Data processing: artificial intelligence – Neural network
Reexamination Certificate
1998-09-03
2001-06-12
Powell, Mark R. (Department: 2122)
Data processing: artificial intelligence
Neural network
C706S015000
Reexamination Certificate
active
06247001
ABSTRACT:
BACKGROUND OF THE INVENTION
The invention relates to a method for training a neural network with training data that characterize a financial market. Attempts are constantly being made to predict changes in a financial market, in order in this way to achieve an optimal asset allocation, also known as portfolio management. By an asset allocation is meant the investment of liquid capital in various trade options, such as for example, stocks, futures, bonds, or also foreign currencies, thus in all possibilities offered by a financial market as a whole. A portfolio is formed with the goal of achieving a maximum return for a predeterminable risk that the investor is willing to take within a predeterminable time interval.
From E. Elton et al., Modern Portfolio theory and Investment Analysis, John Wiley & Sons, Inc., 4
th
edition, New York, ISBN 0-471-53248-7, pages 15-93, 1981, basic principals of asset allocation and investment analysis are known. From this article, a model is likewise known, called the two-point model, for the specification of an achievable return dependent on a risk taken by the investor in the investment opportunity. However, the attempt to make predictions concerning changes in the financial market on the basis of this model is very imprecise, since neither the chronological aspect of the changes of the financial market nor the transaction costs that arise in trades on the financial market are taken into account.
A further disadvantage that can be seen in the model described in this article is that actual data concerning changes of the financial market can in no way be taken into account. This leads to an inflexible, imprecise statement concerning changes of the financial market, based on the model.
In addition, what is known as a reinforcement learning method is known for example from A. Barto et al., Learning to Act Using Real-Time Dynamic Programming, Department of Computer Science, University of Massachusetts, Amherst, Mass. 01003, pages 1-65, January 1993.
SUMMARY OF THE INVENTION
The underlying problem of the invention is to train a neural network while taking into account parameters that characterize the financial market.
In general terms the present invention is a method for training a neural network, in which the following steps are iteratively executed:
a state vector is determined that has elements that characterize a financial market;
for the state vector, an evaluation relating to predetermined evaluation variables is determined; and
using a reinforcement learning method, weights of the neural network are adapted, at least on the basis of the evaluation of this state vector and on the basis of a determined evaluation of at least one following state vector.
In the inventive method, the asset allocation is modeled as a Markov decision problem. Each state of the Markov decision problem is described by a state vector containing elements that characterize the financial market. Since given a large number of influencing variables, the Markov decision problem is very high-dimensional, the reinforcement learning method is used for the training of the neural network. In the use of this method, both the state vector of a first time step and also the state vector of a following time step are taken into account.
This procedure for training a neural network has a number of advantages, in particular with respect to a further application of the neural network trained in this way in order to make statements concerning the changes in the financial market.
A considerable advantage of the inventive method is that chronological changes of the financial market are taken into account in the model.
Furthermore, data that have actually arisen in the past are taken into account in the training. The training data used for the training of the neural network, i.e. the respective state vector, can be used for the training without an actual outlay of capital by the investor for trades on the financial market. Through the use of a large amount of past data of the financial market, the model adapted on the basis of the data becomes very flexible with respect to changes of the financial market.
A further advantage of the inventive method is that the evaluation variables used for the evaluation of the state vectors are very flexible, and are thus easily exchanged or, respectively, expanded, dependent on the respective specific aspect of application. For example, here the classical evaluation variables of return and risk, or also further evaluation variables, such as an inflation rate or, for example, also safe custody fees, can be taken into account.
Advantageous developments of the present invention are as follows:
The state vector respectively has at least one of the following variables: at least one rate of at least one stock index; at least one indication of at least one bond market interest rate; at least one exchange rate indication for at least one first currency into at least one second currency; a gold price; and variables that describe a price-earnings ratio of at least one business enterprise.
The state vector respectively has at least one indication concerning a state of capital of an investor.
The predetermined evaluation quantities describe at least an investment risk and/or at least one determined return of at least one type of investment.
The predetermined evaluation variables describe at least transaction costs of at least one investment type. This takes into account transaction costs that arise due to trades on the financial market.
Given several possible following state vectors, a probability of occurrence is determined for all following state vectors. The evaluation is determined by means of summation of the evaluations of all possible following state vectors, multiplied by the respective probability of occurrence of the state vector. As a result a subsequent time step, are determined, whereby it is possible to train the neural network on the basis of several different successive states.
A reduction factor is provided, by means of which, as the number of iterative steps executed increases, the evaluations as a result of the at least one following state vector are reduced. As a result, an aspect is taken into account by means of which, as the number of iteration steps executed increases, evaluation results, for example the return for returns lying far in the future, are evaluated lower than the returns to be achieved in the near future.
The present invention is also a method for determining an investment decision using a neural network trained with a method described above, in which the following state vector that results given the respective investment decision is determined for at least two of all possible investment decisions, and in which that investment decision is recommended to a user that leads to a higher evaluation of the following state vector with respect to the evaluation variables.
The present invention is also a method for determining an investment decision using a neural network trained with a method described above in which the following state vector that results given the respective investment decision is determined for at least two of all possible investment decisions, and in which that investment decision is taken that leads to a higher evaluation of the following state vector with respect to the evaluation variables. Investment decisions are made using the neural network trained in the way described above, and these investment decisions are evaluated. In this way, a criterion is determined for estimating the quality of an investment decision in a financial market. For the investor, this can be an important aid in decisions relating to investments on the financial market, and thus in the formation of his specific asset allocation.
The present invention is further a method for determining an investment strategy using a neural network trained with a method described above in which several sequences of investment decisions, with various values of the evaluation variables, are determined, and in which a sequence of investment decisions that is opt
Neuneier Ralph
Tresp Volker
Holmes Michael B.
Powell Mark R.
Schiff & Hardin & Waite
Siemens Aktiengesellschaft
LandOfFree
Method of training a neural network does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Method of training a neural network, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Method of training a neural network will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2537257