Starting from this fact, design a simple model to measure
47
momentum (e.g., 6-month return). Then sort stocks by this metric and buy the ones that score well.
The worst thing to do is take a predefined model and see if it
works. Has a 30-day, 200-day moving average crossover been
predictive of VIX futures? What if we change the first period to 50
days? I don't know or care.
Fundamental Analysis
Fundamental analysis aims to predict returns by looking at
financial, economic, and political variables. For example, a
fundamental stock analyst might look at earnings, yield, sales, and
leverage. A global macro trader might consider GDP, currency
levels, trade deficit, and political stability.
Fundamental analysis, particularly global macro, is particularly
susceptible to subjectivity. It also tempts otherwise intelligent
people to make investment decisions based on what they read in
the Wall Street Journal or The Economist. It is exceedingly
unlikely that someone can consistently profit from these public
analyses, no matter how well the story is sourced or how smart the
reader is.
Consider these statements from “experts”:
“Financial storm definitely passed.”
—Bernard Baruch, economic advisor to presidents
Woodrow Wilson and Franklin Roosevelt in a cable to
Winston Churchill, November 1929
Stocks dropped for the next 3 years, with the Dow losing 33% in
1930, 52% in 1931, and 23% in 1932.
“The message of October 1987 should not be taken lightly. The
great bull market is over.”
—Robert Prechter, prominent Elliot wave theorist and
pundit, in November 1987. The Dow rallied for 11 of
the next 12 years, giving a return (excluding dividends)
of over 490%.
“A bear market is likely… It could go down 30% or 40%.”
—Barton Biggs, chief strategist for Morgan Stanley,
October 27, 1997
48
The Dow had its largest one-day gain on October 28 and
continued to rally hard for the next 6 months.
In most situations it is just mean to make fun of people's mistakes.
We all make mistakes. But the people I have quoted have
proclaimed themselves experts in a field where real expertise is
very, very rare.
And evidence of this is more than anecdotal.
The poor prediction skill of experts is a general phenomenon. Gray
(2014) summarizes the results of many studies that show that
simple, systematic models outperform experts in fields as diverse
as military tactics, felon recidivism, and disease diagnosis.
Expertise is needed to build the models, but experts should not
make case-by-case decisions.
Koijen et al. (2015) show that surveys of economic experts
(working for corporations, think tanks, chambers of commerce,
and NGOs) have a negative correlation to future stock returns.
They were also contraindicative for the returns of currencies and
bonds. This effect applies across 13 equity markets, 19 currencies,
and 10 fixed income markets. A simple “fade the experts” strategy
would have given a Sharpe ratio of 0.78 from 1989 to 2012.
Financial advisors are equally bad. Jenkinson et al. (2015) look at
the performance of advisors in picking mutual funds. They
conclude with, “We find no evidence that these recommendations
add value, suggesting that the search for winners, encouraged and
guided by investment consultants, is fruitless.” And fund
managers themselves can't consistently beat the averages. Due to
costs, most managers underperform and there is no correlation
between performances from one year to the next. So, managers
can't pick stocks and it is pointless to try to pick good managers.
It is also likely that much of the alpha generated by fundamental
analysis is smart beta, compensation for exposure to a certain risk
factor. There is absolutely nothing wrong with this. Trading profits
are profits, no matter whether they are due to smart beta or alpha.
But before we ascribe a trader's results to skill, we should know
what is causing the profits. Beta should cost a lot less than alpha.
Conclusion
49
It is difficult to make money in financial markets. The EMH isn't
completely true, but it is closer to being correct than to being
wrong. If a trader can't accept this, she will see edges in noise and
consequently overtrade. Behavioral finance, technical analysis,
and fundamental analysis can all be used as high-level organizing
principles for finding profitable trades, but each of these needs to
be believed only tentatively, and the most robust approach is to
look for phenomena that are independently clear. For example,
momentum can be discovered through technical analysis but also
understood as a behavioral anomaly. The observable phenomenon
must come before any particular method.
Summary
Exceptions to the EMH exist but they are rare.
Exceptions will either be inefficiencies, temporary phenomena
that last only until enough people notice them, or poorly
priced risk premia.
Risk premia will persist and can form the core of a trader's
operations but the profits due to inefficiencies will decay
quickly and need to be aggressively exploited as soon as they
are found.
A promising trading strategy is one whose basis is independent
of the specific methods used to measure it. Start with
observation, then move to quantification and justification.
50
CHAPTER 3
Forecasting Volatility
All successful trading involves making a forecast. Some traders
(for example, trend followers) say they don't forecast, they react. I
don't know why they say this, but in any case, they are wrong. The
moment a trader enters an order, she has implicitly made a
forecast. Why would you get long if you didn't think the market
was going up? No matter how it was arrived at, the forecast is, “the
market is going up.” Except for a pure arb (which are practically
extinct), to get positive expectation we need to make a forecast
that is both correct and more correct that the consensus.
In this chapter we will concentrate on making correct forecasts of
volatility. But, first, here are some principles that are applicable to
any financial forecasting:
Pick a good problem. Some things are impossible to forecast.
No one can predict the price of AAPL in 25 years. Some things
are hard to predict. Forecasting the S&P 500 index in two days
is a hard problem. Some things are trivial to predict. The FED
funds rate in the next day is almost certainly going to be
unchanged. Aim to find problems that are solvable but are
hard enough that you will be able to profit from the
predictions. Volatility is a perfect candidate for this.
Actively look for comparable historical situations. What
happens when the government shuts down? What is the link
between recessions and the stock market? This is a good
>
general principle, but it is also vital if you are looking for
catalysts that could lead to volatility explosions. Good periods
to be short volatility can often be deduced from financial data
alone, but long trades generally need a catalyst (that isn't
priced in) to be successful. Don't trust your intuition or what
you think is true. These will be biased by your experiences,
environment, and political persuasion. If you don't have data,
you don't have knowledge.
“When my information changes, I alter my conclusions.
What do you do, sir?”
—J. M. Keynes
51
Aim to balance being conservative and reactive. All good
investors have a Bayesian model in their head and update their
forecasts as new information arrives but you also shouldn't
update too aggressively.
Actively look for counterarguments. Every person has biases.
If you are convinced that every article you read is a harbinger
of chaos, be open to the possibility that you are wrong. And the
same holds if you are a habitual volatility seller.
“It is impossible to lay down binding rules, because two
cases will never be exactly the same.”
—Field Marshall Helmuth von Moltke
Remember that there are no certainties when predicting the
future.
Model-Driven Forecasting and Situational
Forecasting
An effective way to learn is to organize our current knowledge.
Sometimes this makes it clear that we don't understand certain
things, but even if no such gaps become apparent the thought that
goes into a classification scheme is helpful. Science often starts by
classifying knowledge. We knew about species before we knew
about speciation through natural selection. We knew that
elements could be grouped into the periodic table before we knew
about atomic structure. We knew about dominant and recessive
genes before we knew about DNA structure.
We have already classified trading opportunities into inefficiencies
and risk premia. This distinction is important on a strategic level.
Mispriced risk premia can last for a long time. A business can be
built on harvesting risk premia. Inefficiencies aren't likely to be as
persistent. These need to be aggressively traded while they last,
and we can assume that they won't last long.
There is also an important classification of trades at the tactical
level (strategy defining the high-level goals and tactics being the
methods we use to reach them). Trades are either model driven or
event driven.
With a model-driven trade, we have a theoretical model of a
situation that lets us calculate a fair value or edge. At any moment,
52
we will have an opinion based on the model. For example, if we
have an option pricing model, we can continually generate a
theoretical value for all of the options on a given underlying. An
event-driven trade is based on a specific unusual situation.
Noticing that implied volatility declines after a company releases
earnings would be the basis for an event-driven trade.
All types of trading, investing, or gambling can be classified like
this. In blackjack, card counting is model driven. The player's
counting scheme assigns a value for each card that is dealt. As
cards are played, the player updates the count, and modifies her
edge estimate accordingly. At any point in the dealing, she will
know what her edge is. But there is an event-driven method as
well: ace tracking. Ace tracking is based on the fact that shuffles
aren't perfect randomizers. Cards that are close together in one
shoe will tend to stay close together in the next shoe. So an ace
tracker notes the cards that are located close to aces and when
those same cards are dealt in the next shoe she knows that there is
a heightened chance of an ace being close. Because the player
advantage in blackjack is due to the 2–1 blackjack payout, having a
better than random idea where aces are is enough to give a
significant edge. Ace tracking can be more effective than card
counting.
Stock investing can be similarly classified. We can rank stocks
with a factor model such as Fama-French-Carhart, or we could
buy stocks that have had positive earnings surprises. In horse
racing, an example of a model would be Beyer's speed figures,
whereas an event-driven strategy would be to back horses in
certain traps.
Both of these approaches have strengths and weaknesses.
For a model-driven approach to be effective we need a good
model. Some situations lend themselves to this more than others.
For example, there are very good option valuation models, but
stock valuation methods are crude. Sometimes, the work required
to build a model is not worth it. But if we have a model, we will
always be able to trade. We will have a theoretical value for every
trade opportunity. This means the approach scales very well and
has great breadth of applicability. We will also be able to scale our
trades based on our perceived edge.
The largest problem with this approach is that models have to be
vastly simplified views of reality. Often a model's apparent
53
effectiveness isn't due to its effectiveness but more because data gathering and processing is being rewarded. In the 1980s
collecting daily closing prices and calculating volatility from those
was enough to give a volatility edge in the options markets. Now
that this data is free and easy to automatically process, it seems
like there is no edge left in this volatility arbitrage model. But
there never was any edge in the model at all: the edge was in data
collection and processing.
Event-driven trades have two great strengths. The process for
finding and testing them is very simple. What happens to stocks in
the three days after a FED meeting? What does the VIX do on
Mondays? Are teeny options overpriced? All we need to test these
ideas is data and a spreadsheet.
Most important, trades that are based on specific events or
situations can be very profitable. I have one trade that has literally
never lost money. It only sets up a few times a year and is quite
liquidity constrained, but it has an unblemished record. This
profitability is probably linked to the fact that there is huge
uncertainty about why this situation is lucrative.
A drawback of event-based trades is that we must wait for the
event, and some events don't occur very often. It is hard to
structure a business based on a strategy that might not trade for
years at a time. Also, it is often hard to know why the trade exists.
This isn't always true. For example, ace tracking is profitable for a
very clear reason. But sometimes even a trade with compelling
statistics has no obvious reason. I don't do trades if I have no idea
why they exist, but sometimes it is easy to come up with a post-
hoc reason. For example, many sports fans have convinced
themselves that home field advantage is due to travel fatig
ue. This
seems plausible, but it is wrong. Even when teams share the same
ground the home team has an advantage. There is no magic
answer to this dilemma. The weaker the evidence for a cause is,
the stronger the statistical evidence needs to be.
Related to this problem is that if we have only a vague idea for why
a trade works, we will have a hard time knowing if it has stopped
working or if we are just experiencing a bad period. This is
particularly true if the proposed reason is a psychological one. It is
always tempting to ascribe any anomaly to psychology. This
inevitably leads to overconfidence in the trade. After all, human
54
psychology isn't going to change so why would these trades ever
stop working?
Finally, we will often have no way to differentiate between “good”
and “bad” trades of the same class. If all we know is that selling
options over earnings is profitable, we won't know if it is better to
sell Apple options or IBM options. This makes sizing difficult. We
only have statistics for the entire class of trades. We need to be
very conservative.
Models can give a false sense of security. No model can account
for everything. No event has a single cause. Most events have
many causes. A situational strategy directly acknowledges this
uncertainty and generally traders who are comfortable with
uncertainty will do best. So, although creating models is not a bad
idea, you also need to become comfortable trading with the
ambiguity inherent with specific events.
Our focus in this book is finding situations where we can do better
than the consensus. This is covered in detail in Chapter Five. The
ease of finding and manipulating financial data has considerably
lessened the efficacy of forecasting volatility using time series
models (the primary forecasting tool used in Volatility Trading
[Sinclair, 2013]), but measuring and forecasting volatility in this way is still necessary for trade sizing and allocation.
Econometricians are still writing endless papers about different
members of the GARCH family, but there have been no
fundamentally different advances in volatility measurement and
forecasting in the last 20 years.
For a summary of these time-series methods refer to Sinclair
Positional Option Trading (Wiley Trading) Page 6
