P REDICTING THE D AILY R ETURNS USING F INANCIAL Q UANTITATIVE D ATA AND ASX A NNOUNCEMENTS Zhendong Zhao (4238 8910) Supervisor: Mark Johnson.

PREDICTING THE DAILY RETURNS USING FINANCIAL QUANTITATIVE DATA AND ASX ANNOUNCEMENTS

Zhendong Zhao (4238 8910)

Supervisor: Mark Johnson

OUTLINE

Motivations

Framework & formulation

Dataset & features

Experiments

MOTIVATIONS – TO PREDICT THE DAILY RETURNS Previous works

Use either textual or financial quantitative features

Our workUse heterogeneous features (both textual and financial quantitative features)

Textual Features(Announcements,

Financial News, etc.)

Financial Quantitative Features(past daily returns,

past trading value, etc.)

The daily Returns

EXAMPLES

THE DAILY RETURNS (LOG)

Where is the close price of today, and is the close price of yesterday.

For example: = 1$ = 1.2$

OUTLINE

Motivations

Framework & formulation

Dataset & features

Experiments

PROPOSED FRAMEWORK

A regression model using heterogeneous features

Textual Features(Announcements,

Financial News, etc.)

Regression Algorithms

The dailyReturns

Financial Quantitative Features(past daily returns,

past trading value, etc.)

FORMULATION

OUTLINE

Motivations

Framework & formulation

Dataset & features

Experiments

DATASET & FEATURES

Corpus:Half year (2010) ASX announcements, 27,580 in total. 80% for developing algorithms and 20% for testing.

Features: Textual features:

o Unigram of announcements

Quantitative features:• Past daily returns (~• Stock price;• price sensitive label;• whether published in trading time;• Past trading value;• Decile by capital

OUTLINE

Motivations

Framework & formulation

Dataset & features

Experiments

EXPERIMENTS

Objectives: to find1. The best features;

Combined vs. individual features;

2. The best textual features; Unigram vs. sentiment features;

3. The best regression solver; Equal vs. unequal penalty factors on quantitative features.

OBJECTIVE 1 -- THE BEST FEATURES COMBINED VS. INDIVIDUAL FEATURES

Textual Features(Announcements,

Financial News, etc.)

Quantitative Features(Panel data)

Regression Algorithms

The Stock Returns

OBJECTIVE 1 -- THE BEST FEATURES COMBINED VS. INDIVIDUAL FEATURES

OBJECTIVE 2 -- THE BEST TEXTUAL FEATURESUNIGRAM VS. SENTIMENT FEATURES

Unigram features (all words in corpus)• Huge size of vocabulary (100,000 features)• but sparse for each document

Sentiment features (negative, positive, uncertainty)• Smaller size of features• But may loss information

Vs.

OBJECTIVE 2 -- THE BEST TEXTUAL FEATURESUNIGRAM VS. SENTIMENT FEATURES

OBJECTIVE 3 -- THE BEST REGRESSION SOLVER EQUAL VS. UNEQUAL PENALTY FACTORS

Quantitative Features (dense)

Textual Features (sparse)

Equal penalty factors

Vs.Quantitative

Features (dense)

Textual Features (sparse)

Quantitative penalty factors

Textual penalty factors

OBJECTIVE 3 -- THE BEST REGRESSION SOLVER EQUAL VS. UNEQUAL PENALTY FACTORS

Q & A

Thanks!