue, then
(x
T
x
T
)[E(A
1
A
1
)]
n
vec(xx
T
) ||
2n
|x|
4
e
.
Divide both sides by
n
> (E(A
1
A
1
))
n
and then let n tend to innity, to nd 0 ||
2
/ < 1,
or ||
2
< . This is true for all eigenvalues of EA
1
and all > (E(A
1
A
1
)), which gives
(E(A
1
))
2
< , and nishes the proof.
Note that it is not always true that ({A
n
}) log (EA
1
) for matrices {A
n
} that have both
positive and negative entries; this happens in the second numerical example below.
Theorem 11. (a) If
(E|A
1
|) < 1, E|B|
e
<
then (27) is stable. Moreover, EX
t
is nite and satises
EX
t
= E(A
1
)EX
t1
+EB
t
. (31)
lim
t
EX
t
= (I
+m,+m
E(A
1
))
1
EB
1
= 0. (32)
(b) (Conlisk [7]) Suppose {(A
n
, B
n
)} are i.i.d. and have nite second moments. Then a suf-
cient condition for the system (27) to be stable is
(E(A
1
A
1
)) < 1.
When this is the case, the rst and second moments of X
t
are nite, the rst moments satisfy
(32) and second moments satisfy
vec E(X
t
X
T
t
) = E(A
n
A
t
) vec E(X
t1
X
T
t1
) + [E(B
t
A
t
) +E(A
t
B
t
)] vec EX
t1
+ vec E(B
t
B
T
t
)
lim
t
vec E(X
t
X
T
t
) = (I
(+m)
2
,(+m)
2 E(A
1
A
1
))
1
vec E(B
1
B
T
1
).
Part (b) of the theorem shows that the second-order conditions
log (E(A
1
A
1
)) < 1, E|
2
1
| < ,
ensure that Theorem 9 applies. Part (a) shows that the corresponding rst-order conditions have
the same consequence.
24
4.3. Numerical Experiments. In this section we will show plots of

n
=
1
n
log A
n
A
1

as n grows. This process is used to estimate the Lyapunov exponent ({A

n
}). In all our exper-
iments, we used simulation techniques for variance reduction and numerical stability methods
in order to increase accuracy of the estimation. We devised statistical tests in order to determine
the stopping criterion and we used asymptotic normality in order to produce statistical con-
dence intervals. The variables c
t
are i.i.d. and have a scaled beta distribution. The support of
this distribution is bounded and it is relatively easy to generate in simulations. The details of
the numerical methodology are beyond the scope of this paper and will be reported elsewhere.
0 10 20 30 40 50 60 70 80 90 100
!0.025
!0.02
!0.015
!0.01
!0.005
0
FIGURE 10. Top Lyapunov exponent estimation: a stable system.
m = 4, = 2
E(c
t
) = 2
log (E(A
t
)) = .0064
1
2
log (E(A
t
A
t
)) = 0.278
(0.0115, 0.0058)
Figure 10 shows the estimation process {
n
} for a stable system. Here c
t
is a scaled Beta(2,5)
distribution over the interval (0,7); its mean is 2 and variance is 1.25. The other parameters are
shown. In this case, the deterministic system driven by E(A
t
) has a negative spectral radius (i.e.
the average process (25) is stable). The bound
1
2
log (E(A
t
A
t
)) is also indicated. The
condence interval for the estimate shows that it is likely that < log (E(A
t
)).
The Lyapunov exponent and stability of the system do not depend on the particular zero-mean
iid sequence {
t
} (provided it has nite mean). Figure 11 shows a realisation of the price
sequence

t+
= c
t+
m

j=0

tj
+
t+
, N(0,
2
).
for system in Figure 10. As would be expected, increasing
2
has the effect of increasing the
variance of the price (the two plots have different scales).
Our next experiment shows an unstable system. Figure 12 shows the plot of the estimator
process
n
. The parameters are shown on the right. In this case the condence interval indicates
that > log (E(A
t
)). Here the distribution of c
t
is a scaled Beta(2,5) over the inteval (0,6.3),
with variance 1.0125.
In unstable cases the price sequence behaves more erratically, as would be expected.
m = 6, = 4
E(c
t
) = 1.8
log (E(A
t
)) = .0019
1
2
log (E(A
t
A
t
)) = 0.1967
(0.0155, 0.0191)
5. ACKNOWLEDGEMENTS
We warmly thank Donald MacLaren (Economics, University of Melbourne), Anne van den
Nouweland (Economics, University of Oregon) as well as Merab Menabde and Peter Stone
(BHP, Melbourne) for their helpful comments.
25
0 10 20 30 40 50 60 70 80 90 100
0.4
0.6
0.8
1
1.2
1.4
1.6
(a) = 0.01
0 10 20 30 40 50 60 70 80 90 100
0
1
2
3
4
5
6
7
8
(b) = 0.1
FIGURE 11. Price sequence for stable system.
0 10 20 30 40 50 60 70 80 90 100
0.004
0.006
0.008
0.01
0.012
0.014
0.016
0.018
0.02
FIGURE 12. Top Lyapunov exponent estimation: an unstable system
6. CONCLUSION
We have formulated a more general version of the cobweb model, that includes production lags
and explicit forecasting of prices. Power demand and supply functions, together with a focus
on log-prices, lead to more or less tractable difference equations for the price. The classical
cobweb theorem is shown to have extensions in those situations. Complex analysis helps to
understand what happens when parameters are changed. We have studied the effect of price
forecasting on stability; when the averaging period m is greater than one, stability requires less
stringent conditions on the elasticities than in the classical cobweb theorem. Increasing the
production lag may or may not lead to instability, but letting tend to innity leads to cycles
of constant amplitude. The random case is expressed as a bilinear model, and connects this
problem with recent work on chaos (Lyapunov exponent of random matrices). In this respect
we have provided some results and proofs that may be new.
REFERENCES
[1] Atkins, K., Marathe, A. and Barrett, C. (2007). A computational approach to modeling commodity markets,
Computational Economics 30: 125142.
[2] B elair, J. and Mackey, M.C. (1989). Consumer memory and price uctuations in commodity markets: An
integrodifferential model, Journal of dynamics and differential equations 1(3): 299328.
26
[3] Bougerol, P. and Picard, N. (1992). Stationarity of GARCH processes and some nonnegative time series,
Journal of Econometrics 52(1-2): 115127.
[4] Bougerol, P. and Picard, N. (1992). Strict stationarity of generalized autoregressive processes, Annals of
Probability 20(4): 17141730.
[5] Brandt, A. (1986). The stochastic equation Y
n+1
= A
n
Y
n
+ B
n
with stationary coefcients, Adv. Appl.
Prob. 18: 211-220.
[6] Chiarella, C. (1986). The cobweb model: Its instability and the onset of chaos, Economic Modelling 15(4):
377-384.
[7] Conlisk, J. (1974). Stability in a random coefcient model, International Economic Review 15: 529-533.
[8] van Doorn, J. (1975) Disequilibrium Economics, Macmillan studies in Economics, Macmillan Press LTD,
London.
[9] Dunsby, A., Eckstein, J. and Gaspar, J. (2008) Commodity Investing: Maximizing Returns Through Funda-
mental Analysis, Wiley.
[10] Evans, G.W and Honkapohja, S. (1988). Economic dynamics with learning: new stability results, Review of
Economic Studies 65: 2344.
[11] Ezekiel, M. (1938). The cobweb theorem, Quarterly Journal of Economics 52: 255280.
[12] Goldberg, S. (1986) Introduction to Difference Equations, Dover Publications Inc, New York.
[13] Goodwin, R. (1947). Dynamical coupling with special reference to markets having production lags, Econo-
metrica 15(3): 181204.
[14] Horn, R.A., and Johnson, C.R. (1985). Matrix Analysis. Cambridge University Press.
[15] H. Furstenberg and H. Kesten (1960). Products of Random Matrices, Ann. Math. Statist. 31(2): 457-469.
[16] Kesten, H. and Spitzer, F. (1984), Convergence in distribution of products of random matrices, Wahrschein-
lichkeitstheorie und Verw.(67): 363-386.
[17] J. F. C. Kingman (1973). Subadditive Ergodic Theory, Ann. Probability, 1: 883-909.
[18] Meadows, D.L. (1970) Dynamics of Commodity Production Cycles, Wright-Allen Press.
[19] Metzler, L. (1941). The nature and stability of inventory cycles. Review of Economics and Statistics 23(3):
113-129.
[20] Muth, J.F. (1961). Rational expectations and the theory of price movements, 29, 3: 315-335.
[21] Ng, S. and Ruge-Murcia, F. (2000). Explaining the persistence of commodity prices, Computational Eco-
nomics, 16: 149 - 171.
[22] Nicholls, D.F. and Quinn, B.B. (1982) Random Coefcient Autoregressive models: An Introduction, Lecture
Notes in Statistics, Springer, New York.
[23] Philips, A.W. (1957) Stabilization policy and the time-forms of lagged responses. Economic Journal
67(June): 265-277.
[24] Pryor, F.L., and Solomon, F. (1982) Commodity cycles as random processes. European Review of Agricul-
tural Economics 9: 327-346.
[25] Remmert, R. (1991) Theory of Complex Functions, Springer, New York.
[26] Samuelson, P. (1973) Mathematics of speculative price. SIAM Review 15(1): 2-42.
[27] Sterman, J.D. (2000). Business Dynamics: Systems Thinking and Model for a Complex World. McGraw Hill,
Boston, Massachussets.
[28] Smyth, D.J. (1973). Effect of public price forecasts on market price variation: a stochastic cobweb model,
American Journal of Agriculture, 55, 1: 83-88.
[29] Turnovsky, S.J. (1968). Stochastic stability of short-run market equilibrium under variations in supply,
Quarterly Journal of Economics, 82: 666-681.
[30] Vanneste, J. (2010). Estimating generalized Lyapunov exponents for products of random matrices, Phys. Rev.
E, 81(3): 036701.
[31] Vervaat, W. (1979). On a stochastic difference equation and a representation of non-negative innitely divis-
ible random variables, Adv. Appl. Probab. 11: 750783.
[32] Wheaton, W. (1999). Real estate cycles: some fundamentals, Real Estate Economics 27(2): 209-230.
DANIEL DUFRESNE, CENTRE FOR ACTUARIAL STUDIES, UNIVERSITY OF MELBOURNE
[email protected]
FELISA V AZQUEZ-ABAD, DEPARTMENT OF COMPUTER SCIENCE, HUNTER COLLEGE, CUNY
[email protected]
27