Sie sind auf Seite 1von 176

Quantified News based Trading:

is it the next big thing in algorithmic


trading ?
Rajib Ranjan Borah
Nov 8, 2013
Princeton UChicago Quant Trading Conference

Agenda
Background - how is news quantified
Profitability using quantitative news analysis

Machine learning techniques for designing quant news strategies


Q&A

How is news quantified Profitability

Machine learning techniques QA

Agenda
Background - how is news quantified
Profitability using quantitative news analysis

Machine learning techniques for designing quant news strategies


Q&A

How is news quantified Profitability

Machine learning techniques QA

The world runs on information and few areas as directly so as in


finance

How is news quantified Profitability

Machine learning techniques QA

Historical Perspective
1. Rothschild:

How is news quantified Profitability

Machine learning techniques QA

Historical Perspective
1. Rothschild:
A family network spread across Europe (Frankfurt, London,
Paris, Naples, Vienna)

How is news quantified Profitability

Machine learning techniques QA

Historical Perspective
1. Rothschild:
A family network spread across Europe (Frankfurt, London,
Paris, Naples, Vienna) enabled obtaining financial
information before peers

How is news quantified Profitability

Machine learning techniques QA

Historical Perspective
1. Rothschild:
A family network spread across Europe (Frankfurt, London,
Paris, Naples, Vienna) enabled obtaining financial
information before peers
Knowledge of Battle of Waterloo result one full day before
others

How is news quantified Profitability

Machine learning techniques QA

Historical Perspective
1. Rothschild:
A family network spread across Europe (Frankfurt, London,
Paris, Naples, Vienna) enabled obtaining financial
information before peers
Knowledge of Battle of Waterloo result one full day before
others largest private fortune in the world

How is news quantified Profitability

Machine learning techniques QA

Historical Perspective
1. Rothschild:
A family network spread across Europe (Frankfurt, London,
Paris, Naples, Vienna) enabled obtaining financial
information before peers
Knowledge of Battle of Waterloo result one full day before
others largest private fortune in the world
2. Reuters:

How is news quantified Profitability

Machine learning techniques QA

Historical Perspective
1. Rothschild:
A family network spread across Europe (Frankfurt, London,
Paris, Naples, Vienna) enabled obtaining financial
information before peers
Knowledge of Battle of Waterloo result one full day before
others largest private fortune in the world
2. Reuters:
News service used pigeons & telegraph in 1850s to become
fastest news disseminator

How is news quantified Profitability

Machine learning techniques QA

Historical Perspective
1. Rothschild:
A family network spread across Europe (Frankfurt, London,
Paris, Naples, Vienna) enabled obtaining financial
information before peers
Knowledge of Battle of Waterloo result one full day before
others largest private fortune in the world
2. Reuters:
News service used pigeons & telegraph in 1850s to become
fastest news disseminator
Continued focus on being the fastest news source

How is news quantified Profitability

Machine learning techniques QA

Historical Perspective
1. Rothschild:
A family network spread across Europe (Frankfurt, London,
Paris, Naples, Vienna) enabled obtaining financial
information before peers
Knowledge of Battle of Waterloo result one full day before
others largest private fortune in the world
2. Reuters:
News service used pigeons & telegraph in 1850s to become
fastest news disseminator
Continued focus on being the fastest news source $12.4
billion conglomerate

How is news quantified Profitability

Machine learning techniques QA

Historical Perspective
How have things progressed since 1850s ?

How is news quantified Profitability

Machine learning techniques QA

Historical Perspective
How have things progressed since 1850s ?
1850s : Carrier pigeons

How is news quantified Profitability

Machine learning techniques QA

Historical Perspective
How have things progressed since 1850s ?
1850s : Carrier pigeons
1860s : Telegraph

How is news quantified Profitability

Machine learning techniques QA

Historical Perspective
How have things progressed since 1850s ?
1850s : Carrier pigeons
1860s : Telegraph
1960s : Teleprinter

How is news quantified Profitability

Machine learning techniques QA

Historical Perspective
How have things progressed since 1850s ?
1850s : Carrier pigeons
1860s : Telegraph
1960s : Teleprinter
1980s : Electronic network (and internet)

How is news quantified Profitability

Machine learning techniques QA

Historical Perspective
How have things progressed since 1850s ?
1850s : Carrier pigeons
1860s : Telegraph
1960s : Teleprinter
1980s : Electronic network (and internet)
2000s : Machine Readable News a.k.a. Quantitative News

How is news quantified Profitability

Machine learning techniques QA

What is Quantitative News Trading?


News is the first order factor that affects prices, volume,
volatility of stocks, currencies, commodities, etc

How is news quantified Profitability

Machine learning techniques QA

What is Quantitative News Trading?


News is the first order factor that affects prices, volume,
volatility of stocks, currencies, commodities, etc
Computer programs that scan news articles & quantify them

How is news quantified Profitability

Machine learning techniques QA

What is Quantitative News Trading?


News is the first order factor that affects prices, volume,
volatility of stocks, currencies, commodities, etc
Computer programs that scan news articles & quantify them
-> can respond to price moving factors faster than humans

How is news quantified Profitability

Machine learning techniques QA

What is Quantitative News Trading?


News is the first order factor that affects prices, volume,
volatility of stocks, currencies, commodities, etc
Computer programs that scan news articles & quantify them
-> can respond to price moving factors faster than humans
-> can monitor a vaster amount of news reports than humans

How is news quantified Profitability

Machine learning techniques QA

What is Quantitative News Trading?


News is the first order factor that affects prices, volume,
volatility of stocks, currencies, commodities, etc
Computer programs that scan news articles & quantify them
-> can respond to price moving factors faster than humans
-> can monitor a vaster amount of news reports than humans
This field is known as Quantitative News Trading

How is news quantified Profitability

Machine learning techniques QA

What is Quantitative News Trading?


News is the first order factor that affects prices, volume,
volatility of stocks, currencies, commodities, etc
Computer programs that scan news articles & quantify them
-> can respond to price moving factors faster than humans
-> can monitor a vaster amount of news reports than humans
This field is known as Quantitative News Trading
During the 200 milliseconds a human is reading the latest news headline, a
trading program will have downloaded the entire article, analyzed its
meaning, & traded based on the content

How is news quantified Profitability

Machine learning techniques QA

What is Quantitative News Trading?


News is the first order factor that affects prices, volume,
volatility of stocks, currencies, commodities, etc
Computer programs that scan news articles & quantify them
-> can respond to price moving factors faster than humans
-> can monitor a vaster amount of news reports than humans
This field is known as Quantitative News Trading
Apart from trading, quantification of news is also utilized in
Media evaluation
Market research
Brand & reputation management
Political analysis

How is news quantified Profitability

Machine learning techniques QA

What is Quantitative News Trading?


Sample output of a News Analytics feed: News
represented by numbers

How is news quantified Profitability

Machine learning techniques QA

What is Quantitative News Trading?


News is the first order factor that affects prices, volume,
volatility of stocks, currencies, commodities, etc
Computer programs that scan news articles & quantify them
-> can respond to price moving factors faster than humans
-> can monitor a vaster amount of news reports than humans
This field is known as Quantitative News Trading
Apart from trading, quantification of news is also utilized in
Media evaluation
Market research
Brand & reputation management
Political analysis

How is news quantified Profitability

Machine learning techniques QA

What is Quantitative News Trading?


News is the first order factor that affects prices, volume,
volatility of stocks, currencies, commodities, etc
Computer programs that scan news articles & quantify them
-> can respond to price moving factors faster than humans
-> can monitor a vaster amount of news reports than humans
This field is known as Quantitative News Trading
During the 200 milliseconds a human is reading the latest news headline, a
trading program will have downloaded the entire article, analyzed its
meaning, & traded based on the content

How is news quantified Profitability

Machine learning techniques QA

What is Quantitative News Trading?


News is the first order factor that affects prices, volume,
volatility of stocks, currencies, commodities, etc
Computer programs that scan news articles & quantify them
-> can respond to price moving factors faster than humans
-> can monitor a vaster amount of news reports than humans
This field is known as Quantitative News Trading
During the 200 milliseconds a human is reading the latest news headline, a
trading program will have downloaded the entire article, analyzed its
meaning, & traded based on the content

How is news quantified Profitability

Machine learning techniques QA

What is Quantitative News Trading?


News is the first order factor that affects prices, volume,
volatility of stocks, currencies, commodities, etc
Computer programs that scan news articles & quantify them
-> can respond to price moving factors faster than humans
-> can monitor a vaster amount of news reports than humans

How do you quantify news reports and articles ?


This field is known as Quantitative News Trading
During the 200 milliseconds a human is reading the latest news headline, a
trading program will have downloaded the entire article, analyzed its
meaning, & traded based on the content

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 1. Sentiment


News articles are assigned a score called sentiment

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 1. Sentiment


News articles are assigned a score called sentiment
Sentiment says whether the article has a positive / negative or
neutral tone

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 1. Sentiment


News articles are assigned a score called sentiment
Sentiment says whether the article has a positive / negative or
neutral tone
(Sale of Apple iPhones drop = -ve sentiment)

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 1. Sentiment


News articles are assigned a score called sentiment
Sentiment says whether the article has a positive / negative or
neutral tone
(Sale of Apple iPhones drop = -ve sentiment)
Sentiment at document level is different from sentiment at
entity level

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 1. Sentiment


News articles are assigned a score called sentiment
Sentiment says whether the article has a positive / negative or
neutral tone
(Sale of Apple iPhones drop = -ve sentiment)
Sentiment at document level is different from sentiment at
entity level
(Samsung beats Apple in smart phone sales = -ve sentiment for
entity named Apple, +ve sentiment for Samsung)

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 1. Sentiment


How is sentiment scored ?

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 1. Sentiment


How is sentiment scored ?
Naive parser: based on word count of ve / +ve keywords

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 1. Sentiment


How is sentiment scored ?
Naive parser: based on word count of ve / +ve keywords
Company Xs sales were good
Company Xs sales were fantastic
(Both have one positive keyword, but the intensity differs)

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 1. Sentiment


How is sentiment scored ?
Naive parser: based on word count of ve / +ve keywords
Discriminated parser: weighted word count
The results were good, not bad.
The results were bad, not good.
(Both score equally. Both have the same words - but mean completely
opposite)

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 1. Sentiment


How is sentiment scored ?
Naive parser: based on word count of ve / +ve keywords
Discriminated parser: weighted word count
Grammatical parser: which verbs work on which objects.
check linguistic semantics

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 1. Sentiment


How is sentiment scored ?
Naive parser: based on word count of ve / +ve keywords
Discriminated parser: weighted word count
Grammatical parser: which verbs work on which objects.
check linguistic semantics
Machine Learning: From the data and the answers, try to find
the factors

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 1. Sentiment


How is sentiment scored ?
Naive parser: based on word count of ve / +ve keywords
Discriminated parser: weighted word count
Grammatical parser: which verbs work on which objects.
check linguistic semantics
Machine Learning: From the data and the answers, try to find
the factors
Generate bag-of-words: distance of subject from these sentiment
words
Overfitting (and large vector sets), hitch-hiking and ignorance of
linguistic structure

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 1. Sentiment


Scoring sentiments: grammatical parsing

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 1. Sentiment


Scoring sentiments: grammatical parsing
A database of words & phrases against which the article is
searched

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 1. Sentiment


Scoring sentiments: grammatical parsing
A database of words & phrases against which the article is
searched
Which verbs act on which objects

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 1. Sentiment


Scoring sentiments: grammatical parsing
A database of words & phrases against which the article is
searched
Which verbs act on which objects
Phrases which use adjectives & adverbs emphasize
sentiments, therefore greater weightage

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 1. Sentiment


Scoring sentiments: grammatical parsing issues

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 1. Sentiment


Scoring sentiments: grammatical parsing issues
Linguistic structures like negation, double negation, sarcasm,
intensification, hanging lemma

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 1. Sentiment


Scoring sentiments: grammatical parsing issues
Linguistic structures like negation, double negation, sarcasm,
intensification, hanging lemma
(negation: Company X did not become the best in the world
double negation: Company X did not do bad
sarcasm: With such an attitude, X is sure to become the best firm
intensification: Company X did terribly well
hanging lemma: Company X loses lawsuit against company Y. They will
have to pay $1billion USD )

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 1. Sentiment


Scoring sentiments: grammatical parsing issues
Linguistic structures like negation, double negation, sarcasm,
intensification, hanging lemma
(negation: Company X did not become the best in the world
double negation: Company X did not do bad
sarcasm: With such an attitude, X is sure to become the best firm
intensification: Company X did terribly well
hanging lemma: Company X loses lawsuit against company Y. They will
have to pay $1billion USD )

Word Sense Disambiguation - same word, different meanings

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 1. Sentiment


Scoring sentiments: grammatical parsing issues
Linguistic structures like negation, double negation, sarcasm,
intensification, hanging lemma
(negation: Company X did not become the best in the world
double negation: Company X did not do bad
sarcasm: With such an attitude, X is sure to become the best firm
intensification: Company X did terribly well
hanging lemma: Company X loses lawsuit against company Y. They will
have to pay $1billion USD )

Word Sense Disambiguation - same word, different meanings


Company X received a fine
X is doing fine
X sells fine grained sand, etc

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 2. Relevance


Is Sentiment good enough to quantify a news report?

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 2. Relevance


Is Sentiment good enough to quantify a news report?
A news article might:
be predominantly about a company

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 2. Relevance


Is Sentiment good enough to quantify a news report?
A news article might:
be predominantly about a company
mention that company and others as well

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 2. Relevance


Is Sentiment good enough to quantify a news report?
A news article might:
be predominantly about a company
mention that company and others as well
mention that company in passing in the article

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 2. Relevance


Is Sentiment good enough to quantify a news report?
A news article might:
be predominantly about a company
mention that company and others as well
mention that company in passing in the article
Relevance measures how relevant a news article is for a
particular company

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 2. Relevance


How is relevance scored ?

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 2. Relevance


How is relevance scored ?
How many companies are mentioned in the news article

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 2. Relevance


How is relevance scored ?
How many companies are mentioned in the news article
Is the company mentioned in the headline as the
subject/object
(Headline:UBS downgrades HSBC is not relevant to UBS)

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 2. Relevance


How is relevance scored ?
How many companies are mentioned in the news article
Is the company mentioned in the headline as the
subject/object
(Headline:UBS downgrades HSBC is not relevant to UBS)

In which sentence number is the company first mentioned

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 2. Relevance


How is relevance scored ?
How many companies are mentioned in the news article
Is the company mentioned in the headline as the
subject/object
(Headline:UBS downgrades HSBC is not relevant to UBS)

In which sentence number is the company first mentioned


Length of the article & how many times is the firm mentioned

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 2. Relevance


How is relevance scored ?
How many companies are mentioned in the news article
Is the company mentioned in the headline as the
subject/object
(Headline:UBS downgrades HSBC is not relevant to UBS)

In which sentence number is the company first mentioned


Length of the article & how many times is the firm mentioned
Number of sentiment words & total words in article

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 2. Relevance


How is relevance scored ?
How many companies are mentioned in the news article
Is the company mentioned in the headline as the
subject/object
(Headline:UBS downgrades HSBC is not relevant to UBS)

In which sentence number is the company first mentioned


Length of the article & how many times is the firm mentioned
Number of sentiment words & total words in article
Two firms mentioned in a news article can both have a
relevance of 1.0 (HP & Compaq announce merger)

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 2. Relevance


Issues with calculating relevance

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 2. Relevance


Issues with calculating relevance
Requires synonym database:

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 2. Relevance


Issues with calculating relevance
Requires synonym database:

IBM
International Business Machines
I.B.M.
Big Blue

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 2. Relevance


Issues with calculating relevance
Requires synonym database:

IBM
International Business Machines
I.B.M.
Big Blue

BAML
Bank of America
Merrill Lynch
Bank of America Merrill Lynch
Merrill
BoA

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 3. Novelty


Often the news article is not reported in its entirety, but in
multiple spurts

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 3. Novelty


Often the news article is not reported in its entirety, but in
multiple spurts

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 3. Novelty


Often the news article is not reported in its entirety, but in
multiple spurts

Alert
News Article
Update
Append

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 3. Novelty


Often the news article is not reported in its entirety, but in
multiple spurts

Alert
News Article
Update
Append

Moreover, multiple news


sources report same news

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 3. Novelty


Often the news article is not reported in its entirety, but in
multiple spurts

Alert
News Article
Update
Append

Moreover, multiple news


sources report same news
News also cause price
changes which themselves
become news

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 3. Novelty


If we do not keep track & respond to repeated instances of
the same news

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 3. Novelty


If we do not keep track & respond to repeated instances of
the same news => we will end up repeating our actions
manifold for the same event

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 3. Novelty


If we do not keep track & respond to repeated instances of
the same news => we will end up repeating our actions
manifold for the same event
Therefore every news article should be checked for newness
or novelty before responding

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 3. Novelty


How is novelty measured ?

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 3. Novelty


How is novelty measured ?
The keywords in the current news article are compared to
historical articles about that company for similarity of digital
fingerprints

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 3. Novelty


How is novelty measured ?
The keywords in the current news article are compared to
historical articles about that company for similarity of digital
fingerprints
A linked articles count is generated

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 3. Novelty


How is novelty measured ?
The keywords in the current news article are compared to
historical articles about that company for similarity of digital
fingerprints
A linked articles count is generated
Novelty is reported for
Within same news feed novelty (i.e. all Bloomberg news articles only)

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 3. Novelty


How is novelty measured ?
The keywords in the current news article are compared to
historical articles about that company for similarity of digital
fingerprints
A linked articles count is generated
Novelty is reported for
Within same news feed novelty (i.e. all Bloomberg news articles only)
Across all news feeds novelty (i.e. across Reuters, Dow Jones,
Bloomberg articles)

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 4. Market Impact


Different types of news articles have different impacts on the
price of the asset

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 4. Market Impact


Different types of news articles have different impacts on the
price of the asset
Another aspect of relevance is the likely market impact of the
news article

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 4. Market Impact


Different types of news articles have different impacts on the
price of the asset
Another aspect of relevance is the likely market impact of the
news article
Market Impact is therefore a function of the type of news

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - News Types


Types of news:
Accounting news

Earnings
Trading updates (broker action, market commentary)
Guidance
Financial issues (buybacks, dividends, equity offerings, etc)
Regulatory filings

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - News Types


Types of news:
Accounting news

Earnings
Trading updates (broker action, market commentary)
Guidance
Financial issues (buybacks, dividends, equity offerings, etc)
Regulatory filings

Strategic news

M&A
Restructuring
Product, customer, competition related
Corporate Governance

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - News Types


Types of news based on time of news report
Asynchronous / unexpected
Synchronous / fixed releases

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - Key Factors


While the following are the four key inputs:

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - Key Factors


While the following are the four key inputs:

Sentiment
Relevance
Novelty
Market Impact

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - Key Factors


While the following are the four key inputs:

Sentiment
Relevance
Novelty
Market Impact

Some news analytics based strategies use other factors as well

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 5. Volume


The number of news articles on the same topic can be a useful
input to validate the impact

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 5. Volume


The number of news articles on the same topic can be a useful
input to validate the impact
Volume of news in Social Media also checked sometimes

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 5. Volume


The number of news articles on the same topic can be a useful
input to validate the impact
Volume of news in Social Media also checked sometimes
News Analytics strategies also check market based qualitative
parameters along with news -> these help check if reaction to
news is not already factored in
Trading Volume in last 24 hours (and historical average volume)
Price change in last 24 hours

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - Key Factors


While the following are the four key inputs:

Sentiment
Relevance
Novelty
Market Impact

Some news analytics based strategies use other factors as well


Volume

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 6. Social Media


Long term trading strategies try to gauge market sentiment from
the plethora of information in the social media front

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 6. Social Media


Long term trading strategies try to gauge market sentiment from
the plethora of information in the social media front
Search engine volume counts (e.g. Google Trends) - global
search for news keywords.
Can be used to confirm market impact of news

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 6. Social Media


Long term trading strategies try to gauge market sentiment from
the plethora of information in the social media front
Search engine volume counts (e.g. Google Trends) - global
search for news keywords.
Can be used to confirm market impact of news
Facebook, Twitter - user sentiment evaluated at macro level.

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - 6. Social Media


Long term trading strategies try to gauge market sentiment from
the plethora of information in the social media front
Search engine volume counts (e.g. Google Trends) - global
search for news keywords.
Can be used to confirm market impact of news
Facebook, Twitter - user sentiment evaluated at macro level.
Many tools use certified twitter/facebook feeds only

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - Key Factors


While the following are the four key inputs:

Sentiment
Relevance
Novelty
Market Impact

Some news analytics based strategies use other factors as well


Volume
Social Media

How is news quantified Profitability

Machine learning techniques QA

Quantifying News Market Psyche


News Analytics tools calculate Market Psychology Indices evaluating broad psychological sentiments from global news

How is news quantified Profitability

Machine learning techniques QA

Quantifying News Market Psyche


News Analytics tools calculate Market Psychology Indices evaluating broad psychological sentiments from global news
Country : sentiment, conflict, fear, joy, optimism, trust,
uncertainty, urgency, violence, government corruption,
government instability, social unrest, default, inflation, credit
tightening, etc

How is news quantified Profitability

Machine learning techniques QA

Quantifying News Market Psyche


News Analytics tools calculate Market Psychology Indices evaluating broad psychological sentiments from global news
Country : sentiment, conflict, fear, joy, optimism, trust,
uncertainty, urgency, violence, government corruption,
government instability, social unrest, default, inflation, credit
tightening, etc
Equity: Gloom, Anger, Innovation, Stress, Optimism, Earnings
Expectations, Market Risk, Market Forecast

How is news quantified Profitability

Machine learning techniques QA

Quantifying News Market Psyche


News Analytics tools calculate Market Psychology Indices evaluating broad psychological sentiments from global news
Country : sentiment, conflict, fear, joy, optimism, trust,
uncertainty, urgency, violence, government corruption,
government instability, social unrest, default, inflation, credit
tightening, etc
Equity: Gloom, Anger, Innovation, Stress, Optimism, Earnings
Expectations, Market Risk, Market Forecast
Currency: Forecast, Currency Peg Instability, Carry Trade

How is news quantified Profitability

Machine learning techniques QA

Quantifying News Market Psyche


News Analytics tools calculate Market Psychology Indices evaluating broad psychological sentiments from global news
Country : sentiment, conflict, fear, joy, optimism, trust,
uncertainty, urgency, violence, government corruption,
government instability, social unrest, default, inflation, credit
tightening, etc
Equity: Gloom, Anger, Innovation, Stress, Optimism, Earnings
Expectations, Market Risk, Market Forecast
Currency: Forecast, Currency Peg Instability, Carry Trade
Agriculture: Acreage cultivated, weather damage, subsidies,
production volume, supply vs demand, surplus vs shortage,
price up

How is news quantified Profitability

Machine learning techniques QA

Quantifying News Market Psyche

How is news quantified Profitability

Machine learning techniques QA

Quantifying News Market Psyche

How is news quantified Profitability

Machine learning techniques QA

Agenda
Background - how is news quantified
Profitability using quantitative news analysis

Machine learning techniques for designing quant news strategies


Q&A

How is news quantified Profitability

Machine learning techniques QA

Is it profitable ?
Are computers smart enough to read news and make profitable
trades?

How is news quantified Profitability

Machine learning techniques QA

Is it profitable ?
Are computers smart enough to read news and make profitable
trades?

How is news quantified Profitability

Machine learning techniques QA

Is it profitable ?
Are computers smart enough to read news and make profitable
trades?

News sentiments reversed


trend first -> prices
reversed trend next

How is news quantified Profitability

Machine learning techniques QA

Where Quantified news work


Machines are faster at responding to events than humans

How is news quantified Profitability

Machine learning techniques QA

Where Quantified news work


Machines are faster at responding to events than humans

Machines can process a much vaster amount of information


without any fatigue

How is news quantified Profitability

Machine learning techniques QA

Where Quantified news work


Machines are faster at responding to events than humans
Low latency event based trading (first to respond)

Machines can process a much vaster amount of information


without any fatigue

How is news quantified Profitability

Machine learning techniques QA

Where Quantified news work


Machines are faster at responding to events than humans
Low latency event based trading (first to respond)

Machines can process a much vaster amount of information


without any fatigue
Analyze broad spectrum of news to formulate broad views

How is news quantified Profitability

Machine learning techniques QA

Where Quantified news work


Machines are faster at responding to events than humans
Low latency event based trading (first to respond)

Machines can process a much vaster amount of information


without any fatigue
Analyze broad spectrum of news to formulate broad views

How is news quantified Profitability

Machine learning techniques QA

Where Quantified news work

Analyze broad spectrum of news to formulate broad views

How is news quantified Profitability

Machine learning techniques QA

Where Quantified news work

Analyze broad spectrum of news to formulate broad views

How is news quantified Profitability

Machine learning techniques QA

Where Quantified news work

Analyze broad spectrum of news to formulate broad views

How is news quantified Profitability

Machine learning techniques QA

Where Quantified news work

Low latency event based trading (first to respond)

How is news quantified Profitability

Machine learning techniques QA

Where Quantified news work

Low latency event based trading (first to respond)


For synchronous (fixed releases) expected events (earnings
releases/ economic figures)
Company figures provided in xml format instead of text

How is news quantified Profitability

Machine learning techniques QA

Where Quantified news work

Low latency event based trading (first to respond)


For synchronous (fixed releases) expected events (earnings
releases/ economic figures)
Company figures provided in xml format instead of text
Economic figures provided in binary format instead of textual
news articles

How is news quantified Profitability

Machine learning techniques QA

Where Quantified news work

Low latency event based trading (first to respond)


For synchronous (fixed releases) expected events (earnings
releases/ economic figures)
Company figures provided in xml format instead of text
Economic figures provided in binary format instead of textual
news articles
For asynchronous / unexpected news
Are quantification algorithms robust enough to calculate
trust-worthy sentiment, relevance, novelty scores ?

How is news quantified Profitability

Machine learning techniques QA

Opportunities : initial under-reaction


Quantified news driven trades work even when the trade is done
at the end of the day
(under-reaction to news immediately. Tetlock, et al)

How is news quantified Profitability

Machine learning techniques QA

Late endof day response also profitable


Trading the news immediately = very profitable
At a broad level there is underreaction to news => entering into
trades at the end of the day also makes profits

How is news quantified Profitability

Machine learning techniques QA

Long short strategy returns

How is news quantified Profitability

Machine learning techniques QA

Filtering sentiments increase profits

Increasing threshold from 90 to


95 percentile increases returns
from 55 to 138 bps in 3 days

How is news quantified Profitability

Machine learning techniques QA

Certain sectors more profitable

Moving from Non-Cyclicals to


Financials increased the profit
from 135BP to 147BP

How is news quantified Profitability

Machine learning techniques QA

Sensitivity of different sectors

Sectors like Pharma, Defense, Auto, Energy, Banking more sensitive to news

How is news quantified Profitability

Machine learning techniques QA

Small cap firms more profitable


Smaller Cap firms show greater response to extreme sentiment
news event
(bigger firms have greater scrutiny)

How is news quantified Profitability

Machine learning techniques QA

Filter & trade fewer stocks


More is not better. Quality over quantity
Trading only stocks with very high sentiment/relevance is
better

How is news quantified Profitability

Machine learning techniques QA

Hedged (market-neutral) is better


Long +ve sentiment stocks only
OR
Short -ve sentiment stocks only. Will fail in different regimes

Being long +ve sentiment stocks & short -ve sentiment stocks
at the same time gives consistent returns

How is news quantified Profitability

Machine learning techniques QA

Volatility regimes and news


Volatile vs stable Economic regimes

In more volatile markets people tend to react less strongly to


positive news and react more strongly to negative news

How is news quantified Profitability

Machine learning techniques QA

Surprises are more profitable


Bigger moves happen when there is news in
Stocks with low beta (i.e. surprises happen to sleepy stocks)

How is news quantified Profitability

Machine learning techniques QA

Surprises are more profitable


Bigger moves happen when there is news in
Stocks with low beta (i.e. surprises happen to sleepy stocks)

How is news quantified Profitability

Machine learning techniques QA

Surprises are more profitable


Bigger moves happen when there is news in
Stocks with low beta (i.e. surprises happen to sleepy stocks)
VIX is low (i.e. surprises during calm times)

How is news quantified Profitability

Machine learning techniques QA

Surprises are more profitable


Bigger moves happen when there is news in
Stocks with low beta (i.e. surprises happen to sleepy stocks)
VIX is low (i.e. surprises during calm times)

How is news quantified Profitability

Machine learning techniques QA

Surprises are more profitable


Bigger moves happen when there is news in
Stocks with low beta (i.e. surprises happen to sleepy stocks)
VIX is low (i.e. surprises during calm times)
When markets are improving (i.e. surprise to mostly long
position holders)

How is news quantified Profitability

Machine learning techniques QA

Surprises are more profitable


Bigger moves happen when there is news in
Stocks with low beta (i.e. surprises happen to sleepy stocks)
VIX is low (i.e. surprises during calm times)
When markets are improving (i.e. surprise to mostly long
position holders)

How is news quantified Profitability

Machine learning techniques QA

Surprises are more profitable


Bigger moves happen when there is news in
Stocks with low beta (i.e. surprises happen to sleepy stocks)
VIX is low (i.e. surprises during calm times)
When markets are improving (i.e. surprise to mostly long
position holders)

How is news quantified Profitability

Machine learning techniques QA

Strategy variation - sentiment changes


Instead of absolute sentiment scores, look at changes in
sentiment scores of firms
Bought stocks with highest increase in sentiment
Shorted stocks with highest decrease in sentiment

How is news quantified Profitability

Machine learning techniques QA

Strategy variation - bottom fishing


Bottom - fishing / turnaround stories
Buying stocks with reversal in sentiment from grossly
negative (a lot of the stocks turned out to be buybacks)

How is news quantified Profitability

Machine learning techniques QA

Strategy variation - trading volatility


News articles definitely lead to increased volatility, even
though direction of move might be difficult to predict
through news analytics
Take vega positions (var-swaps) using options in anticipation
of increased volatility

How is news quantified Profitability

Machine learning techniques QA

Generating Alpha
Soft (opinion based) vs. Hard (fact based) news
Hard news has a stronger short term reaction than soft news

Source: RavenPack, FactSet, Macquarie Research, September 2012

How is news quantified Profitability

Machine learning techniques QA

Generating Alpha
Scheduled/expected vs. Unscheduled/unexpected
Investors react more strongly to unscheduled/ unexpected
news than scheduled/ expected

Source: RavenPack, FactSet, Macquarie Research, September 2012

How is news quantified Profitability

Machine learning techniques QA

Generating Alpha
Forecast vs Actual earnings
Investors react more strongly to forecasts than actual earnings
news

Source: RavenPack, FactSet, Macquarie Research, September 2012

How is news quantified Profitability

Machine learning techniques QA

Generating Alpha
Guidance vs Actual earnings
Investors react more strongly to guidance to actual earnings

Source: RavenPack, FactSet, Macquarie Research, September 2012

How is news quantified Profitability

Machine learning techniques QA

Generating Alpha
News type Event Study Results

Source: RavenPack, FactSet, Macquarie Research, September 2012

How is news quantified Profitability

Machine learning techniques QA

To summarize
News Analytics works best with

Small cap stocks


Sectors like pharma, banking, etc
Stocks with low beta
When VIX is low
When markets are improving
Hard news (vis-a-vis Soft news)
Unscheduled news events (vis-a-vis scheduled news events)
Being market-neutral
Doing fewer stocks, but those with stronger signals

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - Where it fails ?


On Sep. 7, 2008 Googles newsbots picked up an old 2002
story about United Airlines possibly filing for bankruptcy

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - Where it fails ?


On Sep. 7, 2008 Googles newsbots picked up an old 2002
story about United Airlines possibly filing for bankruptcy

UAL stock dived immediately

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - Where it fails?


News analytics were taught that Osama-Bin-Laden, and
killed had -ve sentiments for the markets

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - Where it fails?


News analytics were taught that Osama-Bin-Laden, and
killed had -ve sentiments for the markets

On May 2 2012 when news reporting Osama Bin-Landen


killed were published, news bots treated this as a negative
news article and sold stocks

How is news quantified Profitability

Machine learning techniques QA

Quantifying News - Where it fails?


News analytics were taught that Osama-Bin-Laden, and
killed had -ve sentiments for the markets

On May 2 2012 when news reporting Osama Bin-Landen


killed were published, news bots treated this as a negative
news article and sold stocks

The two examples cited and their impacts show the extent to
which people have embraced news analytics to automate
trading

How is news quantified Profitability

Machine learning techniques QA

Quantifying News challenges


Languages like Chinese and Japanese with large number of
alphabetic symbols and complex grammar

How is news quantified Profitability

Machine learning techniques QA

Quantifying News challenges


Languages like Chinese and Japanese with large number of
alphabetic symbols and complex grammar
However, there is a lot of development in this domain already

How is news quantified Profitability

Machine learning techniques QA

Quantifying News challenges


Languages like Chinese and Japanese with large number of
alphabetic symbols and complex grammar
However, there is a lot of development in this domain already
The ever increasing volume of news articles from increased
news sources, and from increased volumes in social media

How is news quantified Profitability

Machine learning techniques QA

Agenda
Background - how is news quantified
Profitability using quantitative news analysis

Machine learning techniques for designing quant news strategies


Q&A

How is news quantified Profitability

Machine learning techniques QA

Machine Learning methodologies


Traditional approach => formulate hypothesis based on
experience/expertise, validate statistically using historical data

How is news quantified Profitability

Machine learning techniques QA

Machine Learning methodologies


Traditional approach => formulate hypothesis based on
experience/expertise, validate statistically using historical data
Machine learning approach =>

How is news quantified Profitability

Machine learning techniques QA

Machine Learning methodologies


Traditional approach => formulate hypothesis based on
experience/expertise, validate statistically using historical data
Machine learning approach => output + raw data fed into a
system. System reports factors within data that lead to output

How is news quantified Profitability

Machine learning techniques QA

Machine Learning methodologies


Traditional approach => formulate hypothesis based on
experience/expertise, validate statistically using historical data
Machine learning approach => output + raw data fed into a
system. System reports factors within data that lead to output
Three broad approaches
Tree
Forest
Planet

How is news quantified Profitability

Machine learning techniques QA

Machine Learning - TREE method


Output: Post-event abnormal results
Input: Quantitative news analytics

How is news quantified Profitability

Machine learning techniques QA

Machine Learning - TREE method


Output: Post-event abnormal results
Input: Quantitative news analytics

How is news quantified Profitability

Machine learning techniques QA

Machine Learning - TREE method


Output: Post-event abnormal results
Input: Quantitative news analytics
Issues: Overfitting
(works with training data
does not work on real data)

How is news quantified Profitability

Machine learning techniques QA

Machine Learning - TREE method


Output: Post-event abnormal results
Input: Quantitative news analytics
Issues: Overfitting
(works with training data
does not work on real data)
Solution: Pruning

How is news quantified Profitability

Machine learning techniques QA

Machine Learning - TREE method


Output: Post-event abnormal results
Input: Quantitative news analytics
Issues: Overfitting
(works with training data
does not work on real data)
Solution: Pruning

How is news quantified Profitability

Machine learning techniques QA

Machine Learning - FOREST method


Multiple factors might impact output

How is news quantified Profitability

Machine learning techniques QA

Machine Learning - FOREST method


Multiple factors might impact output
Instead of one tree to solve everything,
have a forest of trees

How is news quantified Profitability

Machine learning techniques QA

Machine Learning - FOREST method


Multiple factors might impact output
Instead of one tree to solve everything,
have a forest of trees

How is news quantified Profitability

Machine learning techniques QA

Machine Learning - FOREST method


Multiple factors might impact output
Instead of one tree to solve everything,
have a forest of trees
Each tree has a vote in the output.
Weightage of vote depends on accuracy
of that tree

How is news quantified Profitability

Machine learning techniques QA

Machine Learning - PLANET method


Instead of linear relationships between input and output,
Planet breaks the variable space into sections, fits linear
functions within those sections

How is news quantified Profitability

Machine learning techniques QA

Machine Learning - PLANET method


Instead of linear relationships between input and output,
Planet breaks the variable space into sections, fits linear
functions within those sections

How is news quantified Profitability

Machine learning techniques QA

Agenda
Background - how is news quantified
Profitability using quantitative news analysis

Machine learning techniques for designing quant news strategies


Q&A

How is news quantified Profitability

Machine learning techniques QA

Contacts
For 4-month Executive Program in Algorithmic Trading:
contact@quantinsti.com
E-PAT: 4 month weekend online program (3hrs every Sat + Sun)
Statistics
Quant Strategies
Technology (coding on algorithmic trading platform)
For algorithmic trading advisory: contact@iragecapital.com
To reach me directly: rajib.borah@iragecapital.com

How is news quantified Profitability

Machine learning techniques QA

Text Mining: An example


Converting a line
of news into
metadata to be
used for analysis
or trade

How is news quantified Profitability

Machine learning techniques QA

Calculating Novelty
Novelty: It indicates Novelty by comparing current
news with cache of previous news
Paul C. Tetlock defined Staleness measure as follows
Define the unique single word and two-word set for ith stock as

for past 10 news staleness:

Das könnte Ihnen auch gefallen