PREDICTING IMPACT OF NEWS ON STOCK PRICE: AN EVALUATION OF NEURO FUZZY SYSTEM Congress on Evolutionary Computation (CEC 2007) Presented by CUI, Weiwei.

PREDICTING IMPACT OF NEWS ON STOCK PRICE: AN EVALUATION OF NEURO FUZZY SYSTEM

Congress on Evolutionary Computation (CEC 2007)

Presented by CUI, Weiwei

In COMP630P 2009 - HKUST

OUTLINE

Introduction and literature review Specification of neuro fuzzy networks News Coding Experiment and data Discussion of findings Conclusion and Comments

INTRODUCTION

News implicitly affects financial markets News investors stock price Political, economic, financial, macro, micro… Released when the security markers are open or

closed No attempt to study the impact of all news in total

Neural Fuzzy (NF) Systems Predicting complex, non-linear relationships Multiple variables No specific pattern of distribution of data

NF systems are different Different levels of competences and capabilities

OBJECTIVE OF PAPER

Evaluate the effectiveness of four NF systems Feed Forward Neural Network (FFNN) Adaptive Neuro Fuzzy Inference System (ANFIS) Radial Basis Function Network (BRFN) Rough Set Based Pesudo Outer Product Rule

(RSPOP) Apply these four NF systems on the same

dataset Recommend a system for more detailed

analysis based on the experimental results

PAST STUDIES

Pure expert analysis “The number of Dow Jones announcements and the

aggregate measures of securities market activity such as trading volumes and market returns are related”

- Mitchell and Mulherin (1994)

“The arrival of public information in the U.S. Treasury Market sets off a two stage adjustment process for prices, trading volume, and bid-ask spreads”

- Fleming and Remolona (1999)

“Investors in Asian markets tend to react more significantly to negative stock news originating from US sources than they do to positive news”

- Doong et al. (2005)

NF SYSTEMS V.S. STATISTICAL MODELS

NF networks have proven to be better Soft computing approaches synthesizing human

ability to process uncertain, imprecise, and incomplete information to make decisions

High-level linguistic model instead of low-level complex mathematical expressions

Ability to self-adjust the parameters and derive intrinsic relationships between selected inputs and outputs

OUTLINE

Introduction and literature review Specification of neuro fuzzy networks News Coding Experiment and data Discussion of findings Conclusion and Comments

SPECIFICATION OF NF SYSTEMS

Feed Forward Neural Network (FFNN) Radial Basis Function Network (BRFN) Adaptive Neuro Fuzzy Inference System

(ANFIS) Rough Set Based Pesudo Outer Product Rule

(RSPOP)

FFNN ANFISBRFN RSPOP

FEED FORWARD NEURAL NETWORK

Multilayer Perceptron (MLP) Most popular type of neural

networks Back-propagation to update

the weights Simplest form of a MLP model

Benchmark? Not good at prediction of a

time series data Influence of the anterior data?

RADIAL BASIS FUNCTION NETWORK

First used to solve interpolation problems

Fitting a curve exactly through a set of points Weighted distances are

computed between the input x and a set of prototypes

These scale distances are then transformed through a set of nonlinear basis functions h, and these outputs are summed up in a linear combination with the original inputs and a constant.

ADAPTIVE NEURO FUZZY INFERENCE SYSTEM

Combine world-fuzzy logic systems and neural networks Representing prior expert knowledge into a set of

fuzzy membership functions Reducing the optimization search space Adapting the back-propagation to automate

fuzzy controller parametric tuning tuning

Layer 1: Fuzzy member functionLayer 2: MultiplicationLayer 3: NormalizationLayer 4: Production of the input and a first order polynomialLayer 5: Sum

ROUGH SET BASED PESUDO OUTER PRODUCT RULE

Combine the concept of rough set theory and presudo outer product rule Automatically formulate

the fuzzy rules from the numberical training data

No initial rule base needs to be specified

Layer 1: Each input node represents an input linguistic variableLayer 2: Each input label node represents a fuzzy member functionLayer 3: Each rule node represent an if-then fuzzy rulesLayer 4: Each output label node represents a fuzzy member functionLayer 5: Each output node represents an output linguistic variable

COMPARISON

Prior Knowledge

Layer #

Type Advantage

FFNN No need 3 Numerical Simplest

RBFN No need 3 Numerical Interpolation problem

ANFIS Need 5 Linguistic Use prior knowledge to reduce optimization search space

RSPOP No need 5 Linguistic Reduction of attributes and fuzzy rules

TIME SERIES PREDICTION USING NN

Represent target values by the successive relative changes in prices since the previous time point rather than absolute prices after a fixed time horizon

General n-dimensional discrete time dynamic system:

Reconstruct the phase space form the time series data by delay coordinates

OUTLINE

Introduction and literature review Specification of neuro fuzzy networks News Coding Experiment and data Discussion of findings Conclusion and Comments

NEWS CODING

2-Value News Coding Method (2-NCM) Binary coding: There is news for the day or there

is no news Penta Coding Method (PCM)

Categorical info: Classify the contents of news and to ascertain the impact of different categories of news items

2-VALUE NEWS CODING METHOD

Let L be the set of news on the company and T be the Time for which news data is classified

The coding is decided manually based on the headlines extracted from database

PENTA CODING METHOD (PCM)

News category (priority in ascending order): No news LC – News pertaining directly to Company operations

Splits, dividends, bonus, successfulness of product launch LP – Performance related news

Quarterly or annual financial report LM – Macro-environmental changes

Interest rate change Government or regulatory policy news

LO – Other news Major stock index rise/fall without any particular reason Natural or man-made disasters

PENTA CODING METHOD (PCM)

Let L be the set of news on the company and T be the Time for which news data is classified

OUTLINE

Introduction and literature review Specification of neuro fuzzy networks News Coding Experiment and data Discussion of findings Conclusion and Comments

DATA: STOCK PRICES AND NEWS

DBS = Development Bank of Singapore UBO = United Overseas Bank ExMobile = Exxon Mobil

(News was obtained by running a single keyword search with the company names)

EXPERIMENT

Two measures of performance were used: Root mean square error Pearson’s coefficient of correlation

Two results of 2-NCM and PCM were benchmarked against the results form their corresponding setup with only stock prices as inputs

OUTLINE

Introduction and literature review Specification of neuro fuzzy networks News Coding Experiment and data Discussion of findings Conclusion and Comments

RESULTS ON DBS AND UOB

2-NCM: No significant advantage Low interpretability of the news input: a binary

input along with a set of prices PCM: Also no significant advantage

Small amount of training data available to the network

Databases do not keep sufficient information fora small stock like DBS

Singapore is a very controlledmarket

PCM ON APPLE AND EXXON MOBILE

Results are positive FFNN is a primitive model Consistent improvement across RBFN, ANFIS,

and RSPOP Error down by 1.1% for Apple, 1.49% for Exxon

CHANGE IN STOCK PRICE PREDICTION

Legend C: error reduction by $1.72 on 19 Oct. Code 3 news: performance related news Benchmark model is right about the movement

direction

CHANGE IN STOCK PRICE PREDICTION

Legend A: error reduction by $1.13 on 28 Dec. Code 5 news: other news

‘US stock Index Futures Decline; Home Depot, Apple Fall’

Stock price had moved up by $4.03, but benchmark model shows none

CHANGE IN STOCK PRICE PREDICTION

Legend K: error reduction by $0.4 on 29 Jun. Code 4 news: Macro-environmental changes

Apple started investigating stock option grants Not inputting impact direction, it might be dicey

for the network to predict correctly

CHANGE IN STOCK PRICE PREDICTION

Error increase: Legend H: lawsuit Legend D: ‘Reports Findings of Stock Option’ Legend E: ‘Google Inc. CEO Joins Apple

Computer’

CHANGE IN STOCK PRICE PREDICTION

All reductions are at points where the stock has taken a sharp jerk

It is not predictable based on historical past patterns

OUTLINE

Introduction and literature review Specification of neuro fuzzy networks News Coding Experiment and data Discussion of findings Conclusion and Comments

CONCLUSION

Propose, implement , and evaluate the impact of news on stock prices on a short term

News input could increase accuracy in most cases, or at least maintain the performance of the current models.

Two facts increase the prediction accuracy: Large database of news Volatility exhibited by price fluctuations

FFNN degrade results, RSPOP is best

COMMENTS

Many pages for introduction; a few words about experiments; almost no experimental details; results and conclusion are too obvious

Poorly written (typos, missing labels, copied sentences from references)

Problems: Manual coding? PCM Categories are based on? News can override one another? Just considering the news type? What about

sentiment?