Book Read Free

Complexity and the Economy

Page 12

by W Brian Arthur


  creates new predictors by “mutating” the values in the predictor array, or by

  “recombination”—combining part of one predictor array with the comple-

  mentary part of another.

  The expectational system then works at each time with each agent observ-

  ing the current state of the market, and noticing which of his predictors match this state. He forecasts next period’s price and dividend by combining statistically the linear forecast of the H most accurate of these active predictors, and given this expectation and its variance, uses Eq. (5) to calculate desired stock holdings and to generate an appropriate bid or offer. Once the market clears,

  the next period’s price and dividend are revealed and the accuracies of the

  active predictors are updated.

  As noted above, learning in this expectational system takes place in two

  ways. It happens rapidly as agents learn which of their predictors are accurate and worth acting upon, and which should be ignored. And it happens on a

  slower time scale as the genetic algorithm from time to time discards non-

  performing predictors and creates new ones. Of course these new, untested

  predictors do not create disruptions—they will be acted upon only if they

  prove accurate. This avoids brittleness and provides what machine-learning

  theorists call “gracefulness” in the learning process.

  We can now discern several advantages of this multibit, multipredictor

  architecture. One is that this expectational architecture allows the market to have potentially different dynamics—a different character—under different

  states or circumstances. Because predictors are pattern-recognizing expec-

  tational models, and so can “recognize” these different states, agents can

  “remember” what happened before in given states and activate appropriate

  forecasts. This enables agents to make swift gestalt-like transitions in forecasting behavior should the market change.

  Second, the design avoids bias from the choice of a particular functional

  form for expectations. Although the forecasting part of our predictors is

  linear, the multiplicity of predictors conditioned upon the many combina-

  tions of market conditions yield collectively at any time and for any agent a

  nonlinear forecasting expression in the form of a piecewise linear, noncon-

  tinuous forecasting function whose domain is the market state space, and

  whose accuracy is tuned to different regions of this space. (Forecasting is, of course, limited by the choice of the binary descriptors that represent market

  conditions.)

  Third, learning is concentrated where it is needed. For example, J = 12

  descriptors produces predictors that can distinguish more than four thousand

  different states of the market. Yet, only a handful of these states might occur often. Predictor conditions that recognize states that do not occur often will be used less often, their accuracy will be updated less often and, other things being equal, their precision will be lower. They are, therefore, less likely to survive in the competition among predictors. Predictors will, therefore, cluster

  [ 50 ] Complexity and the Economy

  in the more visited parts of the market state space, which is exactly what we want.

  Finally, the descriptor bits can be organized into classes or informa-

  tion sets which summarize fundamentals, such as price-dividend ratios or

  technical-trading indicators, such as price trend movements. The design allows us to track exactly which information—which descriptor bits—the agents are

  using or ignoring, something of crucial importance if we want to test for the

  “emergence” of technical trading. This organization of the information also

  allows the possibility of setting up different agent “types” who have access to different information sets. (In this chapter, all agents see all market information equally.)

  A neural net could also supply several of these desirable qualities. However,

  it would be less transparent than our predictor system, which we can easily

  monitor to observe which information agents are individually and collectively

  using at each time.

  4. COMPUTER EXPERIMENTS: THE EMERGENCE

  OF TWO MARKET REGIMES

  Experimental Design

  We now explore computationally the behavior of our endogenous-expectations

  market in a series of experiments. We retain the same model parameters

  throughout these experiments, so that we can make comparisons of the

  market outcomes using the model under identical conditions with only con-

  trolled changes. Each experiment is run for 250,000 periods to allow asymp-

  totic behavior to emerge if it is present; and it is run 25 times under different random seeds to collect cross-sectional statistics.

  We specialize the model described in the previous section by choosing

  parameter values, and, where necessary, functional forms. We use N = 25

  agents, who each have M = 100 predictors, which are conditioned on J = 12

  market descriptors. The dividend follows the AR (1) process in Eq. (4), with

  autoregressive parameter ρ set to 0.95, yielding a process close to a random walk, yet persistent.

  The 12 binary descriptors that summarize the state of the market are the

  following:

  1–6 Current price × interest rate/dividend > 0.25, 0.5, 0.75, 0.875, 1.0, 1.125

  7–10 Current price > 5-period moving average of past prices (MA), 10-period MA, 100-period MA, 500-period MA

  11

  Always on (1)

  12

  Always off (0)

  a sse t Pr icing under endogenous exPectat ion s [ 51 ]

  The first six binary descriptors—the first six bits—reflect the current price in relation to current dividend, and thus, indicate whether the stock is

  above or below fundamental value at the current price. We will call these

  “fundamental” bits. Bits 7–10 are “technical-trading” bits that indicate

  whether a trend in the price is under way. They will be ignored if useless,

  and acted upon if technical-analysis trend following emerges. The final two

  bits, constrained to be 0 or 1 at all times, serve as experimental controls.

  They convey no useful market information, but can tell us the degree to

  which agents act upon useless information at any time. We say a bit is “set”

  if it is 0 or 1, and predictors are selected randomly for recombination, other things equal, with slightly lower probabilities the higher their specificity—

  that is, the more set bits they contain (see Appendix A). This introduces a

  weak drift toward the all-# configuration, and ensures that the informa-

  tion represented by a particular bit is used only if agents find it genuinely

  useful in prediction. This market information design allows us to speak of

  “emergence.” For example, it can be said that technical trading has emerged

  if bits 7–10 become set significantly more often, statistically, than the con-

  trol bits.

  We assume that forecasts are formed by each predictor j storing values for the parameters a , b , in the linear combination of price and dividend, E [ p +

  j

  j

  j

  t + 1

  d | I ] = a ( p + d ) + b . Each predictor also stores a current estimate of its t + 1

  t

  j

  t

  t

  j

  forecast variance. (See Appendix A).

  Before we conduct experiments, we run two diagnostic tests on our

  computer-based version of the mo
del. In the first, we test to see whether

  the model can replicate the rational-expectations equilibrium (r.e.e.) of

  standard theory. We do this by calculating analytically the homogeneous

  rational-expectations equilibrium (h.r.e.e.) values for the forecasting param-

  eters a and b (see Appendix A), then running the computation with all predictors “clamped” to these calculated h.r.e.e. parameters. We find indeed that such predictions are upheld—that the model indeed reproduces the h.r.e.e.—

  which assures us that the computerized model, with its expectations, demand

  functions, aggregation, market clearing, and timing sequence, is working

  correctly. In the second test, we show the agents a given dividend sequence

  and a calculated h.r.e.e. price series that corresponds to it, and test whether they individually learn the correct forecasting parameters. They do, though

  with some variation due to the agents’ continual exploration of expectational

  space, which assures us that our agents are learning properly.

  The Experiments

  We now run two sets of fundamental experiments with the computerized

  model, corresponding respectively to slow and medium rates of exploration

  [ 52 ] Complexity and the Economy

  by agents of alternative expectations. The two sets give rise to two different regimes—two different sets of characteristic behaviors of the market. In the slow-learning-rate experiments, the genetic algorithm is invoked every

  1,000 periods on average, predictors are crossed over with probability 0.3,

  and the predictors’ accuracy-updating parameter θ is set to 1/150. In the medium-exploration-rate experiments, the genetic algorithm is invoked every

  250 periods on average, crossover occurs with probability 0.1, and the pre-

  dictors’ accuracy-updating parameter θ is set to 1/75.10 Otherwise, we keep the model parameters the same in both sets of experiments, and in both

  we start the agents with expectational parameters selected randomly from

  a uniform distribution of values centered on the calculated homogeneous

  rational-expectations ones. (See Appendix A.) In the slow-exploration-rate

  experiments, no non-r.e.e. expectations can get a footing: the mar-

  ket enters an evolutionarily stable, rational-expectations regime. In the

  medium-exploration-rate experiments, we find that the market enters a com-

  plex regime in which psychological behavior emerges, there are significant

  deviations from the r.e.e. benchmark, and statistical “signatures” of real financial markets are observed.

  We now describe these two sets of experiments and the two regimes or

  phases of the market they induce.

  The Rational-Expectations Regime

  As stated, in this set of experiments, agents continually explore in prediction space, but under low rates. The market price, in these experiments, converges

  rapidly to the homogeneous rational-expectations value adjusted for risk,

  even though the agents start with nonrational expectations. In other words,

  homogeneous rational expectations are an attractor for a market with endog-

  enous, inductive expectations.11 This is not surprising. If some agents forecast differently than the h.r.e.e. value, then the fact that most other agents are

  using something close to the h.r.e.e. value will return a market-clearing price that corrects these deviant expectations: There is a natural, if weak, attraction to h.r.e.e. The equilibrium within this regime differs in two ways from

  the standard, theoretical, rational-expectations equilibrium. First, the equilibrium is neither assumed nor arrived at by deductive means. Our agents instead

  arrive inductively at a homogeneity that overlaps that of the homogeneous,

  10. At the time of writing, we have discovered that the two regimes emerge, and the results are materially the same, if we vary only the rate of invocation of the genetic algorithm.

  11. Within a simpler model, Blume and Easley (1982) prove analytically the evolutionary stability of r.e.e.

  a sse t Pr icing under endogenous exPectat ion s [ 53 ]

  theoretical rational expectations. Second, the equilibrium is a stochastic one.

  Agents continually explore alternatives, albeit at low rates. This testing of

  alternative explorations, small as it is, induces some “thermal noise” into the system. As we would expect, in this regime, agents’ holdings remain highly

  homogeneous, trading volume remains low (reflecting only variations in fore-

  casts due to mutation and recombination), and bubbles, crashes, and technical

  trading do not emerge. We can say that in this regime the efficient-market

  theory and its implications are upheld.

  The Complex or Rich Psychological Regime

  We now allow a more realistic level of exploration in belief space. In these

  experiments, as we see in Figure 1, the price series still appears to be nearly identical to the price in the rational—expectations regime. (It is lower because of risk attributable to the higher variance caused by increased exploration.)

  On closer inspection of the results, however, we find that complex patterns

  have formed in the collection of beliefs, and that the market displays charac-

  teristics that differ materially from those in the rational-expectations regime.

  For example, when we magnify the difference between the two price series,

  we see systematic evidence of temporary price bubbles and crashes (Figure 2).

  We call this new set of market behaviors the rich psychological, or complex,

  regime.

  This appearance of bubbles and crashes suggests that technical trading, in

  the form of buying or selling into trends, has emerged in the market. We can

  check this rigorously by examining the information the agents condition their

  forecasts upon. Figure 3 shows the number of technical-trading bits that are

  used (are 1’s or 0’s) in the population of predictors as it evolves over time. In both sets of experiments, technical-trading bits are initially seeded randomly in the predictor population. In the rational-expectations regime, however,

  technical-trading bits provide no useful information and fall off as useless

  predictors are discarded. But in the complex regime, they bootstrap in the

  population, reaching a steady-state value by 150,000 periods. Technical trad-

  ing, once it emerges, remains12.

  Price statistics in the complex regime differ from those in the

  rational-expectations regime, mainly in that kurtosis is evident in the com-

  plex case (Table 1) and that volume of shares traded (per 10,000 periods) is

  about 300% larger in the complex case, reflecting the degree to which the

  agents remain heterogeneous in their expectations as the market evolves. We

  12. When we run these experiments informally to 1,000,000 periods, we see no signs that technical-trading bits disappear.

  [ 54 ] Complexity and the Economy

  100

  95

  90

  85

  80

  Price

  75

  70

  65

  60

  253000

  253050

  253100

  253150

  253200

  Time

  Figure 1:

  Rational-expectations price vs. price in the rich psychological regime. The two price series are generated on the same random dividend series. The upper is the homogeneous r.e.e.

  price, the lower is the price in the complex regime. The higher variance in the latter case causes the lower price through risk aversion.

  note that fat tails and high volume are
also characteristic of price data from actual financial markets.

  How does technical trading emerge in psychologically rich or complex

  regime? In this regime the “temperature” of exploration is high enough to

  100

  R.e.e Price

  80

  60

  40

  20

  Price Difference

  0

  –20

  253000

  253050

  253100

  253150

  253200

  Time

  Figure 2:

  Deviations of the price series in the complex regime from fundamental value. The bottom graph shows the difference between the two price series in Figure 1 (with the complex series rescaled to match the r.e.e. one and the difference between the two doubled for ease of observation). The upper series is the h.r.e.e. price.

  a sse t Pr icing under endogenous exPectat ion s [ 55 ]

  600

  500

  Complex Case

  400

  300

  Bits Used

  200

  100

  R.E.E. Case

  0

  0

  50000

  100000

  150000

  200000

  250000

  Time

  Figure 3:

  Number of technical-trading bits that become set as the market evolves (median over 25

  experiments in the two regimes).

  offset, to some degree, expectations’ natural attraction to the r.e.e. And so, subsets of non-r.e.e. beliefs need not disappear rapidly. Instead they can

  become mutually reinforcing. Suppose, for example, predictors appear early on

  that, by chance, condition an upward price forecast upon the markets showing

  a current rising trend. Then, agents who hold such predictors are more likely

  to buy into the market on an uptrend, raising the price over what it might

  otherwise be, causing a slight upward bias that might be sufficient to lend validation to such rules and retain them in the market. A similar story holds for

  predictors that forecast reversion to fundamental value. Such predictors need

  to appear in sufficient density to validate each other and remain in the popu-

  lation of predictors. The situation here is analogous to that in theories of the origin of life, where there needs to be a certain density of mutually reinforcing RNA units in the “soup” of monomers and polymers for such replicating

 

‹ Prev