TABLE OF CONTENTS

Enhancing Equity
Trading Models with
Corporate &
Macro Sentiment
Abnormality

Executive Summary

Introduction

Data Description

Indicator Methodology

Evaluating MSI

Conclusion

12

Appendix

13

By:
Peter Hafez
Director of Quantitative Research
RavenPack

October 2014

Contact the authors

This White Paper is not intended for trading purposes. The White Paper is not appropriate for the purposes of making a decision to carry out a transaction or trade. Nor does it provide any form of
advice (investment, tax, legal) amounting to investment advice, or make any recommendations regarding particular financial instruments, investments or products. RavenPack may discontinue or
change the White Paper content at any time, without notice. RavenPack does not guarantee or warrant the accuracy, completeness or timeliness of the White Paper. For more detailed disclaimer,
please refer to the back cover of this document.

www.ravenpack.com
th
New York 535 Fifth Ave., 4 Floor, New York, NY 10017 | TEL: +1 (646) 277-7339 | research@ravenpack.com
RavenPack Quantitative Research 2014 All Rights Reserved. No duplication or redistribution of this document without written consent

RavenPack Quantitative Research

Executive Summary
The aim of this paper is to construct a set of indicators or factors that will help a portfolio
manager outperform both a passive buy-and-hold strategy and a price momentum strategy on
the S&P 500. Specifically, we create a set of macro indicators that capture sentiment
abnormality across six topics, namely business, economic, political, environmental, societal and
corporate. Overall, we find that employing our macro sentiment indicators significantly
improves the performance of traditional long-only or long/short models. We also find that
company-specific or corporate sentiment is the key performance driver over a 1-month horizon.

Both our long-only and long/short news-based strategies deliver Information Ratios (IRs)
of 1.45 since September 2007. In comparison, a buy-and-hold S&P 500 strategy and a
pure return-driven model deliver IRs of only 0.23 and 0.73, respectively.

In addition, employing these news sentiment indicators is particularly useful for

managing downside risk as our sentiment models reduce the maximum drawdown by
more than 70% on the S&P 500 compared to a buy & hold strategy.

200%

Cumulative Return

150%
100%
50%
0%
-50%
-100%

S&P500 (Buy&Hold)

Return

Corporate News

Corporate&Macro News + Return

SOURCE: RavenPack, October 2014

About RavenPack Data
RavenPack News Analytics (RPNA) provides real-time structured sentiment, relevance and novelty data for entities
and events detected in unstructured text published by reputable sources. Publishers include Dow Jones
Newswires, Barrons, the Wall Street Journal and over 19,000 other traditional and social media sites. Over 14
years of Dow Jones newswires archive and 7 years of historical data from web publications and blogs are available
for backtesting. RavenPack detects news and produces analytics data on over 34,000 listed stocks from the world's
equity markets, over 2,500 financially relevant organizations, 138,000 places, 150 currencies and 80 commodities.

Page|2

RavenPack Quantitative Research

1. Introduction
One of the challenges for news analytics is to improve the performance of traditional
investment strategies. Hence, the aim of this paper is to construct a set of indicators that will
help a portfolio manager outperform both a passive buy-and-hold strategy and a price
momentum strategy on the S&P 500.
We will call this family of factors Macro Sentiment Indicators (MSIs) and they are designed to
capture the daily news sentiment of an economy across different topics, namely Corporate,
Business, Economy, Politics, Environment and Society.1 We also look at more granular groups
within these topics. For example, in Business we look at credit ratings, in Economy with look at
employment, in Environment we look at natural disasters, in Politics we look at foreign
relations, or in Society we analyze civil unrest.
In this paper we construct these indicators for the United States and apply them to a monthly
directional prediction model on the S&P 500 index. This thematic approach allows us to isolate
from the noise in the large volumes of news out there, and provides insight as to which
themes are dominant at any given time.
Overall, we find strong evidence that tracking news sentiment at a topical level can help predict
forward returns of the S&P 500. We find news sentiment especially useful for managing
downside risk as our sentiment model reduces the maximum drawdown by more than 70% as
compared to a simple buy & hold strategy. Our monthly long-only or long/short strategies have
delivered Information Ratios of nearly 1.5 since September 2007 vs. 0.6-0.7 for our returndriven benchmarks that follow the same model construction approach. In terms of annualized
returns, the long-only strategy has yielded 13.5% vs. 23.0% for the long/short strategy. While
corporate sentiment is found to be the key return driver of market direction, including macro
sentiment in our model further improves strategy performance.
In the following section, we provide a brief overview of the data used in this paper. In Section 3,
we present our sentiment indicator methodology and Section 4 evaluates our indicators in a
dynamic regression framework with the aim of predicting the direction of the US equity market.
Finally, in Section 5 we present our conclusions.

These topics are the high-level groups in the RavenPack event taxonomy.

Page|3

RavenPack Quantitative Research

2. Data Description
This research is based on the Full Edition of RavenPack News Analytics Version 4.0, which
combines the equities and global macro analytics from the Dow Jones, Web, and Press Release
Editions. The Dow Jones Edition analyzes relevant information from Dow Jones Newswires,
regional editions of the Wall Street Journal, Marketwatch and Barron's. The Web Edition
automatically processes content from more than 19,000 sources including industry and
business publishers, national and local news, blogs, regulatory filings, and government updates.
The PR Edition analyzes the news and information from 22 regulatory news feeds and press
release distribution networks, including exclusive content from PRNewswire, Canadian News
Wire, LSE Regulatory News Service, and others.
The MSIs are based on RavenPacks Event Sentiment Score (ESS), which is a granular score
between 0 and 100 that represents the entity-specific sentiment for one of the events defined
in the RavenPack event taxonomy. The score is determined by systematically matching stories
typically categorized by financial experts as having short-term positive or negative financial or
economic impact.
We only consider entities receiving a RavenPack Relevance Score of 100. The Relevance Score
measures how important an entity is within any given news article and can range from 0 to 100,
where 100 is most relevant.
We also weight each ESS score included in the MSI with RavenPacks Global Event Novelty Score
(G_ENS). G_ENS is a measure between 0 and 100 that represents how "new" or novel a news
story is within a 24-hour time window across all news providers covered by RavenPack2. Hence
breaking news gets most weight and subsequent news coverage less.
Finally, as previously mentioned, we utilize RavenPacks Event Taxonomy to isolate the different
potential drivers of overall market sentiment. The taxonomy identifies four levels of
classification, including event categories, types, groups, and topics, see Figure 1. It offers a
balance between levels of information aggregation, for those who look at news from the top
down, and granularity, for those who look at news from the bottom up. For this particular
study, we divide the business topic into two to create a sixth topic focusing purely on
company-specific events. We label this topic Corporate.

For more details on Relevance, ESS, or G_ENS data please refer to Appendix A.

Page|4

RavenPack Quantitative Research

Fig 1: The Structure of RavenPacks Event Taxonomy

Topics (5)
Groups (51)
Types (412)
Categories (2,064)
This figure presents the structure of the RavenPack Event Taxonomy.
SOURCE: RavenPack, October 2014

3. Indicator Methodology
As previously mentioned, the aim of our paper is to create a family of Macro Sentiment
Indicators (MSIs) that capture the sentiment of an economy at the event topic and group levels.
At the highest level, we measure sentiment across the 6 topics: Corporate3, Business, Economy,
Politics, Environment and Society. The indicators are constructed as follows:
Step 1: Create a daily sentiment score at the event group level by taking the cumulative sum of
all event sentiment scores on a given day for an economy (e.g. the United States). ESS is
rescaled to take values between -1 and 1, and is weighted according to the novelty score of the
event.

Where
is the event sentiment score for news event
with representing the set of
news events within a given event group on day .
represents the global event novelty
score for news event . We exclude events that dont contain sentiment (i.e.
) or that
match the entity type CMDT or CURR, since these events do not relate to any particular
country, but rather relate to commodities or currencies. In addition, we remove events with
entity type COMP from the macro package as we only want to model corporate news based
on the equities package.

Note that the corporate indicator is calculated from company-specific events spanning across all topics in the
RPNA Equities package.

Page|5

RavenPack Quantitative Research

Step 2: The daily scores,

, are then normalized by country and per event group using a Z-score

approach. This allows us to measure sentiment abnormality within a given event group and
thereby identify key drivers of current sentiment:

Where and

are the 365 day average and standard deviation of

Step 3: Create a daily topic level sentiment indicator, , by weighting the group level scores
according to their median ESS as observed over the past 365 days. The idea of using the median
ESS score over the last year is to come up with a systematic, non-price related weighting
function. With our suggested approach, we place more weight on event groups that generally
receive more extreme sentiment4:

Where is the median event sentiment score for event group

with representing the
set of event groups belonging to the relevant Topic including Business, Economy, Environment,
Politics, Society and Corporate. Finally,
is the normalized sentiment score for event group
on day .

3.1 Visual Index Representation

In Figures 2 and 3, we plot the 91-day moving average of corporate and economic sentiment
against the S&P 500 cumulative returns.5 To provide further insight into what our sentiment
indicators are measuring, we look at some of the underlying news events driving extreme
sentiment. For example, we find that corporate sentiment deteriorated strongly during the
financial crisis primarily due to companies having fewer corporate actions, paying lower
dividends, less favorable price targets set by analysts, and increasingly filing for bankruptcy. On
the other hand, the later recovery seems to have been driven primarily by improved analyst
expectations, improvements to earnings and dividends, and fewer bankruptcies.

Within each event group, more frequent events will have greater influence on the median score than less
frequent events.
5

The remaining sentiment dimensions are showcased in Appendix B.

Page|6

RavenPack Quantitative Research

Fig 2: The Corporate Sentiment Indicators (91-day Moving Average)

0.2

80%
More favorable analyst ratings and
price-targets, improvements in earnings
and dividends, fewer bankruptcies

Rise in business contracts,

product releases, and more
favorable analyst ratings

60%
40%

0.1
20%
0
0%
-0.1

-0.2

Rise in bankruptcies, price

target downgrades, lower
dividends, and less corporate
equity actions

Less favorable analyst ratings

and price-targets and more
bankruptcy filings

-0.3

Cumulative Return

91-Day Moving Average (Sentiment)

0.3

-20%
-40%
-60%

corporate

S&P 500 (secondary axis)

This figure presents the U.S. corporate sentiment indicator applying a 91 -day moving average.
The indicator is plotted against the cumulative returns of the S&P 500. The evaluated period
covers January 2002 through August 2014. Web content is introduced from January 2008.
SOURCE: RavenPack, Yahoo Finance, October 2014

Looking at the Financial Crisis from an economic point-of-view, the deterioration in sentiment
was driven primarily by lower consumer confidence, a fall in economic growth and a general
increase in recession-guidance. Higher employment and improved economic growth guidance
have contributed to the recent recovery.
Focusing on sentiment since 2013, we can observe that both corporate and economic
sentiment doesnt seem to support, at least visually, the continued strong market performance
realized over the last couple of years. But this is difficult to judge with the naked eye so we
proceed to test this in our predictive model in Section 3.

Page|7

RavenPack Quantitative Research

0.25

80%

0.2

Lower interest rate guidance, higher

economic growth, consumer
confidence and retail-sales
60%

0.15

Higher interest rates

(actual) and (expected),
less favorable GDP outlook

Strong economic growth

guidance and higher
employment

0.1

40%
20%

0.05
0

0%

-0.05

Cumulative Return

91-Day Moving Average (Sentiment)

Fig 3: The Economic Sentiment Indicators (91-day Moving Average)

-20%
-0.1
-0.15

Lower consumer confidence

and economic growth, increase
in recession expectations

Negative economic
growth guidance and
higher interest rate
expectations

-0.2

-40%
-60%

economy

S&P 500 (secondary axis)

This figure presents the U.S. economic sentiment indicator apply ing a 91-day moving average.
The indicator is plotted against the cumulative returns of the S&P 500. The evaluated period
covers January 2002 through August 2014. Web content is introduced from January 2008.
SOURCE: RavenPack, Yahoo Finance, October 2014

4. Evaluating MSI
To evaluate the predictive power of our MSIs, we construct a simple rolling linear regression
model considering a monthly prediction horizon.6 To benchmark the resulting strategies, we
create a simple buy-and-hold S&P 500 strategy as well as a pure price-driven momentum
strategy, including both a long-only and a long/short version.
Given our aim to predict the direction of the S&P 500 index, we create two sentiment-driven
models: 1) considering only corporate sentiment as we assume equity markets are primarily
driven by constituents news; and 2) a complete model including both corporate and macro
sentiment as well as market returns.

The results of a weekly model are included in Appendix C.

Page|8

RavenPack Quantitative Research

4.1 Model description

For model estimation, considering the complete model, we include 7 sets of variables for
Business, Economic, Environment, Politics, Society, Corporate and S&P 500 returns. To allow for
a lead-lag relationship between sentiment and price, we include up to six lagged monthly
variables for each of our indicators.7 Note that each variable is created by taking the average
topic sentiment score over the last 1-month horizon. The structure of our model is as follows:

With

representing the S&P 500 return at time and

representing the set of sentiment

variables with
represents the number of lagged values included in the model.

, while

The final model is selected using a stepwise Akaike Information Criterion (AIC) procedure with
bidirectional elimination. For each monthly step8, an automatic variable selecting is made and a
forward one period return prediction is made. To fit our model, we use the latest 5 years of
data as part of our estimation window - leaving us with 60 monthly data points, and up to 42
potential predictor variables. When evaluating the return-driven and corporate news strategies
only 6 variables are included representing the lagged monthly values.

4.2 Strategy Performance

When constructing our investment strategies, we hold a long position in the S&P 500 when our
1-month forward prediction is positive; while for the long-only and long/short strategies we
either stay in cash, yielding zero interest, or go short when a negative prediction is generated.
The performance of the different strategies is included in Figure 4 assuming zero transaction
costs. Over the back-testing period (September 2007 through August 2014), we find that both
the complete model (corporate & macro news + return), including all sentiment dimensions and
past returns, and the corporate news model, outperform a buy-and-hold S&P 500 strategy.
The sentiment strategies also significantly outperform the return-driven models. Even though
we find most of the performance is captured by corporate sentiment, the additional macro
variables improve the performance increasing the Information Ratio from 1.15 to 1.46 for the
long-only model and from 1.36 to 1.45 for the long/short model. This is compared to 0.23 for
7

For the weekly model in Appendix C, we allow up to 12 lags of each predictor variable with sentiment measured
as an average of the past 1-week.
8

The model is estimated 30 min before market close assuming that we can enter the market at the close price.

Page|9

RavenPack Quantitative Research

the buy-and-hold strategy and 0.73 for the return-based model. From an annualized return
perspective, the sentiment-based strategies yield about 13.5% and over 20% for the long-only
and long/short strategies, respectively.
Fig 4: Performance Statistics Predicting S&P 500 Returns (Monthly Model)
Long-only Monthly Model (6 lags)
Stat
Ann. Return
Ann. Volatility
Information Ratio
Max. Drawdown
Max. Recovery Period

S&P500
4.0%
17.2%
0.23
54.8%
72

Returndriven
7.7%
10.6%
0.73
10.6%
20

Corporate
News
13.6%
11.8%
1.15
11.1%
8

Corporate&Macro
News + Return
13.5%
9.2%
1.46
7.1%
17

Returndriven
10.0%
16.6%
0.60
15.6%
14

Corporate
News
20.9%
15.3%
1.36
9.5%
2

Corporate&Macro
News + Return
23.0%
15.9%
1.45
15.3%
11

L/S Monthly Model (6 lags)

Stat
Ann. Return
Ann. Volatility
Information Ratio
Max. Drawdown
Max. Recovery Period

S&P500
4.0%
17.2%
0.23
54.8%
72

This figure presents the performance statistics of a long -only (top) and long/short (bottom)
strategy with monthly investment horizons . The S&P 500 is a buy -and-hold strategy, while the
return-driven and corporate news strategies include S&P 500 returns or corporate sentiment
only as predictor variables. The corporate&macro news + return model includes all sentiment
indicators as well as S&P 500 returns (6 lags). The evaluated period covers September 2007
through August 2014.
SOURCE: RavenPack, Yahoo Finance, October 2014

The cumulative return profiles of the long-only and long/short strategies are included in Figure
5.

P a g e | 10

RavenPack Quantitative Research

Fig 5: Cumulative Return: Long-only Strategy (Monthly Model)
120%

Long-only

100%

Cumulative Return

80%
60%
40%
20%
0%
-20%
-40%
-60%
-80%

200%

Long/Short

Cumulative Return

150%
100%
50%

0%
-50%
-100%

S&P500 (Buy&Hold)

Return

Corporate News

Corporate&Macro News + Return

This figure presents the cumulative return of a long -only (top) and long/short (bottom) strategy
with monthly investment horizons. The S&P 500 is a buy -and-hold strategy, while the return driven and corporate news strategies include S&P 500 returns or corporate sentiment only as
predictor variables. The corporate&macro news + return model includes all sentiment indicators
as well as S&P 500 returns (6 lags). The evaluated period covers September 2007 through August
2014.
SOURCE: RavenPack, Yahoo Finance, October 2014

P a g e | 11

RavenPack Quantitative Research

When evaluating our strategies, we find that the number of included lags has an impact on the
overall strategy performance especially for the complete model. For instance, if we restrict
the number of lags to include only the past 3 months, the Information Ratio drops to about half
for the complete model (see Appendix D), while the corporate news model seems more stable with only a small drop to about 1.10. One explanation for this might be that corporate news
typically gets absorbed into equity prices faster than macro news, which may require a stronger
longer-term trend to emerge before impacting prices. This seems to be supported by our
weekly model results in Appendix C, where including macro variables only improves the
maximum drawdown profile and not the overall Information Ratio. Out of the 21 and 42
potential predictor variables for the 3-lag and 6-lag monthly models, on average only 9 and 24
variables survive the monthly step-wise AIC procedure.

5. Conclusion
We set out to create a set of macro sentiment indicators that captured sentiment abnormality
across 6 different news topics for an economy, namely business, economic, political,
environmental, societal and corporate. Employing these indicators in long-only and long/short
strategies on the S&P 500 we were able to generate additional value over traditional pricedriven models.
Aggregating company-specific sentiment into a macro corporate indicator is found to be the key
performance driver with noticeably stable results. Adding economic and geopolitical sentiment
indicators further improves the Information Ratio of our monthly model. Our news-based
models are able to achieve Information Ratios of around 1.45 for both a long-only and a
long/short strategy over our backtesting period (September 2007 through August 2014), while
the buy-and-hold S&P 500 strategy and the pure return-driven models only delivered
Information Ratios of 0.23 and 0.73, respectively. Besides the overall performance
improvements, we find news sentiment can be especially useful for managing downside risk as
our sentiment models reduce the maximum drawdown by more than 70% as compared to a
buy & hold strategy on the S&P 500.
In future research, we plan to explore the use of our macro sentiment indicators in other asset
classes including commodities, foreign exchange and fixed income. In addition, we plan to roll
out our sentiment index methodology across other countries and focus on sectors to address
the asset management challenges of constructing country and sector rotation models.

P a g e | 12

RavenPack Quantitative Research

Appendix A:
ESS EVENT SENTIMENT SCORE
A granular score between 0 and 100 that represents the news sentiment for a given entity by
measuring various proxies sampled from the news. The score is determined by systematically
matching stories typically categorized by financial experts as having short-term positive or
negative financial or economic impact. The strength of the score is derived from training sets
where financial experts classified entity-specific events and agreed these events generally
convey positive or negative sentiment and to what degree. Their ratings are encapsulated in an
algorithm that generates a score range between 0-100 where higher values indicate more
positive sentiment while lower values below 50 show negative sentiment.
ESS probes many different sentiment proxies typically reported in financial news and
categorized by RavenPack. The algorithm produces a score for more than 1,200 types of
business, economic, and geopolitical events ranging from earnings announcements to terrorist
attacks. The algorithms can dynamically assign an ESS score based on fixed scores assigned by
experts or by performing analysis and computation when figures or ratings are disclosed in the
story.
For example, the algorithm is capable of interpreting actual figures, estimates, ratings,
revisions, magnitudes and recommendations disclosed in news stories. It can compare actual
vs. estimated figures about earnings, revenues, and dividends and produce an ESS score
based on comparisons. It calculates percentage differences between financial figures and reads
and interprets stock and credit ratings disclosed by analysts. The ESS algorithms can factor
information such as the Richter scale in the case of an earthquake or the number of casualties
in a suicide bombing event. The use of emotionally charged language by authors is also factored
when shaping the strength component of ESS.
The ESS algorithm has embedded information on ratings scales from all major brokerage firms,
investment banks, and credit rating agencies. It uses this information to differentiate and assess
the various actions taken by analysts. For example, the algorithm generates a lower (more
negative) ESS score for stories about an analyst downgrade from a Strong Buy to a Strong Sell
than from a Buy to a Neutral. In the case of stories about financial results or economic
indicators, it computes the percentage change between the disclosed actual figures vs. the
street consensus or any other benchmarks disclosed in the story. For example, a company
beating earnings by 70% will receive a higher (more positive) ESS score than a company
exceeding a benchmark by 1%.

P a g e | 13

RavenPack Quantitative Research

ESS leverages RavenPacks event detection technology and produces an entity specific
sentiment score every time an event category is matched. ESS is based on RavenPack's Expert
Consensus and Event Score Factors methodologies.
RELEVANCE
A score between 0-100 that indicates how strongly related the entity is to the underlying news
story, with higher values indicating greater relevance. For any news story that mentions an
entity, RavenPack provides a relevance score. A score of 0 means the entity was passively
mentioned while a score of 100 means the entity was predominant in the news story. Values
above 75 are considered significantly relevant. Specifically, a value of 100 indicates that the
entity identified plays a key role in the news story and is considered highly relevant.
RavenPacks analysis is not limited to keywords or mentions when calculating relevance.
Automated classifiers look for meaning by detecting the roles entities play in specific events like
acquisitions or legal disputes or when announcing corporate actions, executive changes,
product launches or recalls, among many other categories. An entity will be assigned a high
mark of 100 if it plays a main role in these types of stories (context-aware). If an entity is
referenced in the headline or story body, it will receive a value between 0 and 99 (contextunaware). The score is assigned by a proprietary text positioning algorithm based on where the
entity is first mentioned (i.e. headline, first paragraph, second paragraph, etc.), the number of
references in the text, and the overall number of entities mentioned in the story. Usually, a
relevance value of at least 90 indicates that the entity is referenced in the main title or headline
of the news item, while lower values indicate references further down the story body.
G_ENS GLOBAL EVENT NOVELTY SCORE
A score between 0 and 100 that represents how "new" or novel a news story is within a 24hour time window across all news providers covered by RavenPack. The first story reporting a
categorized event about one or more entities is considered to be the most novel and receives a
score of 100. Subsequent stories from any news provider covered by RavenPack about the
same event for the same entities receive scores following a decay function whose values are
(100 75 56 42 32 24 18 13 10 8 6 4 3 2 2 1 1 1 1 0 ...) based on the number of stories in the past
24 hour window. If a news story is published more than 24 hours after any other similar story, it
will again be considered novel and start a separate chain with a score of 100. Note that for any
particular story, the G_ENS score is based on the number of similar stories in the most recent
24-hour window preceding that story across all news providers covered by RavenPack.

P a g e | 14

RavenPack Quantitative Research

Appendix B:
The business sentiment indicator captures information at a macro level concerning changes to
general supply and demand, products and services, business contracts, credit-ratings, industrial
accidents, regulatory approvals, labor-issues, and technical signals, among others.
Fig 6: The Business Sentiment Indicator (91-day Moving Average)
0.3

60%

0.2
40%
0.15

0.1

20%

0.05

0%

Cumulative Return

91-Day Moving Average (Sentiment)

0.25

80%
Increased investments, more
government contracts, and
general improvement to
supply/demand activity

-20%
-0.05
-40%

-0.1
-0.15

-60%

business

S&P 500 (secondary axis)

This figure presents the U.S. business sentiment indicator applying a 91 -day moving average. The
indicator is plotted against the cumulative returns of the S&P 500. The evaluated period covers
January 2002 through August 2014. Web content is introduced from January 2008.
SOURCE: RavenPack, Yahoo Finance, October 2014

P a g e | 15

RavenPack Quantitative Research

The politics sentiment indicator captures information at a macro level concerning elections,
foreign-relations, public-opinion, and changes to government.
Fig 7: The Political Sentiment Indicator (91-day Moving Average)
0.5

80%

0.3

Obamas approval
rating soars

Obamas approval rating rises

(Highest point in 3 years)

60%

Bushs approval
rating up

40%
0.2
20%

0.1
0

0%

-0.1
-0.2
-0.3

Bushs approval
rating keeps
dropping

Cumulative Return

91-Day Moving Average (Sentiment)

0.4

-20%

Bushs approval
rating at all time low
Obamas approval
rating drops below
50%

Obamas approval
rating plummets
on Jobs Plan

-0.4

Obamas approval -40%

rating drops (flawed
healthcare rollout)

-60%

politics

S&P 500 (secondary axis)

This figure presents the U.S. political sentiment indicator applying a 91 -day moving average. The
indicator is plotted against the cumulative returns of the S&P 500 . The evaluated period covers
January 2002 through August 2014. Web content is introduced from January 2008.
SOURCE: RavenPack, Yahoo Finance, October 2014

P a g e | 16

RavenPack Quantitative Research

The environmental sentiment indicator captures information at a macro level concerning

pollution and natural-disasters.

0.4

80%

0.2

60%

40%

-0.2

20%

-0.4

0%

-0.6

Hurricane
Katrina

-20%

Major
earthquake
in California
The Joplin tornado
(deadliest single U.S.
twister in more than
60 years)

-0.8

Cumulative Return

91-Day Moving Average (Sentiment)

Fig 8: The Environmental Sentiment Indicator (91-day Moving Average)

-1

-40%

-60%

environment

S&P 500 (secondary axis)

This figure presents the U.S. environmental sentiment indicator applying a 91 -day moving
average. The indicator is plotted against the cumulative returns of the S &P 500. The evaluated
period covers January 2002 through August 2014. Web content is introduced from January 2008.
SOURCE: RavenPack, Yahoo Finance, October 2014

P a g e | 17

RavenPack Quantitative Research

The societal sentiment index captures information at a macro level concerning war and conflict,
civil-unrest, crime, health, security and legal issues.
Fig 9: The Societal Sentiment Indicator (91-day Moving Average)
0.1

60%

40%

-0.1
20%
-0.2
0%
-0.3

-0.4

-0.5

-0.6

U.S. soldiers
and marines
killed in Iraq

Epidemic:
H1N1
(Swine flue)

U.S. soldiers
and marines
killed in Iraq

Occupy Wall
Street protest

Hepatitis
Outbreak in
Pennsylvania

-20%

Cumulative Return

91-Day Moveing Average (Sentiment)

Whooping
Cough
-40%
U.S. menigitis Epidemic
(California)
outbreak

-60%

society

S&P 500 (secondary axis)

This figure presents the U.S. societal sentiment indicator applying a 91 -day moving average. The
indicator is plotted against t he cumulative returns of the S&P 500. The evaluated period covers
January 2002 through August 2014. Web content is introduced from January 2008.
SOURCE: RavenPack, Yahoo Finance, October 2014

P a g e | 18

RavenPack Quantitative Research

Appendix C: Weekly Model Performance

When constructing our weekly investment strategy9, we hold a long position in the S&P 500
when our 1-week forward prediction is positive; while for the long-only and long/short strategy
we either stay in cash, yielding zero interest, or go short when a negative prediction is
generated. The performance of the different strategies is included in Figure 10 - assuming zero
transaction costs. Over our back-testing period, we find that both the complete model
(corporate & macro news + return), including all sentiment dimension and past returns, and the
corporate news model outperform a buy-and-hold S&P 500 strategy. Furthermore, the
sentiment strategy also significantly outperforms the return-driven models. Trading on a 1week horizon, we find that trading only on corporate sentiment outperforms the model
including macro variables, i.e. it seems that macro news doesnt add any additional predictive
power for short-term price movements.
Fig 10: Performance Statistics Predicting S&P 500 Returns (Weekly Model)
Long-only Weekly Model (12 lags)
Stat
Ann. Return
Ann. Volatility
Information Ratio
Max. Drawdown
Max. Recovery Period

S&P500
5.6%
18.7%
0.30
60%
323

Returndriven
2.7%
15.7%
0.17
58%
372

Corporate
News
9.8%
12.5%
0.79
19%
92

Corporate&Macro
News + Return
10.4%
13.1%
0.79
14%
21

Returndriven
0.4%
18.6%
0.02
59.0%
372

Corporate
News
14.6%
18.5%
0.79
28.9%
236

Corporate&Macro
News + Return
15.2%
18.6%
0.82
16.5%
118

L/S Weekly Model (12 lags)

Stat
Ann. Return
Ann. Volatility
Information Ratio
Max. Drawdown
Max. Recovery Period

S&P500
5.6%
18.7%
0.30
59.6%
323

This figure presents the performance statistics of a long -only (top) and long/short (bottom)
strategy with weekly investment horizons . The S&P 500 is a buy -and-hold strategy, while the
return-driven and corporate news strategies include S&P 500 returns or corporate sentiment
only as predictor variables. The corporate&macro news + return model includes all sentiment
indicators as well as S&P 500 returns ( 12 lags). The evaluated period covers September 2007
through August 2014.
SOURCE: RavenPack, Yahoo Finance, October 2014

For the weekly model, 3 years of data is used for fitting - leaving 156 data points and 84 potential predictors.

P a g e | 19

RavenPack Quantitative Research

Fig 11: Cumulative Return: Long/Short Strategy (Weekly Model)
120%

Long-only

100%
80%
60%
40%

20%
0%
-20%
-40%
-60%
-80%

150%

Long/Short

100%

50%

0%

-50%

-100%

S&P500 (Buy&Hold)

Return

Corporate News

Corporate&Macro News + Return

This figure presents the cumulative return of a long -only (top) and long/short (bottom) strategy
with weekly investment horizons . The S&P 500 is a buy-and-hold strategy, while the return driven and corporate news strategies include S&P 500 returns or corporate sentiment only as
predictor variables. The corporate&macro news + return model includes all sentiment indicators
as well as S&P 50 0 returns (12 lags). The evaluated period covers September 2007 through
August 2014.
SOURCE: RavenPack, Yahoo Finance, October 2014

P a g e | 20

RavenPack Quantitative Research

Appendix D: Monthly Model Performance (3-Lags)

Fig 12: Performance Statistics Predicting S&P 500 Returns (Monthly Model)
Long-only Monthly Model (3 lags)
Stat
Ann. Return
Ann. Volatility
Information Ratio
Max. Drawdown
Max. Recovery Period

S&P500
3.6%
17.0%
0.21
54.8%
72

Returndriven
5.0%
10.3%
0.49
18.4%
23

Corporate
News
11.3%
10.5%
1.08
9.5%
11

Corporate&Macro
News + Return
7.9%
11.4%
0.69
27.0%
29

L/S Monthly Model (3 lags)

Stat
Ann. Return
Ann. Volatility
Information Ratio
Max. Drawdown
Max. Recovery Period

S&P500
3.6%
17.0%
0.21
54.8%
72

Returndriven
4.9%
14.7%
0.33
21.5%
17

Corporate
News
17.0%
15.3%
1.11
16.4%
7

Corporate&Macro
News + Return
12.2%
16.7%
0.73
18.6%
5

This figure presents the performance statistics of a long -only (top) and long/short (bottom)
strategy with monthly investment horizons . The S&P 500 is a buy -and-hold strategy, while the
return-driven and corporate news strategies include S&P 500 returns or corporate sentiment
only as predictor variables. The corporate&macro news + return model includes all sentiment
indicators as well as S&P 500 returns ( 3 lags). The evaluated period covers September 2007
through August 2014.
SOURCE: RavenPack, Yahoo Finance, October 2014

P a g e | 21

RavenPack Quantitative Research

Terms of Use
This White Paper is not intended for trading purposes. The White Paper is not appropriate for the
purposes of making a decision to carry out a transaction or trade. Nor does it provide any form of advice
(investment, tax, legal) amounting to investment advice, or make any recommendations regarding
particular financial instruments, investments or products. RavenPack may discontinue or change the
White Paper content at any time, without notice. RavenPack does not guarantee or warrant the
accuracy, completeness or timeliness of the White Paper
You may not post any content from this White Paper to forums, websites, newsgroups, mail lists,
electronic bulletin boards, or other services, without the prior written consent of RavenPack. To request
consent for this and other matters, you may contact RavenPack at support@ravenpack.com.
THE WHITE PAPER IS PROVIDED AS IS, WITHOUT ANY WARRANTIES. RAVENPACK AND ITS AFFILIATES,
AGENTS AND LICENSORS CANNOT AND DO NOT WARRANT THE ACCURACY, COMPLETENESS,
CURRENTNESS, TIMELINESS, NONINFRINGEMENT, TITLE, MERCHANTABILITY OR FITNESS FOR A
PARTICULAR PURPOSE OF THE WHITE PAPER, AND RAVENPACK HEREBY DISCLAIMS ANY SUCH EXPRESS
OR IMPLIED WARRANTIES. NEITHER RAVENPACK NOR ANY OF ITS AFFILIATES, AGENTS OR LICENSORS
SHALL BE LIABLE TO YOU OR ANYONE ELSE FOR ANY LOSS OR INJURY, OTHER THAN DEATH OR
PERSONAL INJURY RESULTING DIRECTLY FROM USE OF THE WHITE PAPER, CAUSED IN WHOLE OR PART
BY ITS NEGLIGENCE OR CONTINGENCIES BEYOND ITS CONTROL IN PROCURING, COMPILING,
INTERPRETING, REPORTING OR DELIVERING THE WHITE PAPER. IN NO EVENT WILL RAVENPACK, ITS
AFFILIATES, AGENTS OR LICENSORS BE LIABLE TO YOU OR ANYONE ELSE FOR ANY DECISION MADE OR
ACTION TAKEN BY YOU IN RELIANCE ON SUCH WHITE PAPER. RAVENPACK AND ITS AFFILIATES, AGENTS
AND LICENSORS SHALL NOT BE LIABLE TO YOU OR ANYONE ELSE FOR ANY DAMAGES (INCLUDING,
WITHOUT LIMITATION, CONSEQUENTIAL, SPECIAL, INCIDENTAL, INDIRECT, OR SIMILAR DAMAGES),
OTHER THAN DIRECT DAMAGES, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGES. IN NO
EVENT SHALL THE LIABILITY OF RAVENPACK, ITS AFFILIATES, AGENTS AND LICENSORS ARISING OUT OF
ANY CLAIM RELATED TO THIS AGREEMENT EXCEED THE AGGREGATE AMOUNT PAID BY YOU FOR THE
WHITE PAPER IN THE 12 MONTHS IMMEDIATELY PRECEDING THE EVENT GIVING RISE TO SUCH CLAIM.
BECAUSE SOME STATES OR JURISDICTIONS DO NOT ALLOW THE EXCLUSION OR LIMITATION OF
LIABILITY FOR DAMAGES OR THE EXCLUSION OF CERTAIN TYPES OF WARRANTIES, PARTS OR ALL OF THE
ABOVE LIMITATION MAY NOT APPLY TO YOU.
These Terms of Use, your rights and obligations, and all actions contemplated by these Terms of Use will
be governed by the laws of New York, NY, USA and You and RavenPack agree to submit to the exclusive
jurisdiction of the New York Courts. If any provision in these Terms of Use is invalid or unenforceable
under applicable law, the remaining provisions will continue in full force and effect, and the invalid or
unenforceable provision will be deemed superseded by a valid, enforceable provision that most closely
matches the intent of the original provision.t

P a g e | 22