Construction of Capital Pricing Model Based on Deep Learning and

Conditional Risk

Muyao Liu 1, Songrui Chen 1, Peibei Tang 2, Shi QIU 3

1
School of Business， Nanjing University， Nanjing 210093， China
2
School of Business， Xi’an Jiaotong- Liverpool University， Suzhou215028，
China
3
Jingyue Street, Jilin University of Finance and Economics, Nanguan District,
Changchun city, Jilin Province

Abstract:

The function of capital prices, which is the core of modern finance is questioned in all
areas of society. The relation between predicted return and risk is analyzed using the
traditional pricing model known as (CAPM) capital asset pricing model, which act as
an essential model for the pricing of risky securities however it has serious price
deviations, and capital price deviations can pose various crises to the economic
market and cause deviations in investors' asset valuations. A misleading investor
decision. It leads to improper allocation of social resources, which causes turmoil in
financial markets. Based on this, many researchers have sought ways to solve
problems from various aspects. This article uses deep learning and conditional risk as
a way to study capital prices establishes a model of capital prices and provides the
theoretical basis for subsequent research.

Keywords: capital pricing; deep learning; conditional risk, CAPM, Conditional

Drawdown-at-Risk(CDaR) and Conditional Value at Risk(CVaR).

1. Introduction
The development of the capital asset pricing model is inseparable from the
efficient market hypothesis. As one of its premise hypotheses, the efficient market
hypothesis is still active in various fields of the economic market. Samuelson first
publicized the concept of an effective market to the world in 1965. After continuous
exploration by researchers such as Fama and Jensen, it gradually improved and
formed a complete theoretical system, namely the effective market theory. The
effectiveness of the securities market is accurately divided into several different
levels. Fama believes that in a fully efficient market, all information should be
transparent, consumers can purchase assets at a completely fair price, and the price of
securities can reflect all the information of the company. The information reliability
and comprehensiveness of the existing securities market are not the same. Fama
divides the securities market in accordance to the level of information provided by
market that can be weak effective market, strong effective market and semi-strong
effective market. Since traditional Capital Asset Pricing Model (CAPM) was
proposed, many scholars have continuously verified and questioned it. The article
divides the research results of various scholars from the perspective of whether
investors are heterogeneous or not, and can be roughly classified into two categories,
one category is based on the research under the assumption of non-heterogeneous
investors, and the other is based on the research under the assumption of
heterogeneous investors. In Section 2.1, we defined the CAPM, CDaR, and CVaR. In
Section 2.3, we constructed the proposed model based on deep learning. In Section 3,
we analyzed the performance of the proposed model. And in Section 4, the conclusion
of the paper is proposed.

1.1 Capital Asset Pricing Model (CAPM)

In finance, all investments possess some level of risk. When the risk linked with the
investment is higher the investor hopes to get more returns to balance the higher risks.
The capital Asset Pricing Model (CAPM ) is very helpful for investors to evaluate
investments and make positive decisions regarding investments as it calculates an
approximate return on the grounds of the proportion of risk present in the investment.
The capital Asset Pricing Model (CAPM ) model is the basic model for investors to
determine the estimated outcome of the investment and helps them in analyzing and
maintaining the investment portfolio.
To obtain a capital asset pricing model based on the risk measurement of conditional
risk reduction, you must first clearly define Conditional Risk Reduction (CDaR).
Conditional risk mitigation measurement measures risk in combination with
Conditional Value at Risk (CVaR).

1.2 Conditional Drawdown at Risk (CDaR) and Conditional Value at Risk

(CVaR)

Conditional drawdown at Risk (CDaR) which is also called expected shortfall is very
similar to (CVaR) Conditional Value at Risk and can optimize a portfolio as both
provide risk assessment techniques that can measure mean value in distribution
beyond a specific level. Conditional drawdown at Risk (CDaR) tells us the average
measure of the downside risk for all the occasions where drawdown surpasses a
certain threshold in an investment portfolio. The measure of tail risk in any
investment portfolio is quantized by Conditional Value at Risk (CVaR) which takes
the mean value of the extreme losses in the tail of the dispensation of feasible payback
measures beyond the cutoff points a VAR can have. Conditional Value at Risk (CVaR)
can assess the lower tail loss. The loss function of CDaR is the absolute reduction
(AD) function, which is different from the loss function of CVaR.

The AD function can be defined as follows:

 (1)

,
Function (1) has the following properties:

It’s a non-negative function ;

Translation invariance: That is , for the constant C ;

Positive oddness: That is , for "non-negative " ;

Convexity: which is .
Under the nature of the function, suppose the probability that the drop loss

function is less than a certain threshold is , and have

。Under this definition, the CDaR function is obtained, namely:

(2)
The generality of Equation (2) can be introduced into the capital asset pricing
model. Assume r 1 , r 2 , … , r T is the rate of return on risky assets，w t is the cumulative

rate of return at time t, which is either a function of the sum ， which is：

； Either a function of the product ， which is . For ease of

analysis, the cumulative rate of return is defined as the sum function of the rate of
return on risky assets. The fall of the risky asset at time t is the loss relative to the
highest point of the cumulative rate of return. That is, within the range of the rate of

return , when , there is ; when there is . So the normal decline

is defined as:

(3)
The decline is non-negative and is often drawn as a curve below the horizontal
plane. If at time t, the cumulative rate of return reaches the highest value in the range

of , then d t =0. At the same time, the return accumulation rate vector

w=( w1 , … , wT ) of risk assets in the range of T can be described by the CDaR

function. Assuming that is an integer, within the confidence range , the

amount of decrease in conditional risk can be described by the function , that is,

the maximum amount of decrease under the average confidence level. This
function can be expressed as the sum of declines, namely：

(4)

When d t is one of the integers ， ； Under other

conditions， =0. The maximum drop and the average drop are special cases of the
amount of conditional risk drop, corresponding to situations α =1 and α =0
respectively. Within any confidence interval level, that is, within α ∈ [ 0 ,1 ], there are
ps > 0 sampling paths r 1 , r 2 , … , r T for the rate of return s with probability s ( s=1 , … S ) ,

so as to obtain the rate of return r s1 , r s 2 , … , r sT of the sampling path. So，

and . They are the cumulative rate of return and the decline in

risky assets 。 For the average confidence level , the maximum amount of
decline can be defined as::

(5)

Where

1.3 Literature review

When CAPM was proposed many researchers such as Merton, Campbetaell, Fama,
and French insist that Fama's traditional CAPM is not completely accurate and has
certain asset pricing deviation issues. In response to this problem, many scholars
have carried out different degrees of research. Most scholars conduct analysis and
research from a series of preconditions limited by the existence of traditional CAPM.
Research results have a certain effect on the improvement of CAPM, but there are
many research studies and different perspectives.
Capital Asset Pricing Model (CAPM ) model shows that excluding the momentum
the remaining three factors of a model comprise the market factor which depends on
the equity of the market, and a factor depending on the equity of book-to-market. It
pretty well explain the area of average stock payback. However, with time, the Fama-
 French model fails to apply to a vast variety of asset pricing.
In the field of quantitative finance, the financial limits applied to the machine learning
mechanism are analyzed. Practitioner faces a large number of challenges to make sure
the implementation and adoption of the latest scientific innovations. The use of
Machine Learning techniques i.e state-of-the-art Machine Learning techniques and
high-performance computing (HPC) infrastructures a theoretical analysis of machine
learning programs in comparison to CAPM and it was observed that Machine
Learning algorithms have greater flexibility to accommodate a lot more series feature
to calculate the returns of the target in US equity. In this way, deep learning can
provide better paths for return calculations.
In 2011 Cao et al made the first attempt to evaluate and develop a model based on
deep learning using the bottom-up method for the establishment of asset pricing
models having parameter (input), security payback, and risk factors. It was observed
that using deep learning the stock prices can be forecasted more accurately with a
minimum error occurring in-sample dataset and out sample dataset present in the asset
pricing models.

2 Methodology
2.2 Model Construction Based on Deep Learning
Traditional linear multifactorial models cannot handle high-dimensional enterprise
characteristics. Deep Learning prediction-based algorithms are keenly designed to
deal with huge data, greater dimensionality, and unorganized data which makes them
the perfect technique to solve problems in a variety of fields. The introduction of deep
learning techniques provides a new path for using high-dimensional enterprise
characteristics in multifactorial models. It will be combined with the conditional risk
analysis as the study of conditioning provides a better mapping of risk factors so it
can better control the dependency of assets in the investment portfolio.
The specific steps are as follows:

Step 1:

High-quality company special expedition Action

Primitive import advancing Artificial neural network Depth learning Depth special

expedition , In, ， Is a company characteristic

dimension， Is the number of companies ， ， Is the depth

feature dimension， Represents the number of neural network layers， means

only the input layer and the deep feature forming layer, means that in addition to
the input layer and the deep feature forming layer, there is a hidden layer in between.
Step 2:
Calculate the weight of each stock to the long/short portfolio based on depth feature
A. Weights can use equal weights, and linear rank weights can also use non-linear
rank weights.

(6)

In, ， is the normalized number

function：

(7)
Step 3:
Multiply the stock long-short portfolio weight and the stock's corresponding return to
obtain the depth factor ：
f k =wk ×r (8)

In, .
Once you have the depth factor, combine it with other traditional pricing factors
to build a linear multi-factor pricing model:

(9)
Where 𝛽 is the beta coefficient of the depth factor 𝑓, and γ j is the beta coefficient
of the benchmark factor g j. In this paper, the benchmark factor uses Fama-French's
(1993) three factors. The depth factor and the benchmark factor of time form the
company characteristics based on the lag time.

2.3 Model Parameter Estimation

Estimating model parameters can be done in the deep learning framework. In
deep learning, estimation of model parameters is achieved by optimizing the objective
function. Concerning Lettau and Pelger (2020), the objective function minimizes the
weighted sum of time-series changes and section pricing errors.

(6)
Where, Ri , t is the return observation value of asset portfolio i used for parameter
estimation at time t , ^
Ri , t is the portfolio return calculated by the linear multi-factor

pricing model：
 (7)

is the estimated value of the pricing error of the asset portfolio：

(8)
𝜆 is a model hyper parameter, used to control the trade-off between time-series
changes and pricing errors, and is set to 10. 𝐴, 𝑏, 𝛽, 𝛾 are model parameters. Among
them, the dimensional scales of 𝐴, 𝑏, 𝛽 depend on the method of deep learning and its
architecture. This article uses feedforward neural networks along with (LSTM) long
short-term memory recurrent neural networks. Construction Depth Coefficients; 𝛾 is
the corresponding parameter of the three Fama-French (1993) coefficients.

3 Results and discussion

3.1 Pricing Model performance
Table 1 shows the statistical performance information for the model. This table
shows the time of Fama and French (1993) as a benchmark model, combining layers 1
to 4 of the neural network with depth coefficients 1 to 8 to create a multifactorial
pricing model based on 40 depth coefficients (below), Depth model) Sequence R2
information. The reference model R2 is the absolute value, and the 40 depth model R2
is the rate of change value concerning the reference model. Compared to the
benchmark model, the R2 inside and outside the 40 depth model samples has been
improved to varying degrees. Even with a market combination that has good
benchmark model performance and a return interpretation of the CSI 300 Index
combination, the depth model has a non-negligible increase in model interpretation. In
addition, the Fama-French combination interpretation resulted in relatively poor
performance of the benchmark model, while the depth model resulted in a relatively
significant increase in model interpretation. This information shows that the depth
model is statistically superior to the benchmark model.

Table 1. Time series R2

No of
L=1 L=2 L=3 L=4 L=1 L=2 L=3 L=4
Layers
Factor Bivariate Permutation and Univariate Permutation and
Combination Combination
FF3 0.955 0.955 0.955 0.955 0.948 0.948 0.948 0.948
1 1.13% 1.14% 1.13% 1.05% 1.48% 1.54% 1.51% 1.39%
2 1.32% 1.30% 1.27% 1.12% 1.77% 1.75% 1.52% 1.50%
3 1.54% 2.31% 1.78% 1.28% 2.04% 2.77% 2.37% 1.47%
4 1.84% 2.16% 2.13% 1.25% 2.40% 2.82% 2.76% 1.47%
5 2.17% 2.76% 1.26% 1.38% 2.72% 3.47% 1.54% 1.50%
6 2.09% 2.71% 2.29% 2.39% 2.83% 3.48% 2.78% 3.11%
7 2.53% 2.63% 2.66% 2.90% 3.23% 3.37% 3.20% 3.64%
 8 2.71% 2.77% 2.66% 2.36% 3.36% 3.45% 3.21% 3.06%
Fama-French Combination Industry Portfolio
FF3 0.881 0.881 0.881 0.881 0.813 0.813 0.813 0.813
1 5.60% 5.65% 5.49% 5.10% 0.47% 0.52% 0.50% 0.45%
2 6.00% 5.48% 5.58% 5.38% 1.04% 1.34% 0.87% 0.47%
3 6.19% 6.97% 6.98% 5.57% 1.51% 2.39% 1.75% 0.86%
4 6.91% 6.97% 7.06% 5.17% 1.94% 2.21% 2.35% 0.76%
5 7.37% 7.90% 5.31% 5.00% 2.43% 3.51% 1.23% 0.76%
6 7.28% 7.90% 7.32% 7.42% 2.76% 3.47% 2.57% 2.61%
7 7.96% 7.79% 7.87% 8.18% 3.34% 3.16% 2.91% 3.37%
8 8.03% 7.81% 7.94% 7.78% 3.62% 3.53% 3.18% 2.71%
Market Capitalization Weighted Market Portfolio CSI 300 Index
FF3 0.991 0.991 0.991 0.991 0.975 0.975 0.975 0.975
1 0.08% 0.10% 0.11% 0.09% 0.36% 0.37% 0.40% 0.35%
2 0.10% 0.25% 0.11% 0.10% 0.40% 0.70% 0.37% 0.34%
3 0.26% 0.22% 0.17% 0.10% 0.50% 0.58% 0.54% 0.37%
4 0.33% 0.24% 0.26% 0.00% 0.67% 0.65% 0.74% 0.44%
5 0.28% 0.36% 0.27% 0.21% 0.89% 0.96% 0.53% 0.43%
6 0.32% 0.41% 0.31% 0.19% 0.91% 0.92% 0.74% 0.86%
7 0.33% 0.31% 0.29% 0.34% 0.91% 0.94% 0.69% 0.62%
8 0.42% 0.39% 0.29% 0.20% 1.02% 0.88% 0.83% 0.60%

3.2 Depth Factor Pricing Contribution

Following common practice, multiply the Sharpe ratio calculated for monthly
returns by √ 12 to convert to the annual Sharpe ratio. Table 2 shows the annual Sharpe
ratio of the pricing factor in Zhang Cheng's portfolio. The total annual Sharpe ratio for
the Pharma-French three-factor Chang Chen is 0.53. After adding the depth factor, the
Sharpe ratio of the pricing factor in Zhang Cheng's portfolio has increased
significantly. The lowest improvement rate is 122% and the highest improvement rate
is 472%. Barillas et al's tests have shown that all depth coefficients built by different
neural network layers have a small pricing contribution to the pricing model.
Increasing the depth factor does not affect the marginal price contribution of the
previous depth factor. Similarly, from horizontal comparisons, there is no apparent
monotonous relationship between the number of neural network layers and the depth
factor and pricing performance used to build the depth model.

Table 2. Combined Sharpe Ratio

Number of
Layers L=1 L=2 L=3 L=4

Factor
FF3 0.53 0.53 0.53 0.53
1 1.58*** 1.66*** 1.70*** 1.52***
 2 2.04*** 3.04*** 1.99*** 1.87***
3 1.21*** 2.21*** 1.08*** 1.54***
4 1.71*** 1.81*** 2.83*** 2.47***
5 1.35*** 2.79*** 2.02*** 1.95***
6 1.89*** 1.87*** 2.06*** 1.67***
7 1.43*** 2.08*** 2.01*** 2.35***
8 1.36*** 2.18*** 2.25*** 2.22***

2. Conclusion and future work

This white paper also combines deep learning with a capital asset pricing model
based on examining the significant non-linear relationships between corporate
characteristics and equities and their earnings. Inspired by traditional Pharma-French
three- or five-factor construction methods, we explored the non-linear ordering of
stock returns to build stock price coefficients and created a depth coefficient model
that minimizes mean square error and price error establish. The objective function is
used to build the final detailed pricing factor to explain the expected return on the
stock section. Through empirical studies in these two areas, this article draws the
following conclusions:
(1) The out-of-sample R2 of the nonlinear model is significantly better than the
corresponding value of the linear model, and the out-of-sample R2 of the long-term
and short-term memory neural network is the corresponding value of the feedforward
neural network. The neural network size is not as large as possible. If the neural
network is too large, overfitting will occur.
(2) The returns of long/short portfolios built on different models are almost always
greater than zero. The return rate of long and short combinations constructed by the
linear model is less than 0.02 in most cases, and the return rate of the combinations
constructed by the feedforward neural network model exceeds 0.02 in about half the
time and is constructed by LSTM. In most cases, the combination will exceed 0.02.
The return rate of the long/short portfolio built with the three models is the same as
the time series of index returns in Shanghai and Shenzhen 300, but the returns of the
portfolio built with the models are significantly better than the index returns.
(3) In most cases, the Sharpe ratio of long/short combinations constructed by
nonlinear models is superior to that of linear models. The Sharpe ratio of the long and
short combinations built by the LSTM neural network is higher than the Sharpe ratio
of the feedforward neural network and exceeds the corresponding value of the linear
model during the sample period.
The results obtained from the proposed model are satisfactory as the return rate of
long/short portfolios built on the current model is less than zero however further
iteration in the algorithm to obtain the results where the pricing error is as low as zero
by introducing the backpropagation where feedback derived from continuous changes
in the pricing error is again fitted to model errors at each stage of the model learning
 process further work can be done on this model to make it the best factor model that
explains all existing alphas.

Data Availability Statement

The corresponding author can provide the datasets used and analyzed during the
current study upon reasonable request.
Conflict of Interests
The authors of this paper declare that they have no competing interests.
Funding Statement
The paper didn’t receive any funding.

References
[1] Fama, E.F. and French, K.R. Common risk factors in the returns on stocks and
bonds [J]. Journal of Financial Economics, 1993,33,3–56.
[2] Ang, A., Hodrick, R.J., Xing, Y., and Zhang X. The cross-section of volatility and
expected returns [J]. The Journal of Finance.2006, 61 (1), 259–299.
[3] Mark M. Carhart. On persistence in mutual fund performance [J]. The Journal of
Finance,1994,52(1),57-82.
[4] Novy-Marx, R. The other side of value: The gross profitability premium [J].
Journal of Financial Economics,2013,108(1),1–28.
[5] Frazzini, A. and Pedersen, L. H. Betting against beta [J]. Journal of Financial
Economics.2014,111 (1),1–25.
[6] McLean, R. D. and Pontiff, J. Does academic research destroy stock return
predictability [J]. The Journal of Finance,2016,71 (1), 5–32.
[7] Hou, K., Xue, C., Zhang, L. Digesting anomalies: an investment approach [J].
The Review of Financial Studies, 2015,28(3),650–705.
[8] Green, J., Hand, J., and Zhang X. The characteristics that provide independent
information about average U.S. monthly stock returns [J]. Review of Financial
Studies,2017, 30(12): 4389 – 4436.
[9] Hou, K., Xue, C., and Zhang. Replicating anomalies [J]. The review of financial
studies, 2020,33(5),2019–2133.
[10] Hou, K., Mo, H., Xue, C., and Zhang, L. An augmented q 5 model with expected
growth [J]. Review of Finance.2020 - theinvestmentcapm.com.
[11] Fama, E. F., and French, K.R. Choosing factors [J]. Journal of Financial
Economics, 2018,128(2), 234 – 252.
[12] Asness, C., Moskowitz, T., and Pedersen. L. Value and momentum everywhere
[J]. The Journal of Finance, 2013, 68(3): 929–985.
[13] Asparouhova, E., Bessembinder, H., and Kalcheva, I. Noisy prices and inference
regarding returns[J]. Journal of Finance, 2013, 68: 665–714.
[14] Philip Ndikum. Machine Learning Algorithms for Financial Asset Price
Forecasting [J]. arXiv, 2004,
[15] Usman Ayub, Muhammad Naveed Jan, Muhammad Asim Afridi. Capital Asset
Pricing Model and Artificial Neural Networks[J]. Pakistan Journal of Social
Sciences (PJSS), Vol. 40, No.2 (2020), pp. 673-688

A Case of Pakistan’s
Equity Market

[16] A Case of Pakistan’s

Equity Market