empirical asset pricing via machine learning
TRANSCRIPT
Empirical Asset Pricing via Machine Learning
Shihao Gu1 Bryan Kelly2 Dacheng Xiu1
1University of Chicago Booth School of Business
2Yale School of Management
New Methods for the Cross Section of ReturnsSept. 2018, University of Chicago
The HypeThe Intellectually Honest Answer
4
CNBC, July 12, 2017
Bloomberg, July 15, 2017
Bloomberg, Aug 9, 2017
Same Reporter3 weeks later
Economist, May 2017
Economist, Dec 2017
The Hype
The Intellectually Honest Answer
4
CNBC, July 12, 2017
Bloomberg, July 15, 2017
Bloomberg, Aug 9, 2017
Same Reporter3 weeks later
Economist, May 2017
Economist, Dec 2017
Despite this excitement, understanding of role/value
of these methods in asset pricing setting is barely nascent
What We Do
I Comparative analysis of machine learning methods in context of perhaps most widely
studied problem in finance, measuring equity risk premia
Clearest way to understand the relevance of ML for AP is to apply methods and
compare performance in familiar empirical problem
Two Primary Contributions
1. New benchmark of accuracy in measuring risk premia (aggregate and individual asset)
I Unprecedented out-of-sample predictive R2
I Strategies that leverage ML forecasts earn Sharpe ratios >2.0
2. Synthesize empirical AP with field of machine learning
I Show ML well positioned to push frontier of risk premium measurementI We provide comparative overview of ML methods applied to the two canonical problems of
empirical asset pricing: predicting returns in the cross section and time series
What We Do
I Comparative analysis of machine learning methods in context of perhaps most widely
studied problem in finance, measuring equity risk premia
Clearest way to understand the relevance of ML for AP is to apply methods and
compare performance in familiar empirical problem
Two Primary Contributions
1. New benchmark of accuracy in measuring risk premia (aggregate and individual asset)
I Unprecedented out-of-sample predictive R2
I Strategies that leverage ML forecasts earn Sharpe ratios >2.0
2. Synthesize empirical AP with field of machine learning
I Show ML well positioned to push frontier of risk premium measurementI We provide comparative overview of ML methods applied to the two canonical problems of
empirical asset pricing: predicting returns in the cross section and time series
What We Do
I Comparative analysis of machine learning methods in context of perhaps most widely
studied problem in finance, measuring equity risk premia
Clearest way to understand the relevance of ML for AP is to apply methods and
compare performance in familiar empirical problem
Two Primary Contributions
1. New benchmark of accuracy in measuring risk premia (aggregate and individual asset)
I Unprecedented out-of-sample predictive R2
I Strategies that leverage ML forecasts earn Sharpe ratios >2.0
2. Synthesize empirical AP with field of machine learning
I Show ML well positioned to push frontier of risk premium measurementI We provide comparative overview of ML methods applied to the two canonical problems of
empirical asset pricing: predicting returns in the cross section and time series
Return prediction is economically meaningful.
Fundamental goal of asset pricing is to understand the behavior of risk premia. If expected
returns were perfectly observed, we would still need theories to explain their behavior and
empirical analysis to test those theories
Efficient markets.
Risk premia are notoriously difficult to measure
— market efficiency forces return variation to be dominated by unforecastable news
— our empirical exercise is non-trivial because of the extremely low signal-to-noise ratio
What is Machine Learning?
Definition of “machine learning” is inchoate and often context specific.
Our definition:
i. a diverse collection of high-dimensional models for statistical prediction
+
ii. regularization methods for model selection and mitigation of overfit
+
iii. efficient algorithms for searching among a vast number of potential model specifications
Why Apply Machine Learning to Asset Pricing?
Reason 1: Measuring an asset’s risk premium is fundamentally a problem of prediction
Fama Nobel Lecture: Two pillars of empirical asset pricing research
1. Describe/understand differences in risk premis across assets
2. Describe/understand dynamics of the market equity risk premium
A risk premium is a conditional expectation of a future realized excess return.
ML methods specialize in prediction tasks, thus ideally suited to risk premium measurement.
Why Apply Machine Learning to Asset Pricing?
Reason 2: The collection of candidate conditioning variables for the risk premium is large
I We’ve accumulated a staggering list of return predictors
I They are often close cousins and highly correlated
Traditional prediction methods break down when predictor count approaches the observation
count and/or predictors are highly correlated
I With its emphasis on variable selection and dimension reduction, ML well suited for
such challenging empirical issues
Why Apply Machine Learning to Asset Pricing?
Reason 3: Functional form is unknown and likely complex
I Theoretical literature offers little guidance for winnowing list of conditioning variables and
functional forms
Three aspects of ML suited to problems of ambiguous functional form
1. Suite of dissimilar methods. Casts wide net in model search
2. Nonparametric design to approximate complex nonlinear associations
3. Parameter penalization and conservative model selection criteria complement complexity,
help avoid overfit and false discovery
The (Very Familiar) Empirical Setting
∼100 stock characteristics (usual suspects)
+
∼10 macroeconomic predictors (a la Goyal-Welch)
⇓
Monthly returns on 1) individual stocks and 2) stock portfolios
Which Machine Learning Methods?
I Linear ModelsI OLS(3) includes value, size, momentumI OLS + Elastic Net + Huber’s Loss
I Dimension ReductionI PCA, PLS
I Generalized Linear ModelsI Additive Series Regression
I Regression TreesI Random ForestI Gradient Boosted Regression Trees
I Neural Networks aka “Deep Learning”I up to 5 hidden layersI around 30,000 parameters
Main Empirical Findings
Machine learning holds promise for empirical asset pricing
1. Vast predictor sets viable in linear prediction when penalization used
2. Non-linearities substantially improve predictions
3. Shallow learning outperforms deeper learning
4. Distance between non-linear methods and benchmark widens when predicting portfolios
5. Gains from machine learning forecasts are economically large
6. Most successful predictors: price trends, liquidity, and volatility
Data and Over-arching Model
I Goal: Find best representation of Et(ri,t+1)
I We consider general model
Et(ri,t+1) = g?(zi,t), zi,t = xt ⊗ ci,t ,
I xt vector of macro predictorsI ci,t vector of stock characteristicsI g?(·) functional form approximated by ML
I Our framework nests typical (time-varying) beta-pricing specification
ri,t+1 = Et(ri,t+1) + β′i,t(Ft+1 − Et(Ft+1)) + εi,t+1︸ ︷︷ ︸
Noise
, Et(ri,t+1) = β′i,tλt ,
Training, Validation, and TestingWe divide the 60 years of data into
I 18 years of training sampleI 12 years of validation sampleI and remaining 30 years for out-of-sample testing.
Empirical Assessments
I Predictive performance in statistical terms
I R2OOS = 1 −
∑(i,t)∈T3
(ri,t+1−ri,t+1)2∑
(i,t)∈T3(ri,t+1−0)2
I Predictive performance in economic terms
I Sharpe ratio
I Model comparison
I Diebold-Mariano tests
I Variable importance
I Decrease in R2 from exclusion
I Benchmark predictive model (“OLS-3”)
I Linear model using size, value, and momentum
Individual Stock Return Prediction: Monthly
OLS OLS-3 PLS PCR ENet GLM RF GBRT NN1 NN2 NN3 NN4 NN5
All -4.60 0.16 0.18 0.28 0.09 0.19 0.27 0.30 0.35 0.38 0.39 0.37 0.35
Top 1000 -14.21 0.15 -0.10 -0.05 0.10 0.17 0.62 0.53 0.44 0.58 0.72 0.67 0.69
Bottom 1000 -2.13 0.37 0.29 0.36 0.18 0.28 0.29 0.27 0.41 0.45 0.46 0.42 0.40
OLS-3+H
PLSPC
REN
et+H
GLM
+H
RF
GBRT+H
NN
1N
N2
NN
3N
N4
NN
5
R2 oos
-0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8All
Top
Bottom
Similar for big/small stocks, monthly/annual returns
“Bottom-up” Prediction of Portfolio Returns
Predict portfolio returns by aggregating individual stock return predictions
Given weights wpi,t in portfolio p, and individual return predictions ri,t+1,
rpt+1 =N∑i=1
wpi,t × ri,t+1
Predicting Pre-specified PortfoliosMonthly R2
OLS-3 PLS PCR ENet GLM RF GBRT NN1 NN2 NN3 NN4 NN5
S&P 500 -0.11 -0.86 -2.62 -0.38 0.86 1.39 1.13 0.84 0.96 1.80 1.46 1.60
Big Growth 0.41 0.75 -0.77 -1.55 0.73 0.99 0.80 0.70 0.32 1.67 1.42 1.40
Big Value -1.05 -1.88 -3.14 -0.03 0.70 1.41 1.04 0.78 1.20 1.57 1.17 1.42
Small Growth 0.35 1.54 0.72 -0.03 0.95 0.54 0.62 1.68 1.26 1.48 1.53 1.44
Small Value -0.06 0.40 -0.12 -0.57 0.02 0.71 0.90 0.00 0.47 0.46 0.41 0.53
Big Conservative -0.24 -0.17 -1.97 0.19 0.69 0.96 0.78 1.08 0.67 1.68 1.46 1.56
Big Aggressive -0.12 -0.77 -2.00 -0.91 0.68 1.83 1.45 1.14 1.65 1.87 1.55 1.69
Small Conservative 0.02 0.75 0.48 -0.46 0.55 0.59 0.60 0.94 0.91 0.93 0.99 0.88
Small Aggressive 0.14 0.97 0.06 -0.54 0.19 0.86 1.04 0.25 0.66 0.75 0.67 0.79
Big Robust -0.58 -0.22 -2.89 -0.27 1.54 1.41 0.70 0.60 0.84 1.14 1.05 1.21
Big Weak -0.24 -1.47 -1.95 -0.40 -0.26 0.67 0.83 0.24 0.60 1.21 0.95 1.07
Small Robust -0.77 0.77 0.18 -0.32 0.41 0.27 -0.06 -0.06 -0.02 0.06 0.13 0.15
Small Weak 0.02 0.32 -0.28 -0.25 0.17 0.90 1.31 0.84 0.85 1.09 0.96 1.08
Big Up -1.53 -2.54 -3.93 -0.21 0.40 1.12 0.68 0.46 0.85 1.28 0.99 1.05
Big Down -0.10 -1.20 -2.05 -0.26 0.36 1.09 0.77 0.48 0.89 1.34 1.17 1.36
Small Up -0.79 0.42 -0.36 -0.33 -0.33 0.31 0.40 0.23 0.60 0.67 0.55 0.61
Small Down 0.40 1.16 0.47 -0.46 0.62 0.93 1.20 0.80 0.97 0.97 0.97 0.96
Variable ImportanceNN5
NN4
NN3
NN2
NN1
GBRT+
HRF
GLM
+H
ENet
+HPCR
PLS
mom
1mm
vel1
chm
omm
axre
t
mom
12m
indm
omdo
lvol
secu
redi
nd spre
tvol
turn
ninc
rid
iovo
l
mom
36m
std_
turn
basp
readag
rrd
_mve
mom
6m ep dyco
nvin
d ps illch
csho rd
depr
zero
trade
beta
age
beta
sqor
gcap
cash
debt lgr
bmlev
cash
prch
inv
inve
stbm
_ia
rd_s
ale
roic
sale
inv
oper
prof
roav
olegr
priced
elayms
herf
cfp
sgr
hire
cash
roaq
sale
rec
sale
cash
mve
_ia
gma
grca
pxcu
rrat
absa
ccacc
quick
roeq
pchc
apx_
iagr
ltnoa
cfp_
iata
ngse
cure
dpc
tacc
std_
dolvol tb
chem
pia
aeav
olch
pmia
chtx
rsup
pchs
ale_
pchi
nvt
cinv
est
real
esta
teno
ise5
pchs
ale_
pchx
sga
pchs
alei
nv
pchg
m_p
chsa
lesic2
pchd
epr
divo
pchq
uickear
chat
oia
noise4
stda
ccst
dcf
divi
pchc
urra
tno
ise1
pchs
ale_
pchr
ect
noise3
noise2sin
Marginal Relation: Characteristics and Et(ri ,t+1)
mom1m
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
×10-3
-5
0
5
ENet+H
GLM+H
GBRT+H
RF
NN3
retvol
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
×10-3
-5
0
5
mvel1
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
×10-3
-5
0
5
acc
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
×10-3
-5
0
5
Machine Learning Long-Short PortfoliosOLS-3+H PLS PCR
Pred Avg Std SR Pred Avg Std SR Pred Avg Std SR
Low -0.41 0.19 6.52 0.10 -0.85 -0.05 6.73 -0.03 -0.92 -0.49 7.02 -0.242 -0.08 0.40 5.28 0.26 -0.26 0.31 6.02 0.18 -0.29 0.15 6.20 0.089 1.44 0.76 7.10 0.37 1.50 1.10 5.34 0.71 1.45 1.34 5.44 0.85High 1.81 1.84 8.52 0.75 2.15 1.51 5.79 0.90 2.09 1.91 6.08 1.09
H-L 2.22 1.65 6.41 0.89 3.01 1.56 4.45 1.22 3.01 2.40 4.65 1.79
Enet+H GLM+H RFLow 0.06 -0.31 7.01 -0.15 -0.43 -0.51 6.94 -0.25 0.26 -0.33 7.13 -0.162 0.33 0.39 6.03 0.22 0.02 0.33 6.08 0.19 0.41 0.32 5.68 0.199 1.34 1.03 5.92 0.60 1.37 1.29 5.59 0.80 0.89 1.17 5.64 0.72High 1.65 1.80 7.31 0.86 1.84 1.64 6.34 0.90 1.07 1.83 6.78 0.93
H-L 1.59 2.12 5.48 1.34 2.27 2.15 4.39 1.70 0.81 2.16 5.34 1.40
GBRT+H NN1 NN2Low -0.03 -0.38 6.67 -0.20 -0.47 -0.80 7.47 -0.37 -0.37 -0.82 7.97 -0.362 0.16 0.43 5.66 0.26 0.14 0.21 6.24 0.12 0.19 0.17 6.44 0.099 0.81 1.11 5.26 0.73 1.64 1.21 5.49 0.76 1.40 1.13 5.46 0.72High 1.02 1.70 6.57 0.90 2.46 2.13 7.30 1.01 2.32 2.36 8.03 1.02
H-L 1.04 2.08 4.25 1.70 2.93 2.93 4.81 2.11 2.69 3.18 4.90 2.25
NN3 NN4 NN5Low -0.39 -0.96 7.77 -0.43 -0.28 -0.90 7.87 -0.40 -0.21 -0.76 7.93 -0.332 0.17 0.13 6.42 0.07 0.25 0.18 6.57 0.09 0.25 0.24 6.58 0.139 1.44 1.16 5.50 0.73 1.32 1.22 5.60 0.75 1.30 1.24 5.54 0.77High 2.30 2.23 7.78 0.99 2.28 2.35 7.95 1.02 2.19 2.21 7.78 0.98
H-L 2.69 3.19 4.77 2.32 2.56 3.25 4.79 2.35 2.39 2.97 5.05 2.03
Machine Learning (Value-Weighted) Long-Short PortfoliosOLS-3+H PLS PCR
Pred Avg Std SR Pred Avg Std SR Pred Avg Std SR
Low -0.42 0.39 5.22 0.26 -0.86 0.27 5.57 0.17 -0.90 0.04 5.92 0.032 -0.08 0.60 4.47 0.46 -0.27 0.49 5.10 0.33 -0.29 0.43 5.33 0.289 1.42 0.56 7.50 0.26 1.49 0.94 5.01 0.65 1.47 1.18 4.98 0.82High 1.72 0.90 8.18 0.38 2.03 0.96 5.45 0.61 2.02 1.36 5.61 0.84
H-L 2.14 0.51 6.46 0.27 2.89 0.70 4.35 0.56 2.92 1.32 4.72 0.97
ENet+H GLM+H RFLow 0.05 0.08 5.64 0.05 -0.42 0.11 5.43 0.07 0.28 0.09 6.09 0.052 0.33 0.52 5.07 0.35 0.02 0.46 4.67 0.34 0.41 0.39 5.17 0.269 1.33 0.88 5.59 0.55 1.36 0.97 5.49 0.61 0.89 1.20 5.88 0.70High 1.59 0.80 6.83 0.40 1.76 1.18 6.30 0.65 1.01 1.49 7.18 0.72
H-L 1.54 0.72 5.49 0.45 2.17 1.08 4.52 0.83 0.73 1.40 5.54 0.87
GBRT+H NN1 NN2Low 0.00 0.03 5.76 0.02 -0.40 -0.37 7.16 -0.18 -0.30 -0.50 7.89 -0.222 0.16 0.50 5.00 0.34 0.15 0.40 6.03 0.23 0.18 0.36 6.13 0.219 0.81 0.99 5.08 0.67 1.60 0.94 5.09 0.64 1.40 1.01 5.52 0.63High 0.97 1.20 5.81 0.71 2.18 1.37 6.31 0.75 2.03 1.43 6.95 0.72
H-L 0.97 1.16 4.27 0.94 2.58 1.73 5.62 1.07 2.32 1.94 5.68 1.18
NN3 NN4 NN5Low -0.21 -0.51 7.83 -0.23 -0.29 -0.43 7.74 -0.19 -0.15 -0.36 7.63 -0.162 0.26 0.32 6.39 0.18 0.20 0.39 6.15 0.22 0.26 0.29 6.36 0.169 1.28 1.20 5.79 0.72 1.36 1.07 5.87 0.63 1.26 1.31 5.77 0.79High 1.99 1.58 7.33 0.74 2.02 1.47 7.11 0.72 1.91 1.55 6.90 0.78
H-L 2.20 2.09 5.78 1.25 2.30 1.90 5.83 1.13 2.06 1.91 6.01 1.10
Conclusion
Machine learning holds promise for empirical asset pricing
1. Vast predictor sets viable in linear prediction when penalization used
2. Non-linearities substantially improve predictions
3. Shallow learning outperforms deeper learning
4. Distance between non-linear methods and benchmark widens when predicting portfolios
5. Gains from machine learning forecasts are economically large
6. Most successful predictors: price trends, liquidity, and volatility
Can Machines Learn Finance?
2011: Google Brain launches. Uncertain if deep neural
networks would identify a cat, let alone drive a car
Answer: Yes, but there is much more to learn
I Most anecdotes in low capacity settings (HFT, OTC, etc.)
I We provide first large scale evidence of value for long-term asset management
I These are early days (2011 cat recognition)
Can Machines Learn Finance?
2011: Google Brain launches. Uncertain if deep neural
networks would identify a cat, let alone drive a car
Answer: Yes, but there is much more to learn
I Most anecdotes in low capacity settings (HFT, OTC, etc.)
I We provide first large scale evidence of value for long-term asset management
I These are early days (2011 cat recognition)
Comparing Predictions: Diebold-Mariano TestsPositive numbers = column model outperforms row model
OLS-3 PLS PCR ENet GLM RF GBRT NN1 NN2 NN3 NN4 NN5
+H +H +H +H
OLS+H 3.81 3.82 3.85 3.81 3.83 3.91 3.94 3.96 3.96 3.98 3.97 3.96
OLS-3+H 0.23 1.72 -0.80 0.63 1.55 1.93 1.98 2.83 3.01 2.61 2.63
PLS 1.58 -0.71 0.08 1.39 1.61 1.52 2.29 2.43 2.18 2.15
PCR -1.51 -1.62 0.06 0.48 0.54 1.13 1.20 0.94 0.85
ENet+H 1.00 1.59 1.79 2.09 2.02 2.19 1.92 1.94
GLM+H 1.21 1.59 1.70 2.55 2.76 2.44 2.33
RF 0.66 0.66 1.12 1.30 0.94 0.90
GBRT+H 0.24 0.73 0.83 0.53 0.46
NN1 0.87 1.11 0.49 0.31
NN2 0.10 -1.09 -1.20
NN3 -1.03 -1.92
NN4 -0.47
Predicting Pre-specified Portfolios (Annual R2)
OLS-3 PLS PCR ENet GLM RF GBRT NN1 NN2 NN3 NN4 NN5
S&P 500 -3.31 0.43 -7.17 0.26 2.07 8.80 7.28 9.99 12.02 15.68 15.30 13.15
Big Growth 3.36 4.88 -4.04 3.62 0.49 9.50 5.86 8.76 8.54 12.42 9.95 7.56
Big Value -11.82 -6.92 -10.22 -2.13 2.44 7.14 6.93 7.47 11.06 11.67 13.37 10.03
Small Growth 6.11 10.81 8.94 8.41 4.31 8.05 3.75 7.24 6.37 7.48 6.60 4.81
Small Value 4.25 2.87 3.19 0.21 0.03 6.20 2.13 3.96 5.52 6.84 2.60 7.23
Big Conservative -8.34 -2.42 -9.77 -3.77 5.17 8.44 5.26 -1.31 8.64 9.65 12.47 6.09
Big Aggressive -0.92 1.89 -4.72 1.36 2.00 7.42 6.67 11.00 11.74 13.08 11.27 10.67
Small Conservative 1.30 6.36 5.01 3.19 2.35 4.60 0.62 5.31 5.39 5.97 4.22 4.71
Small Aggressive 5.53 5.12 2.88 1.04 0.37 6.43 3.23 2.50 4.50 5.50 1.47 6.56
Big Robust -7.17 -2.55 -9.18 1.33 5.42 7.61 6.60 12.55 12.04 13.92 15.29 13.35
Big Weak -1.81 3.09 -7.15 -1.02 -1.12 9.62 7.62 4.41 9.95 11.39 11.73 8.40
Small Robust -2.33 0.93 -0.20 0.76 3.72 0.41 -0.87 2.92 3.67 4.47 0.86 4.19
Small Weak 4.72 9.89 5.68 2.15 -1.11 7.53 3.10 -0.48 1.53 2.96 1.61 1.08
Big Up -24.02 -11.77 -19.16 -5.11 0.52 6.15 6.21 4.26 11.44 11.11 14.48 10.62
Big Down -2.32 0.39 -2.79 -0.15 0.71 7.64 5.53 3.58 8.78 9.54 10.32 6.79
Small Up -5.47 3.82 0.71 -2.83 1.57 1.84 -0.19 -4.22 0.70 1.12 -1.42 2.83
Small Down 4.72 5.59 4.84 2.87 0.50 7.23 3.49 3.24 4.63 5.90 3.28 5.22
Cumulative Returns of Long-Short ML Portfolios
1987 1990 1993 1996 1999 2002 2005 2008 2011 2014 2016
S
ho
rt P
osi
tio
n
Lo
ng
Po
siti
on
4
2
0
2
4
6
8
OLS-3+H PLS PCR ENet+H GLM+H RF GBRT+H NN3 SP500-Rf solid = long dash = short
Drawdowns, Turnover, and Risk Adjusted Alpha
OLS-3 PLS PCR ENet GLM RF GBRT NN1 NN2 NN3 NN4 NN5 MOM1m
+H +H +H +H
Drawdowns and Turnover
Max DD 64.74 35.51 34.17 35.25 27.05 50.22 35.24 19.70 23.72 14.84 18.66 21.73 68.91
Max 1M Loss 38.69 25.05 22.22 34.11 16.92 34.94 22.87 16.98 23.72 10.15 18.66 21.71 43.67
Turnover 156.78 76.92 106.79 143.61 129.22 113.91 136.82 110.87 112.08 113.07 114.08 113.57 172.56
Risk-adjusted Performance
Mean Ret. 1.65 1.56 2.40 2.12 2.15 2.16 2.08 2.93 3.18 3.19 3.25 2.97 1.80
FF3 α 1.43 1.39 2.44 1.96 2.15 2.09 2.08 2.89 3.20 3.20 3.24 2.98 1.54
R2 4.92 12.75 11.59 4.35 3.85 8.64 0.31 8.89 7.28 8.45 8.34 8.38 5.92
FF5 α 1.64 0.96 2.01 1.58 1.74 1.82 1.95 2.66 2.97 3.00 3.03 2.73 1.88
R2 7.05 25.10 23.50 10.96 15.28 14.97 4.09 12.59 10.03 10.74 11.11 11.43 10.60
FF5+Mom α 1.88 0.82 1.76 1.34 1.54 1.75 1.78 2.62 2.93 2.98 2.95 2.68 2.14
R2 17.46 32.91 43.77 26.27 31.80 16.59 16.81 13.19 10.59 10.88 13.61 12.25 20.53
Note: All numbers are in percentage.
Which Covariates Predict Returns?
PLS PCR ENet+H GLM+H
levoperprof
retvolchcsho
nincrcashpr
agrrd_mve
dolvolep
mvel1mom6m
spturn
maxretstd_turn
mom12mindmomchmom
mom1m
0.0 0.1 0.2 0.3mom36m
bm_iachinv
lgrbm
deprcashpr
mom6mep
investagr
rd_mvechcshomvel1
spmaxret
indmommom12m
chmommom1m
0.0 0.1 0.2 0.3mom36m
mom6mretvolmvel1
epturn
chinvchmom
nincrps
std_turnsp
chcshoinvest
rd_mvedolvol
agrindmom
mom12mmom1m
0.0 0.2 0.4 0.6mom36m
securedindsic2
chinvlgrep
turnchmom
dolvolcashprchcsho
illinvest
agrrd_mvemaxretmvel1
indmommom12m
mom1m
0.0 0.1 0.2 0.3 0.4 0.5
RF GBRT+H NN2 NN3
mom36mdolvol
illbetasq
betasp
idiovolconvindmom6m
baspreadmom12m
chmomretvolnincr
securedindindmommaxretmvel1
dymom1m
0.00 0.05 0.10rd_mve
turnmom36m
agebeta
mom6mbaspread
idiovolsp
convindchmom
mvel1indmom
retvolnincr
maxretmom12m
securedindmom1m
dy
0.00 0.05 0.10 0.15 0.20betasq
spmom36m
securedindzerotrade
nincrindmomstd_turn
idiovolill
baspreadmom6m
mom12mturn
dolvolmaxretchmom
retvolmvel1
mom1m
0.00 0.05 0.10 0.15 0.20beta
spsecuredind
mom36mzerotrade
nincrindmomstd_turn
illmom12mbaspread
mom6midiovol
turndolvol
chmommaxret
retvolmvel1
mom1m
0.00 0.05 0.10 0.15 0.20
Characteristic Importance over Time by NN319
87198819
89199019
91199219
93199419
95199619
97199819
99200020
01200220
03200420
05200620
07200820
09201020
11201220
13201420
152016
mom
1mm
vel1
mom
12m
chm
omm
axre
t
indm
omre
tvol
dolvol sp
turnagr
ninc
rrd
_mve
std_
turn
mom
6m
mom
36m ep
chcs
ho
secu
redi
ndid
iovo
l
basp
read ill
age
conv
ind rd
depr
beta
beta
sqca
shpr ps
zero
trade dy
orgc
apbm lgr
cash
debt
chin
vin
vest lev
oper
prof
bm_i
asa
lein
veg
rcf
prd
_sal
esg
rro
aqroic
sic2
mve
_ia
ms
quic
khe
rfhi
re
pric
edel
aysa
lere
cro
avol
roeq
grca
pxcu
rrat
cash
std_
dolvol
acc
cfp_
iagr
ltnoa
gma
pcta
ccab
sacc
sale
cash
secu
red
pchd
epr
tang
pchc
apx_
iach
empi
aea
r
pchs
ale_
pchi
nvt
pchs
alei
nvch
txch
pmia
chat
oia tb
aeav
olrs
up
pchg
m_p
chsa
le
pchs
ale_
pchx
sga
cinv
est
pchq
uick
pchs
ale_
pchr
ect
real
esta
te
pchc
urra
tst
dacc
stdc
fdi
vidi
vosin