Model Uncertainty in the Cross Section

Jiantao Huang (HKU) Ran Shi (Colorado Boulder)

Jan 5, 2025

AFA 2025, San Francisco

Motivation
Existing uncertainty measures regarding macroeconomy and asset markets
▶ VXO/VIX index (Bloom, 2009); Macro/real/financial uncertainty (Jurado, Ludvigson, and Ng, 2015;
Ludvigson, Ma, and Ng, 2021); Policy uncertainty (Baker, Bloom, and Davis, 2016); News implied
volatility (Manela and Moreira, 2017) . . .
▶ Time-varying and related to firms’ investment/production/hiring activities
▶ Focus on time-series dimension, e.g., the extent to which financial outcomes fluctuate
▶ Insufficient to capture uncertainty related to asset allocation decisions

Huang and Shi (2025) Model Uncertainty in the Cross Section Jan 5, 2025 1 / 17
Motivation
Existing uncertainty measures regarding macroeconomy and asset markets
▶ VXO/VIX index (Bloom, 2009); Macro/real/financial uncertainty (Jurado, Ludvigson, and Ng, 2015;
Ludvigson, Ma, and Ng, 2021); Policy uncertainty (Baker, Bloom, and Davis, 2016); News implied
volatility (Manela and Moreira, 2017) . . .
▶ Time-varying and related to firms’ investment/production/hiring activities
▶ Focus on time-series dimension, e.g., the extent to which financial outcomes fluctuate
▶ Insufficient to capture uncertainty related to asset allocation decisions

Our paper proposes cross-sectional uncertainty measure for Bayesian “investors”

▶ Asset allocation (e.g., value or momentum funds) =⇒ Need an asset pricing (AP) model
▶ Model uncertainty: Uncertainty regarding which AP model to use
▶ Two-model example: 99-1% (low uncertainty) vs. 50-50% (high uncertainty)
Huang and Shi (2025) Model Uncertainty in the Cross Section Jan 5, 2025 1 / 17
Main Findings

Model uncertainty is sizable: heightened model uncertainty coincides with bad times
▶ E.g., 2008 GFC, model uncertainty is maximized =⇒ A true AP model is elusive!

Model uncertainty shocks carry negative risk premium

Documents strong correlations between model uncertainty shocks and fund flows
▶ Heightened model uncertainty =⇒ persistent flows out of equity to government bonds
▶ Outflows from small-cap and actively managed funds
▶ NO such patterns for VIX & financial uncertainty

Huang and Shi (2025) Model Uncertainty in the Cross Section Jan 5, 2025 2 / 17
Econometric Theory
The framework: linear SDF models
1. N test assets, R ∈ RN , all excess returns
2. p factors, f ∈ Rp , all tradable long-short portfolios : f ⊆ R
3. Linear SDF model: E[R × m] = 0 in which

m = 1 − (f − E[f ])⊤ b

Huang and Shi (2025) Model Uncertainty in the Cross Section Jan 5, 2025 3 / 17
The framework: linear SDF models
1. N test assets, R ∈ RN , all excess returns
2. p factors, f ∈ Rp , all tradable long-short portfolios : f ⊆ R
3. Linear SDF model: E[R × m] = 0 in which

m = 1 − (f γ − E[f γ ])⊤ bγ

4. Model uncertainty: which elements of f determine E[R]?

▶ γj = 1: the j-th factor is included (bj ̸= 0)
▶ γj = 0: bj ≡ 0
iid
5. Data: D = {R1 , . . . , RT } ∼ N (µ, Σ)
6. Model: a restriction on this distribution through the moment condition

Under model Mγ : µ = Cov[R, f γ ]bγ

Huang and Shi (2025) Model Uncertainty in the Cross Section Jan 5, 2025 3 / 17
The framework: Bayesian inference, g-prior

1. Prior: a generalized version of Arnold Zellner’s g-prior for bγ :

Under model Mγ : bγ ∼ N 0, g · Var(bγ ) , g>0

▶ g: “confidence” in prior information

Huang and Shi (2025) Model Uncertainty in the Cross Section Jan 5, 2025 4 / 17
The framework: Bayesian inference, g-prior

1. Prior: a generalized version of Arnold Zellner’s g-prior for bγ :

Under model Mγ : bγ ∼ N 0, g · Var(bγ ) , g>0

▶ g: “confidence” in prior information

2. The likelihood function:

   
T g pγ
P[D | Mγ ] ∝ exp − SR2max − SR2 − log(1 + g)
2 1+g γ 2
▶ In-sample Sharpe ratio vs model dimensionality
▶ GRS test (Gibbons, Ross, and Shanken, 1989) intuition
▶ SR2max enters every model: Irrelevance of test assets (efficient GMM intuition)

Huang and Shi (2025) Model Uncertainty in the Cross Section Jan 5, 2025 4 / 17
Model uncertainty in the cross section: definition

Our cross-sectional uncertainty measure:

1
p log 2 ∑
E =− (log P[Mγ | D]) P[Mγ | D]
γ

1. entropy[Mγ | D] ∈ [0, 1]

2. minimized when D favors one dominant model

▶ for one dominant model to exist: large SR2γ with small pγ

3. maximized when D lends equal evidence across models

Huang and Shi (2025) Model Uncertainty in the Cross Section Jan 5, 2025 5 / 17
Posterior property: Pitfall of g-priors
Theorem
Assume that the observed return data are generated from a true linear SDF model Mγ0 . If
γ0 ̸= 0 (the SDF is not a constant) and f γ0 ⊂ f (the set of factors under consideration include all
true factors), under the g-prior specification with g ∈ (0, ∞), as T → ∞,
1 (factor selection consistency) if factor j belongs to the true model Mγ0 , the posterior
marginal probability of choosing it converges to one in probability:
p
P[γj = 1 | D] → 1;
2 (model selection inconsistency) the posterior probability of the true model will always be
strictly smaller than one, that is, P[Mγ0 | D] < 1 with probability one.

g-priors can identify true factors, at the cost of incorporating redundant ones.
Huang and Shi (2025) Model Uncertainty in the Cross Section Jan 5, 2025 6 / 17
Restoring model selection consistency: mixture of g-priors
Adapt mixture of g-priors proposed by Liang et al. (2008) into SDF models:
1
π (g) = , g>0
(1 + g)2
▶ g/ (1 + g): decides how posterior beliefs about the SDF should be updated

▶ g/(1 + g) ∼ Uniform(0, 1): agnostic regarding this posterior updating process

Bayes factor still has a closed-solution

▶ Remains the same economic tradeoff: increasing in SR2γ but decreasing in pγ

Huang and Shi (2025) Model Uncertainty in the Cross Section Jan 5, 2025 7 / 17
Restoring model selection consistency: mixture of g-priors
Adapt mixture of g-priors proposed by Liang et al. (2008) into SDF models:
1
π (g) = , g>0
(1 + g)2
▶ g/ (1 + g): decides how posterior beliefs about the SDF should be updated

▶ g/(1 + g) ∼ Uniform(0, 1): agnostic regarding this posterior updating process

Bayes factor still has a closed-solution

▶ Remains the same economic tradeoff: increasing in SR2γ but decreasing in pγ

Theorem
p
Under the mixture of g-priors specification, as T → ∞, P[Mγ0 | D] → 1.

Mixture of g-priors is to achieve posterior model selection consistency! Simulations

Huang and Shi (2025) Model Uncertainty in the Cross Section Jan 5, 2025 7 / 17
A misspecified set of factors
f 0 : the true set of factors, f 0 ⊈ f =⇒ the studied factors omit some pricing factors

Huang and Shi (2025) Model Uncertainty in the Cross Section Jan 5, 2025 8 / 17
A misspecified set of factors
f 0 : the true set of factors, f 0 ⊈ f =⇒ the studied factors omit some pricing factors

Theorem
Assume that the observed return data are generated from a true linear SDF m0 = 1 − (f 0 − E[f 0 ])⊤ b0 .
Let f γ0 = f 0 ∩ f ; that is, f γ0 is the subset of f that includes only the true pricing factors. As T → ∞,
p
1. for all j such that γ0,j = 1, P[γj = 1 | D] → 1;

2. model uncertainty measure E satisfies E ≤ (p − pγ0 )/p with probability one.

True factors in f can always be selected =⇒ factor selection consistency Simulations

If E ≈ 1, only two possibilities:

(1) pγ0 ≈ 0 (all factors under study are useless/no strong factors)
(2) Observed data entirely uninformative about the true SDF model
Huang and Shi (2025) Model Uncertainty in the Cross Section Jan 5, 2025 8 / 17
Empirics
Model uncertainty in US stock markets (14 factors) Data Examples

1.0

0.8
Model Uncertainty

0.6

0.4

0.2

0.0
1975-12 1980-12 1985-12 1990-12 1995-12 2000-12 2005-12 2010-12 2015-12 2020-12
Date
Entropy ∈ [0.27, 0.99]: Average (standard dev) = 0.70 (0.21)
Heightened model uncertainty coincides with economic downturns / stock market crashes

Model uncertainty hits upper bounds during 2008–2011 & 2018–2020

▶ Marginal probs of all factors ≈ 50% =⇒ Model/factor selection is elusive in these periods
Huang and Shi (2025) Model Uncertainty in the Cross Section Jan 5, 2025 9 / 17
Model uncertainty in the “Factor Zoo”

A zoo of 153 factors in Jensen, Kelly, and Pedersen (2023)

▶ Categorized into 13 clusters, e.g., accruals, investment, value, seasonality . . .

Yet, handling simultaneously all 153 factors is infeasible

▶ Theoretical property of our Bayesian approach based on small p and large T
▶ Comparing 2153 models computationally impracticable

A practical solution:
▶ Randomly select one factor in each of the 13 cluster themes
▶ Measure model uncertainty based on market factor + 13 randomly selected factors
▶ Repeat this exercise 1000 times

Huang and Shi (2025) Model Uncertainty in the Cross Section Jan 5, 2025 10 / 17
Model uncertainty in the “Factor Zoo” – Continued
1.0

0.8
Model Uncertainty

0.6

0.4

0.2

0.0
1975-12 1980-12 1985-12 1990-12 1995-12 2000-12 2005-12 2010-12 2015-12 2020-12
Date
Even in this exercise, model uncertainty measure
▶ Is strongly time-varying, displaying clear business-cycle patterns
▶ Has stayed at a very high level since the 2008 GFC
Huang and Shi (2025) Model Uncertainty in the Cross Section Jan 5, 2025 11 / 17
Is Model Uncertainty Priced? Robustness Check

Table 1: Risk Premia of Model Uncertainty Shocks: Monthly Estimates

Number of PCs: 5 6 7 8 9 10
λE -0.066*** -0.067*** -0.067*** -0.065*** -0.062*** -0.060***
s.e. 0.017 0.017 0.017 0.017 0.018 0.018
Time-series R2 5.8% 5.8% 5.8% 6.2% 6.2% 6.2%
The table reports the risk premia estimates of model uncertainty shocks (Etar1 ) based on the three-pass method of Giglio and Xiu (2021). In all estimations, we
standardize Etar1 to have a unit variance. In particular, we project Etar1 onto the space of large PCs of 275 Fama-French characteristic-sorted portfolios in the US
market. The number of latent factors ranges from five to 10. If the 90% (95%, 99%) confidence interval of the risk premium does not contain zero, the risk premium
estimate will be highlighted by * (**, ***). We also report the time-series fit in each panel. Sample: 1975/07 - 2020/12.

Investors willing to pay a premium to hedge against heightened model uncertainty

Equate standard deviation of model uncertainty shocks to the market:
√
Annualized risk premium equals ( 12)λE × 17% ≈ −4%
Huang and Shi (2025) Model Uncertainty in the Cross Section Jan 5, 2025 12 / 17
Model Uncertainty and Mutual Fund Flows

Study the dynamic responses of fund flows to uncertainty shocks using VAR:

Y t = B0 + B1 Y t−1 + · · · + Bl Y t−l + controlst + Sϵt

l denotes the lag order and is chosen by AIC/BIC (equal to 1)

Use Cholesky decomposition to identify the dynamic responses (S):

▶ exogenous cause: uncertainty measure as the first element in Y t
▶ propagating mechanism: uncertainty measure as the last element

Huang and Shi (2025) Model Uncertainty in the Cross Section Jan 5, 2025 13 / 17
Relationships with aggregate fund flows

0.0100

0.060
Exogeneous Model Uncertainty
Endogeneous Model Uncertainty
−0.0225

0.035
Impulse Responses

Impulse Responses
−0.0550

0.010
−0.0875

−0.015
Exogeneous Model Uncertainty
−0.1200

−0.040
Endogeneous Model Uncertainty

1 6 11 16 21 26 31 36 1 6 11 16 21 26 31 36

Months Ahead Months Ahead

Equity Fund Flows Fixed-Income Fund Flows

Huang and Shi (2025) Model Uncertainty in the Cross Section Jan 5, 2025 14 / 17
Fixed-income fund flows with different investment objective codes
0.130

0.050
0.05
Exogeneous Model Uncertainty Exogeneous Model Uncertainty Exogeneous Model Uncertainty
Endogeneous Model Uncertainty Endogeneous Model Uncertainty Endogeneous Model Uncertainty
0.095

0.025
0.03
Impulse Responses

Impulse Responses

Impulse Responses
0.060

0.000
0.01

−0.025
−0.01
0.025
−0.010

−0.050
−0.03
1 6 11 16 21 26 31 36 1 6 11 16 21 26 31 36 1 6 11 16 21 26 31 36

Months Ahead Months Ahead Months Ahead

Government bonds Money markets Corporate bonds

Following high model uncertainty, sharp dynamic inflows to government bond funds
No such patterns in money market or corporate bond funds

Huang and Shi (2025) Model Uncertainty in the Cross Section Jan 5, 2025 15 / 17
Equity fund flows with different investment objective codes

0.0100

0.040
0.01

−0.0275
−0.02

0.015
Impulse Responses

Impulse Responses

Impulse Responses
−0.0650

−0.010
−0.05

−0.1025

−0.035
−0.08

Exogeneous Model Uncertainty Exogeneous Model Uncertainty Exogeneous Model Uncertainty

−0.1400

−0.060
−0.11

Endogeneous Model Uncertainty Endogeneous Model Uncertainty Endogeneous Model Uncertainty

1 6 11 16 21 26 31 36 1 6 11 16 21 26 31 36 1 6 11 16 21 26 31 36

Months Ahead Months Ahead Months Ahead

Style equity fund Small-cap funds Large-cap funds

Following high model uncertainty shocks, equity outflows mainly from style and small-cap
funds, instead of large-cap or sector funds

Huang and Shi (2025) Model Uncertainty in the Cross Section Jan 5, 2025 16 / 17
Conclusion

Model uncertainty in cross-sectional asset pricing

▶ varies significantly over time and is persistently high in bad times
▶ commands a significantly negative risk premium

Combined with low marginal factor probabilities: selecting SDF models is elusive
▶ Example periods: 2008 GFC, recent years from 2018–2020

Outflows from equity funds into US government bond funds under high uncertainty:
▶ No such patterns detected using VIX or financial uncertainty VIX financial uncertainty

Huang and Shi (2025) Model Uncertainty in the Cross Section Jan 5, 2025 17 / 17
Appendix
Data Back

14 prominent factors, from July 1972 to December 2020

▶ Fama-French five factors (Fama and French, 2016)
▶ Momentum factor (Jegadeesh and Titman, 1993)
▶ Size, investment, and profitability factors (Hou, Xue, and Zhang, 2015)
▶ Short and long-term behavioral factors (Daniel, Hirshleifer, and Sun, 2020)
▶ HML devil (Asness and Frazzini, 2013)
▶ Quality-minus-junk (Asness et al., 2019)
▶ Betting-against-beta (Frazzini and Pedersen, 2014)

Consider models that contain at most one factor from following categories:
▶ Size (SMB or ME)
▶ Profitability (RMW or ROE)
▶ value (HML or HML Devil)
▶ investment (CMA or IA)
Simulation study: posterior properties without misspecification Back

Three years Five years 50 years

Scenario g=2 g=4 g=16 mix. g g=2 g=4 g=16 mix. g g=2 g=4 g=16 mix. g

Posterior Probabilities of Factors P[γj = 1 | D]:

γ0,j = 1 0.98 0.98 0.98 0.98 1.00 1.00 1.00 0.99 1.00 1.00 1.00 1.00
(0.03) (0.03) (0.03) (0.04) (0.01) (0.01) (0.02) (0.02) (0.00) (0.00) (0.00) (0.00)
γ0,j = 0 0.49 0.45 0.34 0.25 0.48 0.44 0.34 0.21 0.48 0.44 0.33 0.10
(0.07) (0.09) (0.10) (0.10) (0.07) (0.08) (0.10) (0.10) (0.07) (0.08) (0.09) (0.07)

Posterior Probabilities of Models P[Mγ | D]:

Mγ = Mγ0 0.07 0.09 0.18 0.29 0.07 0.10 0.19 0.39 0.08 0.10 0.20 0.67
(0.04) (0.06) (0.12) (0.17) (0.04) (0.06) (0.12) (0.19) (0.04) (0.06) (0.12) (0.22)
Mγ ⊃ Mγ0 0.05 0.05 0.05 0.04 0.06 0.06 0.05 0.04 0.06 0.06 0.05 0.02
(0.01) (0.01) (0.01) (0.01) (0.00) (0.01) (0.01) (0.01) (0.00) (0.00) (0.01) (0.01)

Model Uncertainty Measure E :

0.42 0.40 0.36 0.31 0.39 0.37 0.33 0.26 0.38 0.36 0.32 0.14
(0.04) (0.04) (0.05) (0.05) (0.03) (0.03) (0.03) (0.04) (0.02) (0.03) (0.03) (0.04)

true factors = {market, SMB, HML, MOM, RMW, CMA}

f = {market, SMB, HML, MOM, RMW, CMA, QMJ, FIN, PEAD, BAB}
1,000 simulations; Sample sizes are three, five, and 50 years of daily observations
Simulation study: model uncertainty under misspecification Back

f = {MKT, SMB, HML, MOM, RMW, CMA, QMJ, FIN, PEAD, BAB} excluding the factor in each column
True factors: FF3 plus the omitted factor (the name of which is at the top of each column)
1,000 simulations with T = 750 days, with standard deviations across simulations in parenthesis

Omitted factor: MOM RMW CMA FIN PEAD QMJ BAB

Posterior Probabilities of Factors P[γj = 1 | D]:

γ0,j = 1 0.94 1.00 1.00 1.00 1.00 0.99 1.00
(0.17) (0.02) (0.00) (0.01) (0.01) (0.06) (0.01)
γ0,j = 0 0.49 0.35 0.27 0.43 0.29 0.36 0.38
(0.33) (0.27) (0.21) (0.31) (0.22) (0.25) (0.27)

Model Uncertainty Measure E :

0.46 0.46 0.46 0.43 0.46 0.49 0.47
(0.08) (0.06) (0.05) (0.07) (0.05) (0.06) (0.06)
Upper bound = (p − pγ0 )/p 0.67 0.67 0.67 0.67 0.67 0.67 0.67

Note: Extremely high model uncertainty should not be driven by model misspecification.
Regressions of Model Uncertainty on Contemporaneous Variables
X FinU. MacroU. RealU. EPUI EPUII VIX TS DS

β 0.21 0.17 0.14 0.00 0.00 0.01 -0.03 -0.00

(1.95) (1.53) (1.20) (0.33) (1.07) (2.20) (-3.44) (-0.09)

#obs. 546 546 546 432 432 420 546 546

The table reports results from the following regression:

Et = β 0 + βXt + ρEt−1 + ϵt ,

where the variable Xt represents a) macro, financial, and real uncertainty measures from Jurado et al. (2015)
and Ludvigson et al. (2021) (Fin U, Macro U, and Real U); b) two economic policy uncertainty (EPU) indices
from Baker et al. (2016) (EPU I and EPU II); c) the CBOE VIX index (VIX); d) the term spread between ten-year
and three-month treasuries (TS), e) the default spread between BAA and AAA corporate bond yields (DS). The
t-statistics in parenthesis are computed based on Newey-West standard errors with 36 lags.
Model uncertainty in the cross section: two states Back

0.30 High Model Uncertainty: DEC 2007

0.25 Low Model Uncertainty: FEB 1994
Model Probability

0.20
0.15
0.10
0.05
0.00
0 10 20 30 40 50
Model ID
Is Model Uncertainty Priced? – Robustness Check Back

Control for VIX & financial uncertainty in the AR(1) regression of model uncertainty
=⇒ model uncertainty innovations now orthogonal to both two variables
Table A1: Risk Premia of Model Uncertainty Shocks: Monthly Estimates

Number of PCs: 5 6 7 8 9 10
λE -0.058*** -0.059*** -0.059*** -0.057*** -0.055*** -0.051***
s.e. 0.016 0.017 0.017 0.017 0.017 0.018
Time-series R2 5.0% 5.1% 5.1% 5.3% 5.4% 5.6%
The table reports the risk premia estimates of model uncertainty shocks (Etar1 ) based on the three-pass method of Giglio and Xiu (2021). We add the control for
contemporaneous term spread and financial uncertainty to estimate the AR(1) innovations in model uncertainty. In all estimations, we standardize Etar1 to have a unit
variance. In particular, we project Etar1 onto the space of large PCs of 275 Fama-French characteristic-sorted portfolios in the US market. The number of latent factors
ranges from five to 10. If the 90% (95%, 99%) confidence interval of the risk premium does not contain zero, the risk premium estimate will be highlighted by * (**, ***).
We also report the time-series fit in each panel. Sample: 1975/07 - 2020/12.

Risk premia estimates, as well as the time-series fit, are similar to the baseline case
Equity fund flows to VIX shocks Back

0.0600

0.1300

0.0800
0.0125

0.0725

0.0375
Impulse Responses

Impulse Responses

Impulse Responses
−0.0350

−0.0050
0.0150
−0.0825

−0.0425

−0.0475
Exogeneous Model Uncertainty Exogeneous Model Uncertainty Exogeneous Model Uncertainty
−0.1300

−0.1000

−0.0900
Endogeneous Model Uncertainty Endogeneous Model Uncertainty Endogeneous Model Uncertainty

1 6 11 16 21 26 31 36 1 6 11 16 21 26 31 36 1 6 11 16 21 26 31 36

Months Ahead Months Ahead Months Ahead

Style equity fund Small-cap funds Large-cap funds

0.1400
0.110

0.10
Exogeneous Model Uncertainty Exogeneous Model Uncertainty Exogeneous Model Uncertainty
Endogeneous Model Uncertainty Endogeneous Model Uncertainty Endogeneous Model Uncertainty

0.1025
0.065

0.06
Impulse Responses

Impulse Responses

Impulse Responses
0.0650
0.020

0.02
−0.025

0.0275

−0.02
−0.0100
−0.070

−0.06
1 6 11 16 21 26 31 36 1 6 11 16 21 26 31 36 1 6 11 16 21 26 31 36

Months Ahead Months Ahead Months Ahead

Government bonds Money markets Corporate bonds

Equity fund flows to financial uncertainty shocks Back

0.0300

0.0600

0.0500
0.0025

0.0225

0.0225
Impulse Responses

Impulse Responses

Impulse Responses
−0.0250

−0.0150

−0.0050
−0.0525

−0.0525

−0.0325
Exogeneous Model Uncertainty Exogeneous Model Uncertainty Exogeneous Model Uncertainty
−0.0800

−0.0900

−0.0600
Endogeneous Model Uncertainty Endogeneous Model Uncertainty Endogeneous Model Uncertainty

1 6 11 16 21 26 31 36 1 6 11 16 21 26 31 36 1 6 11 16 21 26 31 36

Months Ahead Months Ahead Months Ahead

Style equity fund Small-cap funds Large-cap funds

0.0600

0.1000

0.0400
Exogeneous Model Uncertainty Exogeneous Model Uncertainty Exogeneous Model Uncertainty
Endogeneous Model Uncertainty Endogeneous Model Uncertainty Endogeneous Model Uncertainty
0.0225

0.0675

0.0125
Impulse Responses

Impulse Responses

Impulse Responses
−0.0150

−0.0150
0.0350
−0.0525

−0.0425
0.0025
−0.0900

−0.0300

−0.0700
1 6 11 16 21 26 31 36 1 6 11 16 21 26 31 36 1 6 11 16 21 26 31 36

Months Ahead Months Ahead Months Ahead

Government bonds Money markets Corporate bonds

Does Model Uncertainty Matter?

Table A2: Out-of-Sample Model Performance

BMA Top 1 Full Model Carhart4 FF5 HXZ4 DHS3

Low Model Uncertainty 2.572 2.565 2.568 1.288 1.624 1.829 2.282
- - - *** *** *** -
Middle Model Uncertainty 1.717 1.653 1.771 0.450 0.677 1.232 1.818
- - - *** *** ** -
High Model Uncertainty 1.251 1.125 1.106 0.564 0.584 0.552 0.897
- * * *** *** *** **

First three columns: (1) BMA: Bayesian model averaging; (2) Top 1: the model with the highest model probability; (3) Full Model: include all 14 factors. We report the
results on testing the null hypothesis that the Sharpe ratio of BMA is equal to the model γ, i.e., H0 : SR2bma = SR2γ . We use the non-parametric Bootstrap to test the
null hypothesis. *, ** and *** denote significance at the 90%, 95%, and 99% level, respectively.

Note: Model uncertainty matters only when it is heightened