Artificial neural network based universal time series

• US 6,735,580 B1
• Filed: 08/25/2000
• Issued: 05/11/2004
• Est. Priority Date: 08/26/1999
• Status: Active Grant

• Current Assignee: Shanghai Buyfund Internet Technology Co., Ltd.
• Original Assignee: Westport Financial LLC
• Inventors: Tang, Yi, Li, Bin, Li, Liang, Wu, Xiaohua
• Primary Examiner(s): Starks, Jr., Wilbert L.
• Assistant Examiner(s): Booker, Kelvin
• Application Number:
  US09/648,878
• Time in Patent Office: 1,355 Days
• Field of Search: 706/30, 706/21, 706/925, 706/15, 706/19, 705/10, 705/35
• US Class Current: 706/21
• CPC Class Codes: G06N 3/049 Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
• CPC Class Codes: Y10S 706/925 Business

Abstract

A neural network based universal time series prediction system for financial securities includes a pipelined recurrent ANN architecutre having a plurality of identical modules to first adjust internal weights and biases in response to a first training set representing a nonlinear financial time series of samples of a financial quantity and a target value, and then determine and store an estimated prediction error of the ANN in order to adjust short time stock price predictions in accordance with the stored prediction error. The prediction system is also designed to output upper and lower prediction bounds within a confidence region.

Claims

1. A computer implemented method of training a recurrent artificial neural network (ANN) for prediction of a response in a time-variable financial series of values comprising the steps of:

(a) collecting historical data comprised of a multiplicity of said values;

(b) determining an expected trend of the data and removing the expected trend from the data, wherein the expected trend of the data is determined by quantities of particular items to be predicted, trading volume and volatility or spread of quantities with respect to time;

(c) feeding a first group of said data to the recurrent ANN in a first training process;

(d) adjusting weights and biases of the ANN in response to the first training process;

(e) repeating steps (b) and (c) with additional groups of said values;

(f) determining an error in successive values from step (d) after said plurality of repetitions, wherein the ANN error represents the difference between each ANN output and a respective target sample;

(g) averaging said error over a plurality of the values;

(h) selecting a particular item for which a prediction of a response is desired;

(i) adjusting an output of ANN for the particular item with the determined average error; and

(j) outputting upper and lower error bounds of the adjusted output defining therebetween a probability range within which the adjusted output is predicted to remain for a period of time without fluctuating beyond the upper and lower error bounds.

2. The method defined in claim 1 wherein the expected trend of the data is determined as E=Σ

w(f(x)−

y)², wherein x is time, y is the quantities of the particular item to be predicted, f is the expected trend to be determined, w is the trading volume divided by the volatility or spread of the quantities y at the time x.

3. The method defined in claim 1 wherein the determined trend is added to the adjusted output before the outputting the lower and upper bounds.

4. The method defined in claim 1 wherein the probability range varies up to 80%, thereby giving at least only 20% chances that the adjusted output falls outside of the range.

5. A computer implemented method of predicting the future behavior of a financial time series comprising the steps of:

(a) collecting a first set representing a nonlinear financial time series of subsequent samples of a financial quantity combined in a plurality of groups, wherein the first sample of a subsequent group is a target sample of the previous group;

(b) feeding the first group of the samples and the first target sample of the first set of samples to an Artificial Neural Network (ANN);

(c) adjusting internal parameters of the ANN to minimize deviation of an ANN output from the target sample;

(d) feeding a plurality of subsequent groups of samples and their respective target samples of the first set until the target sample is the last sample of the set, thereby continuously adjusting the internal parameters of the ANN upon comparing each subsequent ANN output to the respective target sample;

(e) continuously re-feeding the previously processed groups of samples to gradually minimize an error function of the ANN between the ANN'"'"'s output and the respective target value;

(f) feeding a second set including a plurality of groups of samples and target samples to the ANN, thereby adjusting the internal parameters of the ANN in response to comparison between each new ANN output and a respective target sample;

(f′

) simultaneously with the step (f) determining an ANN error function representing the difference between each ANN output and the respective target sample;

(g) averaging the ANN error functions to determine an ANN estimated prediction error;

(h) predicting future results from any processed group of samples different from the first and second sets by adjusting the ANN output with the estimated prediction error; and

(i) outputting upper and lower error bounds of the adjusted ANN output at a given confidence level.

6. The method defined in claim 5 further comprising the steps of determining an expected trend of financial time series of subsequent samples representing each of the processed sets, removing the trend before processing of the groups of the respective set, and adding the trend before predicting the adjusted output.

7. The method defined in claim 6 further comprising the step of responding to any support computer language and applications using a Component Object Module technique.

8. A computer implemented method of training a recurrent artificial neural network (ANN) for prediction of a response in a time-variable financial series of values comprising the steps of:

(a) collecting historical data comprised of a multiplicity of said values;

(b) feeding the first group of said data to the recurrent ANN in a first training process;

(c) adjusting weights and biases of the ANN in response to the first training process;

(d) repeating steps (b) and (c) with additional groups of said values;

(e) determining an error in successive values after said plurality of repetitions, wherein the ANN error represents the difference between each ANN output and a respective target sample;

(f) averaging said error over a plurality of the values;

(g) selecting a particular item for which a prediction of a response is desired; and

(h) adjusting an output of ANN for the particular item with the determined average error.

9. The method defined in claim 8 wherein the step of feeding the first group of said historical data includes a step of feeding a first historically known target which is a value subsequent to the last value of said first group.

10. The method defined in claim 9 further comprising the step of generating an output by the ANN and comparing it with the target value, said weights and biases being adjusted to minimize an error function representing the difference between the output of the first training process and first historically known target.

11. The method defined in claim 8 wherein the first training process is repeated several thousand times, thereby adjusting internal weights and biases to gradually eliminate the error function.

12. The process defined in claim 8 wherein each of the additional groups of said values is fed to the ANN along with a respective target representing a value subsequent to the last one of the additional group, feeding of the additional groups of the values continues until the target is the last value from the collected data.

13. The method defined in claim 8 wherein each of the determined errors in successive values is stored.

14. The method defined in claim 8 wherein the comparison between the successive values and the respective target in step (e) is a one step training process.

15. The method defined in claim 8 wherein the time-variable financial series includes the group consisting of stock price series, bond price series, portfolios of securities, and volatility series.

16. The method defined in claim 8 wherein the adjusted output of ANN for the particular item is represented by a pair of upper and lower error bounds at a given confidence level, said confidence level being up to 80%.

17. An artificial neural prediction network system having a pipelined recurrent ANN (PRNN) architecture designed to train the ANN to predict with confidence, comprising:

a computer;

software executing on the computer for collecting a first set representing a nonlinear financial time series of subsequent samples of a financial quantity combined in a plurality of groups, wherein the first sample of a subsequent group is a target sample of the previous group;

a database having the collected set of samples of the financial security;

software executing on the computer for initially feeding the first group of the samples and the first target sample of the stored first set of samples;

software executing on the computer for processing the first group of samples and for adjusting the internal parameters of the PRNN to minimize deviation of the output from the target sample;

software executing on the computer for feeding a plurality of subsequent groups of samples and their respective target samples of the first stored set until the target sample is the last sample of the set, thereby continuously adjusting the internal parameters of the ANN upon comparing each subsequent output of the ANN to the respective target sample;

software executing on the computer for continuously feeding the previously processed groups of samples to obtain a minimal level of the deviation between the ANN'"'"'s output and the respective target value;

software executing on the computer for storing a second set of groups of samples in the database;

software executing on the computer for determining a prediction error between an output of the ANN in response to feeding each of the groups of samples of the second set and the respective target value;

software executing on the computer for averaging the prediction errors of the second set to obtain an estimated prediction error;

software executing on the computer for predicting future results from any processed group of samples different from the first and second sets, each subsequent future result being adjusted in accordance with the estimated prediction error; and

software executing on the computer for outputting upper and lower error bounds of the adjusted ANN output at a given confidence level.

18. The system defined in claim 17 further comprising software for determining an expected trend of financial time series of subsequent samples representing each of the processed sets, removing the trend before processing of the groups of the respective set, and adding the trend before predicting the adjusted output of a single security or a portfolio of securities.

19. The system defined in claim 17 wherein the given confidence level is up to 80%.

20. The system defined in claim 17 wherein the PRNN architecture uses a Component Object Module technique, so that said system can be readily called by any COM support computer language and applications.

21. A system employing a Recurrent ANN architecture (ANN) for prediction of a response in a time-variable financial series of values comprising:

a computer;

a database having a collected historical data comprised of a multiplicity of groups of successive time variable financial values;

software executed on the computer for continuously feeding a first group of said collected historical data to the ANN and comparing each subsequent output of the ANN with a target value which is selected from the historical data in a first training process;

software executed on the computer for adjusting weights and biases of the ANN in response to a plurality of repetitions of the first training process;

software for continuously feeding at least one second group of said collected historical data to the ANN and comparing its output to a respective target value selected from the historical data in a second training process;

software executed on the computer for adjusting weights and biases of the ANN in response to a plurality of repetitions of the second training process;

software for feeding a third group of said stored data and determining an ANN prediction error function in response to comparing the third group of said data to a respective target value;

software for selecting a particular item for which a prediction of a response is desired; and

software for adjusting the output of ANN for the particular item with the determined prediction error.

22. The system defined in claim 21 further comprising software executed on the computer for outputting upper and lower error bounds of the adjusted ANN output at a given confidence level.

23. The system defined in claim 21 wherein each target value of the respective training process is a value of the historical data which is subsequent to the last value of a previous group of values processed during a respective one of the training processes.