Transcript 2010-06-16-ny-Renick

RESEARCH ON SENTIMENT-BASED SMART MONEY INVESTMENT
MODELS AND IMPLICATIONS FOR INVESTORS
QWAFAFEW NYC CHAPTER MEETING
JUNE 16, 2010
Dirk Renick, PhD
Manager – Quantitative Research, Thomson Reuters StarMine
Why are we researching “smart money”?
Following the dumb money doesn’t seem like such a good idea...
$ Billions
12.00
5500
10.00
5000
4500
8.00
4000
6.00
3500
4.00
3000
NASDAQ
Fund Flows
2.00
2500
0.00
2000
Nov-02
Sep-02
Jul-02
May-02
Mar-02
Jan-02
Nov-01
Sep-01
Jul-01
May-01
Mar-01
Jan-01
Nov-00
Sep-00
Jul-00
May-00
Mar-00
Jan-00
Nov-99
Jul-99
Sep-99
May-99
Mar-99
Jan-99
Nov-98
Jul-98
Sep-98
1000
May-98
(4.00)
Mar-98
1500
Jan-98
(2.00)
Tech sector fund flows 1998 - 2002
Source: Lipper FMI Monthly Technology Fund Flows, Datastream
2
AGENDA
• HOLDINGS MODELS
– TOP HOLDINGS
– SMART HOLDINGS
• INSIDER TRANSACTIONS
• SHORT INTEREST
• CONCLUSION, Q & A
DATA MINING TIP #1: Start with a hypothesis based
on economic intuition, even though mileage may vary.
• Economic intuition usually implies exploiting a behavioral bias
• When you get a test result, you can compare results to your baseline
hypothesis, which is better than not having a well-formed hypothesis
–
Explore what assumptions were wrong in your hypothesis
–
Maybe find out that your hypothesis was only a proxy for another known market
anomaly
–
Sometime you pick up random, interesting tidbits that can inform future hypotheses
4
Popular idea that didn’t show consistent alpha:
fund manager’s “top holdings”
• Hypothesis: Mutual fund managers’ top several positions comprise
their most confident bets and achieve outperformance.
• Fact: Their most overweight positions are largely comprised of stocks
that have undergone a price run-up in the prior 1-2 quarters
• Conclusion: We found inconsistent alpha performance over time, with
a substantial drop between the first and second halves of our data
history
average monthly
return
α4
1997 - 12/2002
0.0131
0.0123
1/2003 - 5/2009
0.0118
0.0016
Time period
5
DATA MINING TIP #2: When the investing
environment changes (gov’t action, etc.) test before
and after separately.
• In the case of investor sentiment models, Reg FD is an
exogenous shock to the disclosure landscape
– Reg FD was enacted, in part, to level the information playing field
between the “smart money” and the retail investor
• In the case of “best ideas” we saw a step change in factor
efficacy
6
“Top Holdings” portfolios show significantly higher
trailing returns than forward returns – it’s momentum
Mean monthly return
Rather than a function of conviction, “Top Holdings” are dominated by
stocks that have run up in price in the trailing 3-6 month period.
Holdings as_of_date (0 day report lag)
Data available date (60 day report lag)
Previous months
Following months
7
Top holdings are more likely to be sold by the fund
• These largest positions are more likely to be sold by the fund in
the following quarter than they are to be bought further.
• In the quarter following a top holding showing up, fund managers are
over 3 times more likely to sell at least 5% of their position, as they
are to buy more.
Top holdings this quarter
% being bought next Q
% being sold next Q
14.6%
46.1%
• This corresponds with the belief that fund managers are rebalancing
to avoid being concentrated in any given position.
8
AGENDA
• HOLDINGS MODELS
– TOP HOLDINGS
– SMART HOLDINGS
• INSIDER TRANSACTIONS
• SHORT INTEREST
• CONCLUSION, Q & A
Holdings data represents a rich dataset to model
institutional and investment management behaviors
Global Ownership Data
• Mutual funds, HF, institutions
• Global securities
• 13 years of data, from 1997
(around 23 GB of data)
Combine with
• Fundamental data
• Analyst Estimates data
• Price data
Understand
Investor
Behaviors
10
What are the
fundamental factor
biases of each
fund/firm ?
Are investors
favoring value or
growth this year?
What stocks are
most attractive to
large numbers of
funds ?
Research into a “Smart Holdings” model of buying and
selling behavior reveals some interesting biases.
Research into historical holdings data uncovered several factors that
help us predict future holdings:
• Funds are attracted to companies that are “like” ones they already
own. Likeness can be on an analyst coverage, sector or
fundamental basis.
• Funds exhibit biases towards buying companies with certain
fundamental factors, and these biases change over time.
• This is different from the definition of “conviction” or “best ideas” in the literature
which attempts to use holding size as an indicator.
The model is based on predicting future change in % institutional
holding for each company rather than price.
If we can accurately predict which stocks are likely to be
bought (sold), we can forecast price increases (decreases).
11
DATA MINING TIP #3: Predicting price changes can
be hard; it may be more fruitful to focus on predicting a
behavioral phenomena that should drive prices.
In the case of looking at holdings, we are focusing on predicting
the future change in institutional holdings
• Similar tactic undertaken in development of the StarMine Analyst
Revisions model which predict future revisions, not price.
You can validate the behavioral hypothesis with a perfect
foresight model
• A perfect foresight model is unachievable but useful to know that you
are going after a worthwhile target.
12
By predicting what factors are most important, you can
model which stocks managers will find most attractive
Public filing
Predict Q1 holdings
one month prior
Changes in % held by institutions
cause price changes
Predicting forward one quarter change
in % Institutional ownership is highly
profitable
A perfect foresight model is off the
charts performance. So this is
worth modeling.
Q1 +45 days
Q0
A good measure of a stock’s popularity
is % Institutional ownership.
Pred. Change
12/31/07 Actual Chg
Rank
Ticker
9/30/07
MSFT
58.60%
61.80%
3.20%
70
IBM
65.40%
64.00%
-1.40%
56
XOM
49.90%
50.90%
1.00%
53
MON
84.90%
85.40%
0.50%
41
C
65.60%
63.40%
-2.20%
10
Spearman rank correlation between F1Q predicted
%inst ownership and actual F1Q % inst ownership 0.14
13
Below we can see how different biases (or investing
styles) come in and out of favor over time.
Price momentum and growth
factors dominate
1999
StarMine PriceMo
LTG
G5 EPS
Debt/Assets
Interest Coverage
2001
EPS_CAGR3
ROE
Profit Margin
Debt/Assets
LTG
2003
ROE
Profit Margin
Interest Coverage
LTG
F12m E/P
ROE, Earnings Quality and other
value factors dominate
2004
ROE
Interest Coverage
F12m E/P
Profit Margin
LTG
2005
2007
ROE
ROE
F12m E/P
F12m E/P
Interest Coverage Interest Coverage
Profit Margin
Profit Margin
StarMine EQ
StarMine PriceMo
in-sample years 1999-2007
The shift from growth to value matches intuition around how
investors approached these different market regimes.
14
Microcap companies, which we define as not being in
the top 98.5% of all market cap, can present different
phenomena and potentially misleading aggregate
results
Average decile returns, all securities
with median market cap of decile
Average decile returns, top 98.5% market cap
with median market cap of decile
2.5%
281M
2.0%
230M
1.5%
69M 106M 126M
1.0%
2.5%
214M
23M
Avg Monthly Return
Avg Monthly Return
3.0%
151M 175M
205M
0.5%
693M
2.0%
902M
1.5%
777M
612M
1.0%
360M
0.5%
431M 505M
690M
569M
234M
0.0%
0.0%
1
2
3
4
5
6
Decile
7
8
9
10
Microcaps included in Universe
1
2
3
4
5
6
Decile
7
8
9
Microcaps excluded
This represents a better
investible universe
10
AGENDA
• HOLDINGS MODELS
– TOP HOLDINGS
– SMART HOLDINGS
• INSIDER TRANSACTIONS
• SHORT INTEREST
• CONCLUSION, Q & A
SOX (and perhaps the Martha Stewart trial!) represented
watershed moments in efforts to model insider trading
Every literature theory we tested has significant performance drop-offs after SOX.
• Maybe due to SOX; maybe market cycles; maybe Martha Stewart (seriously!)
“We find that the cost of equity in a country, after controlling for a number of other variables, does not
change after the introduction of insider trading laws, but decreases significantly after the first
prosecution.”
Bhattacharya, Daouk, Journal of Finance, 2002, “The World Price of Insider Trading”
17
We investigated whether we could still get value out of
insider data and found three dimensions informative
• Breadth (Net Buyer Ratio)
• The more agreement there is among insiders as to buying (selling), the more
bullish (bearish) we should be on the company.
• Buying Depth
• The more $ that insiders are buying, the more bullish we should be on the
company.
• Selling Depth
• The more $ that insiders are selling, the more bearish we should be on the
company
• Adjust for “why is this insider selling”: there are many reasons insiders might sell
shares
• Some of those reasons have little to do with their outlook for the business
• We achieve better results by putting less weight on transactions that look more
like compensation (e.g. an insider who exercises options and immediately sells
them is probably thinking of the option grant as a bonus.)
18
DATA MINING TIP #4: Does your model improve over
a simple model?
Top
Decile
Bottom
Decile
Decile
Spread
IC
StarMine Insider Model U.S.
10.3%
-0.1%
6.9%
0.014
Basic Buy/Sell Ratio Model
8.7%
6.2%
2.2%
0.007
4.7%
0.007
StarMine Value Added
1996-2008; in-sample stocks only
19
DATA MINING TIP #5: Did you rediscover a well
understood investor bias?
Check your model against correlations of “standard” quant
models (could include Beta, MCAP, etc… as well)
ARM
Spearman
Correlation: -0.036
EQ
IV
PriceMo
RV
ValMo
-0.129
0.213
-0.026
0.199
0.144
Combining the Insider model (value) with Price
Momentum achieves better results than either alone,
which matches well-known quant model behavior
• Insiders tend to behave like value investors
• When they violate this tendency, it is an especially powerful signal
(e.g. when insiders sell a stock that has already declined a lot)*
Quintile
1
StarMine
Insider
Trading
Model US
Quintile
1
Quintile
2
Quintile
3
Quintile
4
Quintile
5
-16.5%
StarMine PriceMo
Quintile Quintile Quintile
2
3
4
-9.2%
-6.2%
0.2%
Quintile
5
8.5%
Insider trading spread = 6.5%
-4.2%
Price momentum spread = 18.5%
Simple combination = 29.5%
-7.5%
-6.0%
-2.0%
4.3%
7.4%
-0.4%
-6.4%
-2.6%
-1.1%
5.2%
6.2%
0.9%
-9.5%
-2.3%
0.8%
4.6%
11.5%
0.9%
-6.3%
-4.0%
0.1%
6.3%
13.0%
2.3%
-9.3%
-4.7%
-1.5%
4.4%
9.2%
Annualized excess returns of the StarMine Insider Trading Model US and the StarMine PriceMo model, alone and
in combination, broken out into quintiles, for a 30-day holding period rebalanced monthly from January 1996December 2008. Quintile 1 is the most bearish; quintile 5 the most bullish. Transaction costs are not included.
*consistent with Seyhun, N. 1998, “Investment Intelligence From Insider Trading”, MIT Press
21
AGENDA
• OWNERSHIP ANALYTICS
– TOP HOLDINGS
– SMART HOLDINGS
• INSIDER TRANSACTIONS
• SHORT INTEREST
• CONCLUSION, Q & A
The fraction of U.S. stock shares sold short has
increased steeply over the last 15 years
Short Interest as a % of shares outstanding, top 3000 US stocks by market cap.
10.0%
9.0%
8.0%
% of shares outstanding held short
7.0%
6.0%
15 year average level (3.98%)
5.0%
4.0%
3.0%
2.0%
1.0%
0.0%
1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010
This trend at least in part reflects the increased popularity of market
neutral and 130/30 funds in the U.S. marketplace.
23
What are we trying to do with a short interest model?
Three broad categories of short sellers represented in the raw
short interest data from the exchanges:
• Hedgers – risk management technique to reduce exposure
incurred through long investments (130/30, pairs trades)
• Arbitrage shorts – exploit mispricing between two assets or asset
classes. (convert arbs, M&A arbs)
• Fundamental Shorts - Valuation based shorts based on investors
sentiment that the stock is overvalued by some reasonable amount
or the company is going BK. This is the “smart” money making a
directional bet on stock price.
Our model intelligently separates the Fundamental shorts
from the non-informative arb shorts, and then enhances
signal due to loan availability and sell-side sentiment
24
Can you make money following the shorts?
Yes, and with low turnover and correlation!
Top decile –
lightly shorted
Cumulative return based on
a long minus short strategy
Bottom decile –
heavily shorted
• An investment strategy based on ranking top 3000 market cap US stocks by
short interest as a percent of shares outstanding returned an average of 13% a
year (compared to an average of 5.5% a year for the Russell 3000) from
December 2003 to early 2009.
• Correlations to StarMine’s existing models are low, ranging from 0.09 for EQ to
0.26 for Val-Mo.
• Total average annualized turnover = 113%. Annual turnover of top decile = 38%.
25
Applying economic intuition, you can achieve superior
performance by considering additional data sources
beyond simple short interest level
1.
Intelligently separate value shorts from the arb/hedge shorts.
–
Identify stocks that are likely candidates for risk arbitrage or convertible arbitrage
strategies.
2.
3.
Consider how ‘expensive’/difficult it is to short each stock.
–
Level of institutional ownership is a proxy for cost of borrowing.
–
Consider dividend payments as an additional cost of holding stock short.
Enhancing model performance with an analyst recommendations
“kicker”
–
When analysts have buy recommendations on highly shorted stocks go with the
shorts
4.
Sector affects are present in short interest sentiment
5.
Beware of the dreaded Short Squeeze!
–
Consider risk factors for heavily shorted stocks that are likely to be squeezed.
26
Companies with convertible bond issues see a sharp
rise in short interest level at the announcement date
However, on average, they outperform heavily shorted companies
WITHOUT convertible issues by 38bps/month.
post-issue avg.
(8.6%)
Further 2% increase in short
interest following debt issuance.
pre-issue avg.
(6.5%)
Convertible debt issuers are more
heavily shorted even prior to
issue.
2003-2008 total
universe avg.
(4.1%)
27
Short interest can be misleading when there is convertible
bond arbitrage in play – you should account for that
$1.2B convertible bond
issue announced
100
80
Gilead Sciences(GILD)
$50.00
70
$40.00
60
50
$30.00
Price
Rank of Short Interest/Shout
90
$60.00
40
$20.00
30
20
$10.00
10
0
12/1/2003
12/1/2004
12/1/2005
12/1/2006
baseline mcap neutral rank
12/1/2007
$0.00
12/1/2008
close price
Low short interest rank represents a heavily shorted stock
28
Consider how expensive/difficult it is to short each stock
1.
Fraction of shares owned by institutions as a proxy for borrow rate
2.
Dividends – additional cost of being short
29
Conditioning on Institutional Ownership (as Rebate
Rate proxy)
Short Interest Quintiles
• Stocks with low levels of institutional ownership are more costly to
short, therefore the same short level on these names should be
associated with higher expected profit.
• The most heavily shorted stocks have the largest spreads when they
are also in the lowest 2 quintiles of institutional ownership.
lowest inst
ownership
most heavily shorted
-0.84%
2
-1.14%
3
-0.41%
4
-0.25%
highest inst
ownership
-3.73E-05
si only (collapse
across ownership
-0.53%
2
0.29%
-0.21%
-0.29%
-0.13%
0.22%
-0.02%
3
0.81%
-0.01%
0.18%
0.46%
0.19%
0.32%
4
0.36%
0.20%
0.34%
0.18%
0.17%
0.25%
least shorted
0.45%
0.60%
0.48%
0.46%
0.41%
0.48%
1.29%
1.73%
0.89%
0.71%
0.42%
outperf.
outperf
underperf
underperf
underperf
spread with ownership
buckets:
1.01%
5X5 quintile plot of mean monthly returns to portfolios formed using short interest and fraction of
lowest inst
highest inst
institutional ownership compared to baseline short interest (2003-2008)
counts ownership
6247
most heavily shorted
6272
2
6275
3
2
6250
6278
6279
3
6247
6277
6279
4
6250
6276
6277
ownership
6258
6281
6282
30
Analyst recommendations “kicker”
1. Agreement – we see outperformance at the positive agree corner of
our plot, and underperformance in the negative agree corner.
2. Contrarian – we also see underperformance of the High Short
Interest / High Recommendations bucket (circled in red).
When short-sellers and the sell-side disagree, you’re better off
betting with the shorts.
Short Interest Quintiles
Analyst Recommendation Quintiles
Highest SI = Sell short
Lowest SI = Buy
Quintile Spread (SI)
Best = Buy
-0.57%
-0.11%
0.46%
0.20%
0.60%
-0.41%
-0.31%
0.03%
0.47%
0.58%
-0.24%
-0.14%
-0.04%
0.16%
0.52%
-0.77%
-0.34%
-0.12%
0.13%
0.33%
1.18%
0.99%
0.76%
1.10%
Worst = Sell
-0.60%
-0.25%
0.01%
-0.21%
0.41%
1.01%
si only
-0.52%
-0.23%
0.07%
0.15%
0.49%
1.01%
5X5 quintile plot of mean monthly returns to portfolios formed using short interest and
recommendations (2003-2008)
31
Beware of the dreaded Short Squeeze!
“Short Squeeze” means different things to different people. We look for a
large forward 1-month “draw-up”.
What is F1M draw-up?
The maximum % price increase
from the first day of the period.
e.g., for HOV on 2007-7-31:
F1M draw-up = (16.22 - 11.95)/11.95
= 35.7%
Our goal is to predict the rank of F1M draw-up, rather than an absolute value.
Otherwise, we would have to arbitrarily answer questions such as:
-What price increase comprises a “squeeze”? 15%? 25%? 50%?
-How long does it last? Days? Weeks? Months?
32
We found that Days to Cover, a commonly used
predictor of short squeezes, does not work
• Days to cover = “Short Ratio” = # Shares Short / Avg Daily Volume (T1M)
•
Our primary measure is hit rate (0-1, higher is better). It’s the number of
stocks actually in the top decile of F1M draw-up / number of stocks predicted
to be in the top decile. Basically, it tells you how correct your highestconviction predictions were.
• The hit rate for the short ratio is essentially the same as a random number
generator.
Hit Rate for
Random Model
= 0.1
We have developed an indicator that has a significantly better hit rate.
33
AGENDA
• OWNERSHIP ANALYTICS
– TOP HOLDINGS
– SMART HOLDINGS
• INSIDER TRANSACTIONS
• SHORT INTEREST
• CONCLUSION, Q & A
What did we learn about smart money?
It is possible to follow the smart money to abnormal returns, but...
– With ownership data, you can’t just buy the “best ideas,” largest
holdings, or recent purchases, you must anticipate what is going
to be bought or sold in the future
– With insider trading, you do much better if you isolate behavior
reflecting opinion from compensation
– With short interest, it’s important to separate hedging and
arbitrage shorts from the fundamental bets
35
Questions?
Dirk Renick
[email protected]
36