APPENDIX C
Information Cascades and the Bimodal Distribution of
Economic Recessions in the United States
Paul Ormerod
Volterra Consulting
October 2003
[email protected]
Abstract
The phenomenon of information cascades is widely observed in economic and social
systems. In a highly connected random network of interacting agents whose decisions
are determined by the actions of their neighbours according to a simple threshold
rule, the size distribution of cascades is bimodal. There is an exponential tail at small
cascade size and a second peak at the size of the entire system corresponding to a
single global cascade.
The American economy can be thought of as a highly connected network in terms of
both its technological and informational connections. The cumulative size of
economic recessions, the fall in output from peak to trough, is analysed for the US
economy 1900-2002. A least squares fit of an exponential relationship between size
and the rank of size gives a good description of most of the data. But the observation
for the Great Depression of the 1930s stands out as a very distinct outlier. In other
words, we observe a bimodal relationship of the type predicted by theory.
2
1. Introduction
The phenomenon of information cascades [1] is widely observed in economic and
social systems. Ref. [2] notes that small initial shocks can cascade to affect or disrupt
large systems that have proved stable with respect to similar disturbances in the past.
Examples include financial markets [3], the commercial success or failure of films
[4], and the diffusion of crime [5].
Watts [2] offers a general theoretical model of a possible explanation in terms of a
sparse, random network of interacting agents whose decisions are determined by the
actions of their neighbours according to a simple threshold rule. Two regimes are
identified in which the network is susceptible to very large cascades that occur very
rarely. When cascade propagation is limited by the connectivity of the network, a
power law distribution of cascades is observed. But when the network is highly
connected, the size distribution of cascades is bimodal, with an exponential tail at
small cascade size and a second peak at the size of the entire system corresponding to
a single global cascade.
The purpose of this short paper is to present evidence on the size of recessions in the
United States economy, which is an obvious example of a networked system.
2. The background
Firms in the economy are connected in two distinct ways. First, there are the
technological connections which arise from the need for companies in any particular
sector to use as inputs into their production the outputs of other industries. An
example is the motor vehicle industry, which uses materials produced by other
industries to make cars and trucks. Input-output tables in the national economic
accounts describe these connections at the level of the industry1. Ref [6] shows that in
a modern developed economy, the connections between industries are rather
extensive.
1
similar data is not available at the level of the firm
3
The second type of connection arises from the impact on any given firm of the
decisions and opinions about the future of other firms in the economy. Keynes [7],
for example, attached considerable importance to the state of long term expectations
for the decisions by companies to invest. The more optimistic the overall climate of
opinion, the more likely it is that a firm will carry out an investment decision. Ref [8]
sets out an agent based theoretical model based on this principle which is able to
generate a good approximation to some of the key underlying properties of US
business cycle time series data.
Information about the general climate of expectations is readily available from the
media, through newspapers such as the Wall Street Journal, for example. This is
certainly the case as far as large companies are concerned. In other words, it is as if
all agents are connected to, say, Microsoft or General Motors in terms of gathering
information about the business climate. The information they gather may not be
perfect, but a large amount of it is available.
So on both the technological and informational criteria, the US economy appears to be
highly connected.
The duration of recessions in 17 capitalist economies over the 1871-1994 period is
examined in [9], a recession being defined as a period of at least one year in which the
rate of growth of real GDP was less than zero. The individual country data on
duration was pooled, giving a total number of 336 observations across the 17 country
sample.
A power law relationship between frequency and duration of recessions provides a
good approximation to the data, but suggests that there should have been more
recessions of long duration than have actually existed. However, when recessions
lasting only one year are excluded from the sample, a power law relationship with an
exponent of –3.2 provides an almost perfect fit to the data. Ref [9] concludes that
‘there may be two separate processes going on in the process which generates data on
capitalist recessions. When a recession arises, for whatever reason, agents appear to
have some capacity to react quickly, which often prevents the recession from being
prolonged beyond one year. Once this has happened, however, recessions can take
4
place on all scales of duration’. In other words, there appears to be a bimodal type of
relationship in the duration of recessions in capitalist economies.
Ref [9] fitted power law relationships to the cumulative size of recessions, and
although such a relationship provides a rough approximation to the data, it is much
less satisfactory than the duration relationship.
3. The results
Percentage changes in real annual GDP are analysed for the United States over the
period 1900-2002. The data source from 1929 onwards is [10] and before that date
[11].
The cumulative size of recessions is the variable of interest, in other words the
percentage reduction in the level of GDP from peak to trough. There are in total 19
observations, ranging in size from the fall of barely one fifth of one per cent in GDP
in the recession of 1982, to the fall of some 30 per cent in the Great Depression of
1930-33. Figure 1 plots the histogram of the data, using absolute values.
5
8
6
Frequency
4
2
0
0 5 10 15 20 25 30
Cumulative percentage fall in GDP, absolute value
Figure 1 Histogram of cumulative absolute percentage fall in real US GDP,
peak to trough, during recessions 1900-2002
Most recessions have been fairly small. By far the largest recession was the Great
Depression of the early 1930s, and the second largest was the transition back into a
peace-time economy at the end of World War Two.
The cumulative density function is plotted in Figure 2.
6
1.0
0.8
0.6
Inverse CDF
0.4
0.2
0.0
0 5 10 15 20 25 30
Cumulative percentage fall in real US GDP
Figure 2 Cumulative density function of cumulative absolute percentage fall in
real US GDP, peak to trough, during recessions 1900-2002
A log-log plot of the absolute size of the recession against the log of the rank of the
size shows that although a power law provides an approximate description of this
data, it is not particularly good.
7
3
log of cumulative size of recessions
2
1
0
-1
0.0 0.5 1.0 1.5 2.0 2.5 3.0
log of rank of size
Figure 2 Log-log plot of cumulative size of recession and the rank of the
cumulative size
A formal regression shows that an exponential relationship between size and rank
gives a much better description, with the regression standard error being 0.347 for the
power law and 0.252 for the exponential.
However, the outlying observation of the Great Depression period is predicted very
inaccurately by the exponential fit. Table 1 shows the actual values of the six largest
cumulative recessions, and the values fitted by the exponential relationship between
size and rank. The smaller recessions are fitted very closely by the relationship.
Table 1 Six largest cumulative recessions, percentage change in real GDP,
actual and fitted from exponential least-squares regression of cumulative size on
rank
Actual 3.6 7.9 8.0 8.6 13.7 31.0
Fitted 5.0 6.3 8.0 10.1 12.7 16.1
8
The overall fit to the tail is not perfect, but the Great Depression clearly stands out as
an outlier.
Excluding the observation for the Great Depression makes little difference to the
estimated coefficient on the rank of the cumulative size. Using the whole sample it is
0.234 with a standard error of 0.011 and excluding the 1930s depression it is 0.221
with a standard error of 0.008. But the fit to the data is good, with the multiple R2
being 0.978 excluding the Great Depression observation.
In other words, the bulk of the data is fitted well, though not perfectly, by an
exponential relationship between size and rank of size, but there appears to be a
bimodal distribution. The observation for the Great Depression of the 1930s stands
out as a clear outlier.
The American economy is a highly connected network. In such circumstances, when
it is as if the degree of connectivity between agents is high, ref [2] suggests that a
bimodal distribution of cascades will be observed, with an exponential tail at small
cascade size, and a second peak with a much larger, global cascade size. This appears
to be very close to what we actually observe.
4. Conclusion
In a random network of interacting agents whose decisions are determined by the
actions of their neighbours according to a simple threshold rule, [2] suggests that
when the network is highly connected, a bimodal distribution of cascades will be
observed. There will be an exponential tail at small cascade size and a second peak at
the size of the entire system corresponding to a single global cascade.
The American economy can be thought of as a highly connected network, in terms of
both the technological connections between industries and the access to information
on the decisions and opinions of agents about economic prospects.
9
In such circumstances, when it is as if the degree of connectivity between agents is
high, ref [2] suggests that a bimodal distribution of cascades will be observed, with an
exponential tail at small cascade size, and a second peak with a much larger, global
cascade size. This appears to be very close to what we actually observe.
The cumulative percentage fall in real GDP from peak to trough in American
recessions 1900 – 2002 can be approximated closely by a bimodal distribution. An
exponential relationship between size and the rank of size fits the bulk of the data
well, but the single observation of the Great Depression of the 1930s stands out as a
very clear outlier.
References
1. S.Bikhchandani, D.Hirshleifer and I.Welch, Journal of Political Economy,
100, 992-1006, 1992
2. D.J.Watts, Proceedings of the National Academy of Science, 99, 5766-5771,
2002
3. A.Kirman, Bank of England Quarterly Bulletin, Bank of England, 1995
4. A.S.De Vany and W.D.Walls, Economic Journal, 106, 1493-1514, 1996
5. P.Ormerod, C.Mounfield and L.Smith, Non-linear modelling of crime in the
UK, Home Office Research Series, Home Office, London, UK, 2003 forthcoming
6. W.Cook and P.Ormerod, submitted to Physica A, 2003, pre-print available at
www.volterra.co.uk
7. J.M.Keynes, The general theory of employment, interest and money,
Macmillan, 1936
8. P Ormerod, Physica A, 314, 774-785, 2002
9. P.Ormerod and C.Mounfield, Physica A, 293, 573-582, 2001
10. Bureau of Economic Affairs, www.bea.gov
11. A.Maddison, Monitoring the world economy, OECD, Paris, 1995
10
11
