Transcript Business Cycles, Macro Variables, and Stock Market Returns

Business Cycles, Macro
Variables, and Stock Market
Returns
William Carter, David Nawrocki,
and Tonis Vaga
Agenda
 Introduction and literature review (Jon)
 Relationships between real activity and




stock returns (Jordan)
Multiple phases of the business cycle
(Danielle)
Linear regression analysis (Dmitry)
Application of neural network (Raegen)
Conclusion (Jon)
Introduction
 Business cycle indicators: relevant issue
 Chen, Roll, and Ross; Fama and French; and
Schwert
 Risk premium embedded in expected returns
moves inversely with business conditions
 Whitelaw
 Conditional returns and conditional volatility
change over time with changes in the cycle
 Nawrocki and Chauvet
 Find dynamic relationship between stock market
fluctuations and cycles
Intro. Con’t.
 Perez-Quiros and Timmerman
 Asymmetries in conditional mean and volatility of
excess stock returns around cycle turning points
 Chauvet and Porter
 Suggest non-linear risk measure that allows riskreturn relationship to not be constant over Markov
states
 DeStafano
 Tests four-state model of cycle and dividend
discount model to provide evidence that expected
stock returns vary inversely with economic
conditions
This all suggests…
 Nonlinear financial market dynamic
 Thus requiring a nonlinear methodology
 Between business cycle and stock market
 DeStafano (2004)
 Arbitrarily defined four phases
 Period between NBER peaks and troughs into two
equal periods
Where the authors differ
 Utilizes simple linear models
 Looks for phase transitions
 Provides preliminary definitions of phases
 Then used in the neural network methodology for final
estimates of the phases
 Independent of NBER peaks and troughs
 Not announced until 9-18 months after the
fact
Multiple Phases of the Business
Cycle
 Chauvet and Potter (1998) and Perez-Quiros and
Timmermann (2000) study two phases: expansions and
recessions
 Consistent with the NBER’s definition of business cycle peaks
and troughs
 Chauvet and Potter (1998) note changes in conditional
means and variances well before the peak and trough,
suggesting additional phases of the business cycle
 Four/five-stage models have been proposed by Hunt
(1987), Stovall (1996), DeStefano (2004), Guidolin and
Timmermann (2005), Guidolin and Ono (2006)
Advantages of the Neural Network
 Eliminates problems from traditional
approaches




Linearity assumptions
Data-pooling issues
Data mining
Pre-specification of the model
Relationships between Real Activity
and Stock Returns
 Prior Research:
 Moore (1976) and Sherman (1986) found certain economic
indicators are leading indicators for the business cycle and
security markets
 Chen, Roll, Ross (1986) modeled equity returns using
macroeconomic factors:
 Industrial Production
 Monetary Aggregates
 Debt Market Yields
 Fama & French (1989) measured stock return volatility using the
relationship between returns and real activity
Skewness
 Skewness and volatility has also been tied to the
business cycle
 Schwert (1989) finds stock market volatility increases
during recessions
 Other research has found high variability in the
skewness of stock returns and that it varies
systematically with business conditions
 Skewness becomes more negative during expansions
and less negative or positive during contractions
Prior Research
 Whitelaw (1994) finds that the relationship
between the conditional mean and volatility of
stock returns is nonstationary
 Using a linear relationship between mean and
volatility can lead to incorrect results from GARCH
and ARCH models
 Utilizing a Nonlinear Markov switching
regression:
 Volatility increases during recessions
 Conditional means rise before the end of recessions
 Conditional means decrease before the peak of
expansions
 Sharpe ratios are negative in troughs, positive in
peaks
Prior Research
 Whitelaw (1994) et al. find conditional variance is
countercyclical
 Fama and French (1989) et al. find conditional means
move with the business cycle
 Rapach (2001) finds real stock returns are related to
changes in money supply, aggregate supply, aggregate
spending
 This research suggests that stock market phases are
related to economic fluctuations
Prior Research
 Recent research finds that the power of the
economic factors used for predictions varies
over time and volatility
 Small firms are shown to be strongly affected
during recessions
 Fundamental factors such as DDM are affected
by the business cycle
 Investors discount earnings using short term T-Bill
when the economy is slowing down
 Discount using long term T-Bond rate in the other
states of economy
Method
 Time-invariant forecasting models will not work under
sudden large changes in time series
 Previous research was determined using the NBER
cycle dates, which have a lag of 9 – 18 months
 The Markov switching VAR is used in this study along
with a neural network
 It does not require the form of the regression to be previously
specified
 Allows for a state switching nonlinear model that tests
the significance of the various macroeconomic variables
 The neural network must be provided with an initial set of dates
for the phases and macroeconomic variables for the transistions
Multiple Phases of the Business
Cycle
 Chauvet and Potter (1998) and Perez-Quiros and
Timmermann (2000) study two phases: expansions and
recessions
 Consistent with the NBER’s definition of business cycle peaks
and troughs
 Chauvet and Potter (1998) note changes in conditional
means and variances well before the peak and trough,
suggesting additional phases of the business cycle
 Four/five-stage models have been proposed by Hunt
(1987), Stovall (1996), DeStefano (2004), Guidolin and
Timmermann (2005), Guidolin and Ono (2006)
Stovall’s Business Cycle Phases
 Expansion in 3 phases:
 Recovery from recession – slow growth
 Economic growth picks up vigorously
 Inflation increases
 Recession in 2 phases:
 Decline in economic production
 Economy flattens out and begins to recover
 A simplistic model – Stovall uses the time period
between NBER peaks and troughs, divides each time
period evenly into three and two periods
 Finds that certain sectors perform well during certain
stages
Hunt’s Business Cycle Phases
 Hunt suggests economic variables that drive the transition between phases
 Easeoff




Industrial production slows
Initial unemployment claims increase
Non-farm payrolls turn down
University of Michigan Consumer Sentiment index falls
 Plunge
 Federal Funds rate decreases
 Real monetary base increases
 Interest rate spread narrows
 Revival
 Industrial production increases
 Initial unemployment claims fall
 Non-farm payrolls increase
 Acceleration
 Real monetary base increases
 Consumer Price Index rises
 Early Revival – transition between Plunge and Revival
Hunt’s Business Cycle Phases
 Implemented his model using 12-month rate of change
statistics, followed monthly
 One complete cycle measured from Easeoff to Easeoff
phase
 Each phase exhibited different investment behavior
 Easeoff had significant negative skewness
 Consistent with Alles and Kling’s (1994) finding that skewness
becomes strongly negative during contractions
 Plunge had insignificant skewness
 Revival had initial insignificant skewness, followed by positive
significant skewness
 Acceleration exhibited poor risk-return behavior (high inflation
period)
 Easeoff and revival exhibited the best risk-return behavior
Linear regression analyses
 Two regression analyses performed on monthly time
series for the period 1970-1997 to study relationships
between S&P 500 and variables
 Macroeconomic variables considered





CPI rate of change (CPIROC)
Industrial production rate of change (IP)
Spread between 90-days T-bill and 30-year T note (SPREAD)
Difference between AAA and BAA corporate bonds (AAA_BAA)
Rate of change in real adjusted monetary base lagged 4 month
(REAL_MB)
 Level of housing starts (STARTS)
 Level of manufacturing orders excluding aircraft and parts
(ORDERS)
Regression results for 1971-1997
 Industrial production, manufacturing orders, and housing starts are
significant at 10% confidence level
 The correlation between independent variables is quite low below 0.40.
Only two correlation coefficients were as high as 0.60
 Adjusted R2 below 0.0386 indicates little relationship between variables
Individual regression results for four
business cycle phases
Individual regression results for four
business cycle phases (cont.)
 Impact of variables changes through the phases
of the business cycle
 All of the phase regressions have higher
adjusted R2 compared to the base regression
 The four phase regressions exhibit different
significant independent variables both from each
other and the base regression
 Conclusion: strong support for the hypothesis
that S&P 500 has different phases
Studying Economic Phases with a
Neural Network
 What is a neural network?
 Mimics the structure of the brain. Output is produced
by interconnected nodes in a parallel fashion as
opposed to traditional sequential processing.
 This operation makes the NN more robust and
adaptable to fuzzy logic.
 Here, a neural network is used as computational
architecture to learn from past economic phases and
performance variables. And, then predict unseen
phases in the economy.
Studying Economic Phases with a
Neural Network
 Advantages of using a Neural Network
 Captures all relationships (linear and non)
 A pre-specified regression equation is not required
 This study uses a PNN
 PNN’s use estimated “probability functions to train the
network with a data set.”
 It is an adaptive PNN, meaning that an algorithm
determines a smoothing function for each variable.
The variables can be weighted and insignificant
variables eliminated.
Studying Economic Phases with a
Neural Network
Studying Economic Phases with a
Neural Network
 How it works
 The neural network was trained, using 1971 to 1988, to specify
the phase for the next year.
 After each 12 month period was added on the network retrained
 Testing the neural network
 Known economic phases for Dec 1989 through Dec 1997 were
compared to the neural network’s defined phases
 Linear and nonlinear models differ 37% of the time…indicating
that there is some nonlinear dynamic captured by the NN.
 “There are significant variables and processes in the S&P data
stream that are not strictly linear. Linear models can only
approximate the actual nonlinear process.”
Studying Economic Phases with a
Neural Network
…Since 1997
Summary and Conclusions
 Previous research
 Two market states in economy and US stock
market returns (S&P 500 index)
 Four, possibly five Markov states have been
identified in the business cycle
 Regression analysis and neural network
provide evidence of four distinct market states
 Supports empirical research that delineates 4-5
market states
Summary and Conclusion Con’t.
 Instead of a fundamental variable
approach using earnings and discount
rates (DeStafano)
 Macroeconomic variable approach was
used
 Real time approach
 Even though independent of NBER
 NBER peak occurs in Easeoff/Plunge phases
 NBER trough occurs in Plunge/Revival phases
Summary and Conclusion Con’t.
 This methodology closely corresponds to the
“growth cycle” methodology defined by
Geoffrey H. Moore
 Also supports studies that discovered
nonlinear relationships in financial markets
 Chauvet and Potter (1998)
 Perez-Quiros and Timmermann (2000)
 Echo LeBaron’s warning
 Results with nonlinear measures are not as robust
as results obtained from linear models
Step Back…..
 These different business cycles could be
used for the Coleman Fund
 To switch out of potentially underperforming
sectors
 QInsight has the economy in the plunge
phase
 In general, if these criteria were used we would be
invested in a slightly different combination of
sectors