Go here to get my outline...
https://docs.google.com/open?id=0B1TnOdkv_G2hUXNlb2dmZUdFQms
Or if you insist on reading it in this terribly small screen on the blog it is also below.
Outline
Abstract
The major approaches to forecasting the market and the
concepts needed to grasp a basic understanding are reviewed. After each
approach support and counterarguments as to their use are given. Concluding the
section on prior knowledge the difference between modeling the market as a
dynamical system and as non-stationary system and the significance this
decision has on the development of future agents is discussed. Culminating with
a more in-depth explanation and expansion of social learning and supporting
arguments as well as counter arguments towards its use versus the other
possible methods. The Paper concludes with a short summary of why each method
alone is not optimally effective and why social learning with selective combination
is possibly a more effective method of prediction.
Thesis and support summary
Thesis
A combination of GP and ANNs as described in [1] and expanded with selective
combination as described in [2], [3] may be a more effective
method for market prediction.
Overview of current methods
Concepts and Current Methods
Statistical approaches
What is it?
Formally, Statistical approaches to prediction use
deterministic mathematic models to predict the market.
English Example
It is common knowledge that when a card is drawn from a deck
there is a 1/52 chance that the card drawn will be any given card, and as the
deck gets smaller or the cards that are left become known this probability
improves. When a blackjack player is counting cards he is constructing a
statistical model of the deck in his head so that he knows when there is a low
enough risk for him to bet.
Prior research
Generally in the market when a statistical method is used it
is taking advantage of a statistical phenomenon that is moldable. [4] and [5] are examples of this. In [4] and [5] the models developed take
advantage of the fact that humans will take more aggressive risks then may be
wise.
Why Not
While in [5] the statistical model was
successful for some time, when the humans involved in the trading were notified
of the strategy being used the humans modified how they were trading to
compensate and the statistical model quickly became nearly useless.
·
Generally bad at adapting to new situations. [5]
·
Will never learn from failures.
·
Generally only applicable to deterministic
systems [6]
WHY
·
Deterministic. (You know what it will do when
presented with any given situation)
·
Expressible as a mathematical model.
·
Fairly simple to implement. (This is because the
structure of the model is just a set of formulas no complicated structuring of
data is generally needed.)
Genetic Algorithms
What Is it?
Formally, Genetic Algorithms (abbreviated as GA or GP) are
used to solve problem where there are to many variables to find the exact
answer in a reasonable amount of time. GP accomplishes this by making a set of
guesses as to the answer, evaluating how close each guess is, and then finally
merging the best guesses in a variety of methods.
English example
GP was based off of how animals adapt to new situations over
generations. Evolution is a perfect example of a genetic algorithm, the “best”
animals of each generation mate and produce a set of offspring that is a mixing
of the qualities of both parents then the process is repeated thereby
continually optimizing the quality of the population.
Why Not
·
Can be slow
·
As with any AI GP can get stuck in what it
thinks is the best solution but in reality is not.
·
GP that is not able to adapt to new situations
is not only slow it is inaccurate [7].
WHY
·
It is possible for GP to adapt to new situations
[1], [7].
·
Instances of GP agents can share information and
help each other evolve more quickly (social learning) [1].
·
GP can be constructed in a way that makes it
possible to see what the agent has learned [7].
Prior research
·
[7] explores how to classify GP
algorithms and which classes are most effective, examples of how to implement
GP are also given along with what indicators are used inside of the algorithms.
The paper concludes that a class of GP called SFI (which simply means that they
can adapt to new situations are much
more effective a predicting the market).
·
[1] presents a method of using hybrid
GP and ANN agents in a manner where they are able to share information. This
method proves to be significantly more effective than the traditional manner.
Neural Networks
What Is it?
Neural Networks (abbreviated as ANN) are systems for
learning a pattern from a set of data and then recognizing the pattern again.
This is accomplished with a structure of layers of “neurons”. Each neuron in
each layer is connected to every neuron in the next layer through “synapses”
that are weighted. When a signal enters the network it is converted to a value
by the entry layer and then scaled by each layer and synapses it encounters on
its way out of the network. When the value exits the network if it is a above a
certain threshold it is interpreted as a recognition of the pattern.[8]
English example
ANNs roughly approximate how neurons work in nature. The
simplest way to understand neural networks is to look at diagram of one and
interact with one (insert diagram here and brain.js example from github.com).
Why Not
·
Difficult to explain
·
Difficult to implement
·
Nearly impossible to reclaim learned information
·
Difficult for ANNs to share learned information
·
Can get extremely large
WHY
·
Parallel (each neuron can have it’s own
processer since it does not need to know anything about the state of the other
neurons [8].)
·
The information gathered is Inductive [8] (later evidence is presented
that the market is driven by inductive decisions [9]).
·
Can be quickly trained [1].
Prior research
In the research reviewed for this paper ANNs were almost
always used in conjunction with GP [1], [7], [10], [11]. The research that used ANNs
alone were studying the nature of different type of neural networks [12] or using ANNs to identify
more useful indicators for use in other agents [13].
Dynamic Model vs. Non-Stationary Probabilistic
Dynamic model
This model of the market assumes that the market can
theoretically be deterministically modeled in a manner that accounts for all of
the behaviors expressed [14], [6]. A dynamic model means that the
system is extremely diverse in it’s outputs relative to the variation of its
inputs. Another example of a dynamic model is the weather, a very small change
in climate be it temperature or pressure can have massive effects on the rest
of the system (Think about the proposed effect of global warming, just a few
degrees difference massive change). This model is generally used for
Statistical agents.
Non-Stationary probabilistic model
This model views the market as a subjective entity, it’s nature changes with the beliefs of the individuals
[9]. This model appear here
because economy is driven by humans which make subjective decisions and so it
cannot be accurately predicted how they will react in the long term[9].
WHy use non-Stationary model
Humans drive the market. Subjective decisions are made
constantly. When the standard (Dynamic or Simple) model is forced to take into
account investors viewing the market differently they break down [9], [15].
Basis
Social Learning
The methods presented in [1] compensated for many of the
downfalls of GP.
·
Having multiple agents all learning with
different indicators and then sharing information the efficiency of evolution
was greatly increased. [1]
·
ANNs are effective in short term situations [13] where genetic algorithms alone are less so [7]. Combining the approaches
helps compensate.
·
Each agent managing multiple indicators enables
them to throw out old strategies and avoid local maxima [1], it also classifies the
agents as SFI GPs which are always more effective at market prediction[7].
Argument
Problems with social learning
·
Processing time (and possibly knowledge) is
wasted when ANNs are thrown away.
·
The ANNs only can have a limited view of the
market because it is either dynamical or non-stationary [1], [6], [9]. This may skew the efficacy
of a network
·
Only one ANN gets published at a time, the
reason an agent was successful could have been do the composite effect of
multiple networks.
Improvement
Composite the results of the ANNs using selective
combination (only acknowledging the results of a few of the networks to help
minimize the effect of failures on the group) as presented in [3], and [2] the methods presented in ,
and publish the group of ANNs that are used most instead of just the single
most accurate ANN.
·
Enables the Agents to recognize more complex
patterns
·
ANNs are not wasted are readily.
·
Combining results of ANNs has been shown to
improve overall accuracy and help to avoid local minima [3].
Conclusion
Using Statistical modeling alone is not effective because while
it is effective at exploiting a short term patter, it does not take the subjective
nature of the market into account well and does not evolve with the changes in
behavior that other agents will exhibit [9], [5].
Using ANNs alone works especially in short term predictions [13].
Using GP works [7] when the GP is able to evolve
with the changes in the market. But is inherently slower than ANNs and is not
parallelizable as easily (if at all) as ANNs [8].
The use of social learning and selective combination of ANNs
may result in more effective predictions. This is supported by evidence that
social learning works [1], and that combining neural
networks is effective [16], [2], [3]
[1] G.
Kendall, Y. Su, and G. Kendali, “Learning with imperfections - a multi-agent
neural-genetic trading system with differing levels of social learning,” in Cybernetics
and Intelligent Systems, 2004 IEEE Conference on, 2004, vol. 1, pp. 47-52
vol.1.
[2] A. J.
Sharkey, Combining Artificial Neural Nets: Ensemble and Modular Multi-Net
Systems, 1st ed. Secaucus, NJ, USA: Springer-Verlag New York, Inc., 1999.
[3] Z. Ahmad,
C. Technology, and N. U. Tyne, “A Comparison of Different Methods for Combining
Multiple Neural Networks Models,” pp. 828-833.
[4] G. C.
Calafiore, B. Monastero, and P. Torino, “Experiments on stock trading via
feedback control,” in Information and Financial Engineering (ICIFE), 2010
2nd IEEE International Conference on, 2010, pp. 494-498.
[5] J.
Grossklags and C. Schmidt, “Software agents and market (in) efficiency: a human
trader experiment,” Systems, Man, and Cybernetics, Part C: Applications and
Reviews, IEEE Transactions on, vol. 36, no. 1, pp. 56-67, 2006.
[6] T.
Iokibe, S. Murata, and M. Koyama, “Prediction of foreign exchange rate by local
fuzzy reconstruction method,” in Systems, Man and Cybernetics, 1995.
Intelligent Systems for the 21st Century., IEEE International Conference on,
1995, vol. 5, pp. 4051-4054 vol.5.
[7] M.
Kampouridis, S.-H. Chen, and E. Tsang, “Investigating the effect of different
GP algorithms on the non-stationary behavior of financial markets,” in Computational
Intelligence for Financial Engineering and Economics (CIFEr), 2011 IEEE
Symposium on, 2011, pp. 1-8.
[8] “Introduction
to artificial neural networks,” in Electronic Technology Directions to the
Year 2000, 1995. Proceedings., 1995, pp. 36-62.
[9] W. B.
Arthur, “Complexity in Economic and Financial Markets,” Complexity, pp.
20-25, 1995.
[10] G. Hai-ru,
H.-ru Guo, and Z.-min Li, “A method of improving generalization ability for
neural network based on genetic algorithm,” and Intelligent Systems (ICIS),
2010 IEEE, pp. 4-7, Oct. 2010.
[11] S.
Hayward, “Setting up performance surface of an artificial neural network with
genetic algorithm optimization: in search of an accurate and profitable
prediction of stock trading,” in Evolutionary Computation, 2004. CEC2004.
Congress on, 2004, vol. 1, pp. 948-954 Vol.1.
[12] E. W. Saad,
D. V. Prokhorov, D. C. W. II, and D. C. Wunsch, “Comparative study of stock
trend prediction using time delay, recurrent and probabilistic neural
networks,” Neural Networks, IEEE Transactions on, vol. 9, no. 6, pp.
1456-1470, 1998.
[13] X. Wang,
P. K. H. Phua, and W. Lin, “Stock market prediction using neural networks: Does
trading volume help in short-term prediction?,” in Neural Networks, 2003.
Proceedings of the International Joint Conference on, 2003, vol. 4, pp.
2438-2442 vol.4.
[14] K. S.
Narendra and A. U. Levin, “Regulation of Nonlinear Dynamical Systems Using
Multiple Neural Networks.”
[15] W. B.
Arthur, “Complexity and the Economy,” Science, vol. 284, no. 5411, pp.
107-109, 1999.
[16] C.-Y. Lee
and J.-J. Lee, “Adaptive Control for Uncertain Nonlinear Systems Based on
Multiple Neural Networks,” IEEE Transactions on Systems, Man and
Cybernetics, Part B (Cybernetics), vol. 34, no. 1, pp. 325-333, Feb. 2004.
