2
Agenda
 Introduction
 Problem statement
 Methodology
 Experiment Results
 Conclustion
 Reference

Introduction
3
 Stock forecast, is a branch of economic forecasts,
which use the accurate survey statistics and stock
market information as the basis.
 If we can predict the stock's ups and downs, and
the stock market in a timely manner to reasonable
regulation and with health guide, it will continue to
develop our economy to provide a solid backing.

Introduction(cont)
4
 Existence of high yield and high-risk
characteristics in the stock market.
 people are continuing to explore its internal rules.
 many traditional time series analysis methods.
 Exponential smoothing method, ARMA (Auto
Regressive Moving Average Model) .
 ARCH (Auto Regressive Conditional
Heteroskedasticity Model)

Problem statement
 Paper presents a BP Artificial neural
network prediction modeling method for
forecasting the end of Shanghai index.
 Paper Uses the genetic algorithm to
optimize the BP network parameters,
weight and structure.
5

Artificial Neural Network(Cont.)

Artificial Neural Network(Cont.)
 error back-propagation algorithm
 Its:
 error-correction learning rule.
 BP ANN Is Learning By Two
Teqhniques:
Supervised learning
Unsupervised learning

Genetic Algorithm
10
http://ib-poland.virtualave.net/ee/genetic1/3geneticalgorithms.htm

GA Steps To Optimize BP ANN:
 Intialization:
p , Crossover Scale – pc Crossover
Probability
Pm Mutation Probability
WIH(ji ) Connection Weights of Hidden L.
WHO(ji) Weights of Output L.

GA Steps To Optimize BP ANN:
 Coding :
Real Number Coding .
Initial Pobulation Takes 30

Fitness Function
. F(i) fitness value of indvidual i.
. Sum(E) sum of the squares errors

Fitness Function
 i=1…..N number of chromosomes.
 K= 1……4 for the number of output Layers
 P=1…….5 the study sample size
 T(k) Teacher Signal

Using genetic algorithm to optimize the
weights of the neural network
 1) Initialize: Initialize population P, including
crossover scale, Pc ,Pm and initialization for WIHij
and WHOji, Paper Author use the real number
coding, and the initial population take 30.
 2) Select and Computing fitness: each individual
evaluation function, and sort them; we can choose the
network by the probability value that show in
Formula; 15

Using genetic algorithm to optimize
the weights of the neural network

Using genetic algorithm to optimize the weights of
the neural network
 3) crossover: Individual G i and G i+1 crossover operation with probability Pc
to generate new individuals 'G i and , G i+1.
 4) mutate: Individual Gj mutate by probability Pm, and then produce new
individuals , Gj.
 5) evaluate new pop: Put the individuals into the new population P, and
calculate the new evaluation function of the individual.
 6) decide satisfactory: If you find a satisfactory individual, then the end, or
switch to 3).
 After achieve the required performance indicators, will eventually decode the
group's best individual you can get the optimized network connection weights.
17

Genetic algorithm to optimize the BP
network’s workflow
18

Establish Forecasting Model
 BP ANN :
 3 Layers ANN.
 parameters of Related Training
 start Training
 Use G A To Optimize ANN Weights
 Train the Optimized ANN Again
 Use the Optimized ANN To test Samples.

BP neural Network Weights
Optimization
 We need to use GA for BP weights to
be Optimized.
 Initialize Weights Encoding and Fitness
Calculations.
 Choose new Generation According to
Fitness.
 Repeat until Getting a set of Weights
to meet the Accuracy Req’s.

Training of BP ANN (again)
 Asseign the Weights and Threshold
Optimized to the BP ANN.
 Use training Sample To Train The BP
ANN again .
 Untill NN o/p and Sample o/p Tailed .
 Terminate the Trainig.

Experiment Results(Cont.)
 After a series of training, eventually selected parameters are:
a) Population scale: popu=30
b) Selection rate: opti=0.09
c) Crossover: arithXover
d) Crossover rate: Pc=0.95
e) Mutation: nonUnifMutation
f) Mutation rate: Pm=0.1
g) Genetic generations: gen=120
22

Experiment Results(Cont.)
 The training of the BP neural network after optimized Using genetic algorithm
program assign the weights and threshold (W1, B1, B2 W2) that after
optimized to the BP neural network.
 Use the training sample, to train BP network again, each 2,000 times, until
network output and sample output tallied, terminate the training.
 Stock Index Forecast Using the established GA-BP neural network based
stock index forecasting model to predict the stock price index Output the
results of the model predicted values and target values, and draw curve, to
used to verify the prediction accuracy, operability and practicability.
23

Experiment result(Cont.)
 R .. BP ANN Target Value
 P .. Predicted Value
 E .. Absolute Error.

Conclusion
 GA-BP algorithm used to avoid the BP algorithm into a local minimum,
slow convergence problem, and also to overcome the GA in a similar
form of exhaustive search for optimal solution search time caused by
long, slow shortcomings, is a fast, reliable method.
 Paper results shows that BP neural network using GA for the learning of
rules and to optimize the network weights and weights of the network and
the fixed threshold can improve the accuracy of stock index prediction
model.
27

References
 [1] Shen Bing. Equity Investment Analysis [M]. Chongqing: Chongqing
Publishing House, 2002: 94.
 [2] Chu Yuan. Securities Investment Principles [M]. Beijing: Lixin Accounting
Book Publishing ,2003:74-78.
 [3] Liu Yong. China's stock market and the empirical relationship between
macroeconomic variables [J]. Finance and Trade Economics, 2004 (4): 21-27.
 [4] Zhang Ling Song, Tao Chongen. Stock technical analysis tool [M]. Beijing:
China Encyclopedia Publishing House, 1994: 52-56 .
 [5] Ma Weihua, LI Yu-hong. Stock index futures and stock market development
in China [J]. Finance Teaching and Research, 2004, (5): 50-54.
 [6] E.W. Saad, D. V. Prokhorov, D.C. Wunsch. Comparative Study of Stock
Trend Prediction Using Time Delay, Recurrent and Probability Neural Networks.
IEEE Trans on Nerual Netowrks, 1998, 9(6): 1 4561 470.
28

Stock market analysis using ga and neural network

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to Stock market analysis using ga and neural network (20)

More from Amr Abd El Latief (12)

Recently uploaded (20)

Stock market analysis using ga and neural network