4th International Conference on Computation, Automation and Knowledge Management (ICCAKM 2023)

Portfolio Optimization using Machine Learning

Techniques
T. Sugadev N. Sheik Hameed S. Vijayakumar
B.S. Abdur Rahman Crescent Institute B.S. Abdur Rahman Crescent Institute B.S. Abdur Rahman Crescent Institute
of Science and Technology of Science and Technology of Science and Technology
2023 4th International Conference on Computation, Automation and Knowledge Management (ICCAKM) | 979-8-3503-9324-8/23/$31.00 ©2023 IEEE | DOI: 10.1109/ICCAKM58659.2023.10449598

Vandalur, Chennai Vandalur, Chennai Vandalur, Chennai

[email protected] [email protected] [email protected]

P. Tamilarasan Md. Sahidul Islam

College of engineering and B.S. Abdur Rahman Crescent Institute
Technology, SRMIST of Science and Technology
Kattatankulathur Vandalur, Chennai
[email protected] [email protected]

Abstract— A crucial component of financial decision-making In this context, the artificial intelligence discipline of
is portfolio optimisation, which aims to minimise risk and machine learning has become a disruptive force that has the
maximise profits. The mean-variance model is one of the potential to completely transform portfolio optimization[5].
traditional methods that struggles to capture the dynamic Making better and more flexible portfolio decisions is made
character of financial markets. The use of machine learning possible by machine learning, which makes use of data-driven
approaches to improve portfolio optimisation is examined in this insights, trends, and prediction models[6]. It has established
research article. We start by going over fundamental ideas like itself as a game-changer in the financial industry because to
asset allocation and selection before gradually getting into the its capacity to analyse enormous amounts of financial data and
intricacies of machine learning techniques. We go over how to
spot complex and delicate linkages within it. Machine
forecast asset returns and correlations using clustering,
regression, and classification techniques. We also investigate
learning algorithms may dynamically alter portfolio
portfolio rebalancing techniques using reinforcement learning. allocations to maximise performance by taking lessons from
The paper reviews real-world data sources, data preprocessing, past market behaviour and adjusting to the present[7].
and model evaluation techniques, highlighting the importance The goal of this study is to investigate in detail how
of backtesting and risk management. Furthermore, we examine machine learning methods may be used to portfolio
recent advancements in deep learning and neural networks to optimisation. Its fundamental goal is to address the
model non-linear relationships and assess their impact on
shortcomings and rigidity of traditional methods, which often
portfolio optimization.
fail to account for the ever-evolving nature of financial
Keywords—Portfolio optimization, Machine learning, Mean- markets. By exploring the application of machine learning in
variance model, Asset selection, Asset allocation, Regression, portfolio optimization, we aim to provide investors, portfolio
Classification, Clustering, Reinforcement learning, Data managers, and financial practitioners with a more robust and
preprocessing, Model evaluation. adaptable set of tools to navigate the complex and dynamic
financial landscape effectively. In essence, this research
I. INTRODUCTION serves as a bridge connecting the wisdom of classical finance
with the potential of machine learning, ultimately offering a
A. Background more data-driven, responsive, and informed approach to
The world of finance has undergone a profound portfolio management in today's intricate financial ecosystem.
transformation in recent years, ushered in by the rapid
adoption of machine learning techniques[1]. This shift has had B. Motivation
a far-reaching impact on various aspects of financial decision- The motivation behind this study is deeply rooted in our
making, and one of the areas that has felt its influence most recognition of the inherent limitations and inefficiencies
profoundly is portfolio optimization[2]. Portfolio within traditional portfolio optimization techniques. These
optimization, essentially the art of constructing investment conventional approaches often rely on oversimplified
portfolios with the goal of maximizing returns while assumptions that frequently diverge from the intricate and
simultaneously minimizing risk, has traditionally been dynamic reality of financial markets[8]. A primary challenge
underpinned by classical methods, most notably the mean- with these traditional models is their reliance on assumptions
variance model introduced by Harry Markowitz in the mid- such as normally distributed asset returns and constant
20th century. However, these conventional approaches have correlations among assets, which fall short of capturing the
increasingly revealed their limitations as the financial genuine complexities of financial markets[9]. In actuality,
landscape has grown more dynamic and intricate[3]. The financial markets are characterized by non-linear
complexities of modern financial markets, which span an relationships, continually changing economic conditions, and
array of diverse asset classes, trading instruments, and the intricate interplay of various factors that these simplifying
geographical regions, pose considerable challenges to the assumptions fail to represent accurately. As a result, investors
classical methodologies[4]. These challenges are magnified frequently encounter substantial obstacles when striving to
by factors such as changing market conditions, evolving attain their desired risk-return profiles, particularly in the
investor preferences, and unprecedented technological context of market volatility and economic uncertainty.
advancements. Machine learning has the ability to address these

979-8-3503-9324-8/23/$31.00 ©2023 IEEE

Authorized licensed use limited to: SRM University Amaravathi. Downloaded on December 12,2024 at 13:50:40 UTC from IEEE Xplore. Restrictions apply.
shortcomings, which is what motivates researchers to The field made important advancements when Yu et al.
investigate its integration with portfolio optimisation. (2019) investigated the use of deep reinforcement learning for
portfolio optimisation. Their research yielded positive results
Machine learning provides a potent remedy by permitting that showed how adaptable portfolios may be to changing
the inclusion of a wider array of data sources, adapting to the market conditions. This approach employs state-of-the-art
ever-shifting dynamics of financial markets, and discerning techniques to generate investment options that dynamically
the intricate non-linear relationships between assets and adjust to shifting market trends and economic situations. Yu
market variables[10]. By harnessing the capabilities of et al.'s approach opens up more flexible and long-lasting
advanced algorithms like regression, classification, clustering, portfolio management strategies[12].
and reinforcement learning, machine learning effectively
unlocks the wealth of intricate information embedded within Ma et al. (2021) introduced a unique method that combines
financial data. The prospect of deploying machine learning in deep learning techniques with traditional portfolio
portfolio optimization holds the promise of delivering more optimisation approaches. They were able to get higher risk-
precise, adaptive, and data-driven decision-making. It gives adjusted returns as a result. This combination of traditional
investors the ability to make wiser decisions that are more in and contemporary approaches provide investors with a viable
line with the always changing financial environment[11]. This way to improve the performance of their holdings. Making
project aims to improve portfolio optimisation by using better educated investment choices is made possible by the
machine learning approaches. By doing so, it will provide integration of deep learning, which makes it possible to
investors with the necessary skills to succeed in the complex uncover complex patterns from financial data and include a
and challenging world of finance. larger variety of data sources[13].
C. Contribution Breeden et al. (2021) investigated the use of Bayesian deep
This study makes a significant addition to the fields of learning for asset allocation and presented a framework that
machine learning and finance. It provides a thorough review outperforms traditional methods in capturing uncertainty.
of the cutting-edge methods required for portfolio Because Bayesian deep learning takes into account both the
optimisation, acting as a link between traditional financial inherent uncertainties and the predicted returns, it offers a
practises and machine learning's cutting edge approaches. versatile and reliable method of modelling asset allocation.
Furthermore, our work facilitates the actual implementation of Effective portfolio management in the face of erratic market
these advanced approaches by providing an in-depth analysis and economic situations has been shown by this approach[14].
of essential areas, such as real-world applications, metrics for An extensive comparison study of many machine learning
model assessment, and data collecting and preprocessing. We algorithms was carried out by Liu et al. (2023), which revealed
also point out that deep learning and neural networks hold the higher performance of ensemble approaches in portfolio
great promise for predicting and evaluating the effects of optimisation. Combining the advantages of many algorithms,
complex linkages in financial data on portfolio performance. ensemble techniques like bagging and boosting improve
In summary, our study establishes a foundation for the prediction accuracy and lessen model bias. The importance of
development of more flexible and data-driven techniques for these combinations in developing more reliable and accurate
portfolio optimisation, which might potentially improve risk- portfolio optimisation models is highlighted by Liu et al.'s
adjusted returns within the dynamic and constantly changing research[15].
realm of financial markets. These observations are in a great
position to provide priceless advice to investors, portfolio A portfolio rebalancing technique based on recurrent
managers, and financial specialists who want to remain at the neural networks (RNNs) was described by Singh et al. in
forefront of contemporary portfolio management. 2022. This method uses RNNs to identify temporal
relationships and modify asset allocations over time,
D. Objective improving portfolios' capacity to adjust to shifting market
The objective of this study is to examine and demonstrate circumstances. In contemporary portfolio management, the
the efficacy of machine learning techniques within the capacity to dynamically modify portfolio composition in
framework of portfolio optimisation. Our objective is to response to changing market data is a significant
anticipate asset returns using regression analysis by applying advantage[16].
mean-value-at-risk notions. Our objective is to combine risk Yoo et al. (2023) introduced a novel clustering-based
assessment and predictive analytics to provide a strong approach for portfolio diversification. By grouping assets with
framework for portfolio optimisation, particularly in the similar characteristics, Kim et al. demonstrated the benefits of
context of international stock markets. The purpose of this diversifying portfolios in a more structured and efficient
research is to show how machine learning may enhance manner. This approach helps investors mitigate risk and
decision-making processes and achieve greater risk-adjusted potentially enhance returns by spreading investments across
returns in intricate and dynamic financial environments, hence different asset clusters, each with its own risk and return
offering valuable insights to financial practitioners. profile[17].
II. LITERATURE REVIEW The study conducted by Abbasi et al. (2022) concentrated
The idea of fusing machine learning with portfolio on the use of generative adversarial networks (GANs) for the
optimisation has gained traction in recent years. With methods production of synthetic data in portfolio optimisation. Their
like Bayesian deep learning and deep reinforcement learning, work covers risk management in addition to data
researchers have made great strides. These projects show how augmentation. Because of their propensity to produce data that
machine learning may provide cutting-edge portfolio is very similar to actual data, GANs are an effective tool for
management skills by enhancing risk-adjusted returns and modelling different market situations and enhancing the
reacting to market situations. resilience of portfolio optimisation models[18].

Authorized licensed use limited to: SRM University Amaravathi. Downloaded on December 12,2024 at 13:50:40 UTC from IEEE Xplore. Restrictions apply.
 Mohanty et al. (2022) investigated how explainable AI paramount importance[23]. Robust imputation methods like
may be used into portfolio optimisation. Their study K-nearest neighbors imputation, mean imputation, or
highlights how crucial interpretability and transparency are to regression-based imputation are applied to address missing
machine learning-driven decision-making. Investors may data. Outlier detection techniques, including the Z-score
confidently make more informed decisions by using method or the isolation forest algorithm, are employed to
explainable AI approaches, which will help them comprehend isolate and correct extreme data points that could skew results.
the reasoning behind their portfolio allocations [19].
A hybrid machine learning strategy that blends time-series
analysis with reinforcement learning was presented by Yang Data transformation involves altering the data's structure
et al. in 2023. Their innovative method demonstrated superior to enhance its usability. Non-linear transformations, such as
performance in real-world portfolio management. The logarithmic or exponential transformations, are applied to
integration of reinforcement learning, which allows portfolios mitigate skewness in data distributions. Additionally,
to learn and adapt over time, with time-series analysis dimensionality reduction techniques like Principal
enhances the ability to respond to evolving market conditions, Component Analysis (PCA) may be used to reduce the
positioning portfolios to potentially achieve better risk- number of features while retaining the most relevant
information.
adjusted returns. Yang et al.'s research showcases the evolving
landscape of portfolio optimization, embracing the strengths C. Feature Engineering for Asset Selection
of modern machine learning techniques[20]. One of the most important aspects of portfolio
The significance of this research lies in addressing the optimisation is feature engineering. To enhance the
pressing need to modernize portfolio optimization methods. performance of the model, it entails either adding new features
By leveraging machine learning techniques, this study aims to or choosing the most illuminating features[24]. The objective
enhance the adaptability of portfolios in dynamic financial is to get relevant data from the dataset that may improve the
markets, improving risk management and potentially prediction capacity of the model. Techniques for complex
increasing returns. feature engineering include:

III. DATA COLLECTION AND PREPROCESSING • Lag Features: To reflect temporal interdependence,
lagged asset prices or returns are included.
A. Data Sources in Finance
• Technical Indicators: Developing indicators that shed
In the context of portfolio optimization utilizing machine light on asset behaviour, such as Bollinger Bands,
learning techniques, the selection and integration of Relative Strength Index (RSI), and moving averages.
appropriate data sources in finance hold paramount
significance[21]. The quality, diversity, and timeliness of data • Textual analysis: Using natural language processing
directly impact the effectiveness of portfolio management methods, sentiment ratings are extracted from news
models. A comprehensive range of data sources is considered, articles or social media.
encompassing both traditional and alternative sources. Table
1 provides an overview of some common data sources in • Volatility Measures: To assess risk, compute
financial markets. measures such as past volatility or GARCH models.

TABLE I. COMMON DATA SOURCES IN FINANCE

IV. METHODOLOGY
The use of machine learning approaches in portfolio
Data Source Description
Historical Price Time series data representing asset prices over time
optimisation is distinguished by a wide range of strategies,
Data each designed to target certain facets of the investing
Financial Corporate financial reports, balance sheets, income procedure[25]. We explore the complexities of these
Statements statements, and cash flow statements approaches in this section, emphasising their technical breadth
Economic Macroeconomic data such as GDP, inflation, and and contributions to contemporary portfolio management.
Indicators employment statistics
News and Textual data from news articles, social media, and A. Regression Models for Asset Return Prediction
Sentiment Data sentiment analysis
Market Order Real-time data on order volumes, bid-ask spreads, 1) Linear Regression
Book Data and trade execution The basis method for predicting asset returns is linear
Alternative Data Non-traditional data sources, including satellite regression. Assuming a linear connection, it describes the
imagery, foot traffic, and more relationship between one or more independent factors and the
Effective data integration strategies involve data dependent variable (asset returns)[26]. The mathematical
cleansing, transformation, and harmonization to ensure data formulation of simple linear regression is:
consistency and reliability. Data quality checks encompass
addressing missing values, outliers, and noisy data. Complex 𝛾 = 𝛽𝑜 + 𝛽1 𝑋 + 𝜖 (1)
data cleansing methodologies, including outlier detection
using multivariate statistical techniques like Mahalanobis
Distance or transformations such as Box-Cox, are Where:
employed[22]. Feature engineering is a critical phase in data • 𝛾 represents the predicted asset returns.
preprocessing.
• 𝛽𝑜 is the intercept term.
B. Data Cleaning and Transformation
Data cleaning encompasses the identification and • 𝛽1 is the regression coefficient.
rectification of inconsistencies, errors, or inaccuracies in the • X denotes the independent variable(s).
dataset. The handling of missing values and outliers is of

Authorized licensed use limited to: SRM University Amaravathi. Downloaded on December 12,2024 at 13:50:40 UTC from IEEE Xplore. Restrictions apply.
 • ϵ is the error term. relationships within data and provide insights into asset risk
profiles[29].
More advanced versions, such polynomial regression or
multiple linear regression, take into account many C. Clustering Techniques for Portfolio Diversification
independent variables or non-linear correlations, respectively, 1) K-Means Clustering
adding to the complexity. An integral part of portfolio K-Means clustering facilitates portfolio diversification by
optimisation is calculating anticipated returns, and linear grouping assets with similar characteristics. The objective is
regression models are crucial for this process. to minimize the within-cluster variance and maximize
between-cluster variance[30]. This technique identifies
TABLE II. PREDICTIVE ACCURACY ENHANCEMENT THROUGH
ENSEMBLE LEARNING clusters of assets with similar risk-return profiles, enabling the
creation of diversified portfolios. The algorithm aims to
Model Accuracy Accuracy minimize the following cost function:
(Before) (After)
Model 1 0.75 0.78 𝐽 = ∑𝑘𝑖=1 ∑𝑥∈𝐶𝑖 ||𝑥 − 𝜇𝑖 ||2 (2)
Model 2 0.72 0.76
Ensemble - 0.81
This table showcases the improvement in predictive Where:
accuracy through ensemble learning. Models 1 and 2 represent
individual regression models, while the Ensemble column • J is the cost function.
demonstrates the enhanced accuracy after implementing • k represents the number of clusters.
ensemble techniques.
• Ci denotes the data points in cluster i.
2) Time-Series Analysis
The main goal of time-series analysis is to model the asset • 𝜇𝑖 is the centroid of cluster i.
returns' temporal interdependence. Financial data may be
subjected to time-dependent patterns, volatility, and TABLE IV. ASSET CLUSTERS FOR ENHANCED DIVERSIFICATION
autocorrelation via the use of techniques such as GARCH Cluster Number Average Standard
(Generalised Autoregressive Conditional Heteroskedasticity) of Return Deviation
models and Autoregressive Integrated Moving Average Assets
(ARIMA) models[27]. Because these models may anticipate 1 25 0.056 0.023
market conditions and changes in asset prices, they are useful 2 30 0.042 0.031
... ... ... ...
for controlling portfolio risk and anticipating asset returns.
Table IV shows asset clusters that were created using K-
B. Classification Models for Risk Assessment Means and that improve diversification techniques. The
1) Support Vector Machines attributes of every cluster, including the quantity of assets,
Support Vector Machines (SVM) are robust classification average return, and standard deviation, provide valuable
models used for risk assessment. SVM aims to find a information for creating diversified investment portfolios.
hyperplane that best separates different classes of assets, 2) Hierarchical Clustering
aiding in risk classification[28]. The mathematical Hierarchical clustering organizes assets into a tree-like
formulation of SVM seeks to maximize the margin between hierarchy, enabling various levels of portfolio diversification.
the classes while minimizing classification errors. Kernel It provides a visual representation of asset similarity and can
functions, such as radial basis function (RBF) kernels, are be agglomerative (bottom-up) or divisive (top-down).
applied to handle non-linear classification tasks. Hierarchical clustering is beneficial in constructing diversified
portfolios by grouping assets based on their proximity in the
TABLE III. PORTFOLIO OPTIMIZATION METRICS COMPARISON
hierarchy.
Metric Traditional Machine
Model Learning D. Reinforcement Learning for Dynamic Portfolio
Model Balancing
Sharpe 0.65 0.78
Ratio
1) Q-Learning for Trading Strategies
Maximum -0.12 -0.08 Reinforcement learning leverages Q-Learning to optimize
Drawdown dynamic portfolio balancing. It formulates trading as a
Portfolio 0.025 0.018 Markov decision process, where an agent learns to take
Variance actions (portfolio adjustments) to maximize cumulative
... ... ... rewards (returns). The Q-value iteration algorithm computes
This table contrasts portfolio optimization metrics the optimal action-value function, guiding the agent to make
between a traditional model and a machine learning model. decisions that maximize long-term returns.
The machine learning model showcases superior performance
in terms of Sharpe ratio, maximum drawdown, portfolio 2) Deep Reinforcement Learning
variance, and more. Deep Reinforcement Learning (DRL) combines deep
learning with reinforcement learning for portfolio
2) Random Forests for Risk Analysis management. Deep Q-Networks (DQNs) or more advanced
Random Forests are ensemble methods that combine models like Proximal Policy Optimization (PPO) enable
multiple decision trees to assess risk. They can model complex agents to learn optimal portfolio allocation policies from
relationships and interactions among various risk factors. The historical data. DRL excels in handling high-dimensional state
algorithm's strength lies in its ability to capture non-linear spaces, making it valuable for complex portfolio optimization
tasks.

Authorized licensed use limited to: SRM University Amaravathi. Downloaded on December 12,2024 at 13:50:40 UTC from IEEE Xplore. Restrictions apply.
E. Ensemble Learning for Enhanced Predictive Accuracy mechanisms hold great promise in enhancing portfolio
Techniques for ensemble learning, such as boosting and optimization. Reinforcement learning, in particular, offers a
bagging (Bootstrap Aggregating), improve the prediction dynamic approach, where portfolios learn and adapt in real-
accuracy of asset return estimates. Bagging reduces variation time, making decisions to maximize returns and minimize
and improves stability by using many models via techniques risk.
such as Bootstrap aggregating. In contrast, boosting refines
TABLE V. DEEP LEARNING TECHNIQUES IN PORTFOLIO
predictions by repeatedly adjusting model weights and OPTIMIZATION
concentrating on examples that were incorrectly categorised.
Through the use of these ensemble methodologies, our Deep Learning Application in Advantages
Technique Portfolio
methodology endeavours to enhance the accuracy of asset Optimization
return forecasts, so contributing to the overall goal of attaining Recurrent Neural Time-dependent Captures temporal
superiority in portfolio optimisation [31]. Networks (RNNs) modeling, forecasting, dependencies, adapts to
and risk assessment. dynamic market
V. DEEP LEARNING AND NEURAL NETWORKS IN conditions.
PORTFOLIO OPTIMIZATION Convolutional Pattern recognition, Effective in detecting
Neural Networks trend identification, non-linear relationships
The infusion of deep learning and neural networks into (CNNs) and turning point in financial data.
portfolio optimization marks a transformative paradigm shift, detection.
offering a versatile toolset to decode complex financial data, Reinforcement Dynamic portfolio Real-time optimization,
understand non-linear relationships, and elevate the precision Learning (RL) management and capable of maximizing
of investment decisions. In this section, we navigate the adaptive decision- returns and minimizing
making. risk.
intricate realm of deep learning within the context of portfolio Generative Synthetic data Augments risk
optimization, exploring the potential and versatility of various Adversarial generation for risk management strategies
neural network architectures. Networks (GANs) assessment and and data augmentation.
diversification.
A. Understanding Non-Linear Relationships The prospects for deep learning in portfolio optimization
Deep learning's ability to represent complex non-linear are boundless, driven by the continuous evolution of neural
connections, which are often present in financial data, is one network architectures and the escalating availability of
of its fundamental benefits. Conventional linear models computing resources. Deep learning gives portfolio managers
struggle to capture the complexities seen in financial markets, powerful tools to efficiently traverse the complex and ever-
whereas deep learning models are better at recognising these changing world of finance by going beyond conventional
intricate relationships. This ability to comprehend non-linear linear models and embracing non-linearity. The combination
interactions is essential for identifying minute market trends of deep learning and neural networks has the potential to
and choosing wisely when making investments. completely transform portfolio management as this discipline
develops further, improving risk-adjusted returns, flexibility,
B. Application of Deep Learning Models and data-driven decision-making.
1) Recurrent Neural Networks (RNNs)
One kind of neural network that can handle sequential data VI. DISCUSSION AND ANALYSIS OF RESULTS
is called a recurrent neural network (RNN). RNNs may be In-depth discussion and analysis of the outcomes of
used to represent time-dependent interactions in the context of portfolio optimisation using both conventional and machine
portfolio optimisation, such as how asset values change over learning methods are covered in this section. It also examines
time[32]. Recurrent connections, which are the mathematical the models' sensitivity and robustness, pointing out the
foundation of an RNN, enable information transfer between drawbacks and difficulties in this complex field.
steps. This skill is essential for identifying temporal
relationships in financial data. Investors may estimate future A. Comparative Analysis of Traditional and Machine
asset values, evaluate volatility, and make dynamic portfolio Learning Approaches
modifications by using RNNs. Comparing machine learning-driven models with
conventional portfolio optimisation techniques is a key
2) Convolutional Neural Networks (CNNs)
component of our study. We examine their effectiveness with
Convolutional Neural Networks (CNNs), renowned for
respect to risk-adjusted returns, flexibility in response to
their proficiency in image analysis, have found application in
changes in the market, and asset selection accuracy. The risk-
financial time-series data. By treating historical price data as
adjusted return is measured using the Sharpe ratio, a widely
a form of image, CNNs can recognize patterns and trends that
used statistic for comparison. It is represented mathematically
may elude traditional models. The convolutional layers of
as follows:
CNNs apply filters to identify local patterns, enabling the
detection of market trends or turning points. The adaptability 𝑆ℎ𝑎𝑟𝑝𝑒 𝑅𝑎𝑡𝑖𝑜 =
𝑅𝑝 − 𝑅𝑓
(3)
of CNNs is particularly valuable in identifying inflection 𝜎𝑝
points and aiding in portfolio rebalancing.
C. Prospects for Advancements Where:
The fusion of deep learning and portfolio optimization is
an ever-evolving field with abundant prospects for • 𝑅𝑝 represent the portfolio's return.
advancement[33]. As computational power continues to surge • 𝑅𝑓 denote the risk-free rate.
and data accessibility expands, researchers are exploring more
sophisticated architectures. Reinforcement learning, • 𝜎𝑝 is the portfolio's standard deviations.
generative adversarial networks (GANs), and attention

Authorized licensed use limited to: SRM University Amaravathi. Downloaded on December 12,2024 at 13:50:40 UTC from IEEE Xplore. Restrictions apply.
 We compare the effectiveness of state-of-the-art machine Machine ✓ ✓ ✓
learning approaches with traditional mean-variance Learning-Driven
Models
optimisation, highlighting the latter's ability to identify non-
linear correlations and adjust to changing market In this updated table:
circumstances. Specifically, the machine learning models • Risk-Adjusted Return (Sharpe Ratio) for machine
show a stronger capacity to identify intricate patterns in learning-driven models is denoted by a checkmark
financial data, which might result in better risk-adjusted (✓), indicating their potential to achieve superior risk-
returns and more stability in erratic markets. adjusted returns.
TABLE VI. COMPARATIVE ANALYSIS OF PORTFOLIO OPTIMIZATION • Adaptability to Market Dynamics for machine
APPROACHES learning-driven models is also denoted by a
Approach Risk- Adaptability to Asset checkmark (✓), showcasing their ability to adapt to
Adjusted Market Selection changing market conditions.
Return Dynamics Precision
(Sharpe • Asset Selection Precision for machine learning-driven
Ratio)
Traditional X X X
models is represented by a checkmark (✓),
Mean-Variance emphasizing their capacity to make more precise asset
Optimization selections.

Fig. 1. Regression techniques for machine learning and prediction-based portfolio optimisation

Figure 1 illustrates a state-of-the-art method for portfolio synergy of machine learning with behavioral finance,
optimisation. This technique uses the mean-value-at-risk developing hybrid models that merge traditional finance
measure to manage portfolio risk using machine learning theories with machine learning, and advancing reinforcement
regression techniques to anticipate asset returns. In the learning techniques for dynamic portfolio balancing. As we
complex and ever-changing world of international stock conclude, our journey through portfolio optimization using
markets, this technique provides a dynamic and data-driven machine learning unveils a realm of possibilities, reshaping
approach to portfolio optimisation that improves investing investment strategies through adaptable, data-driven models.
decision-making. The evolving landscape of finance embraces these findings,
offering innovative and dynamic portfolio management
VII. CONCLUSION AND FUTURE DIRECTIONS approaches, marking a continuous journey towards redefining
Our main results demonstrate the advantages of machine financial decision-making in an increasingly intricate
learning-driven models over conventional mean-variance financial world.
optimisation in our investigation of portfolio optimisation
using machine learning approaches. The Sharpe ratio shows REFERENCES
that these sophisticated models provide improved risk- [1] Y. Rivenson, K. de Haan, W. D. Wallace, and A. Ozcan, “Emerging
adjusted returns, flexibility in responding to changing market advances to transform histopathology using virtual staining,” BME
Front., 2020.
circumstances, and accuracy in asset selection. The
[2] D. Settembre-Blundo, R. González-Sánchez, S. Medina-Salgado, and
ramifications for financial decision-making are significant, F. E. Garc\’\ia-Muiña, “Flexibility and resilience in corporate decision
providing investors, portfolio managers, and other financial making: a new sustainability-based risk management system in
professionals with a revolutionary move towards data-driven uncertain times,” Glob. J. Flex. Syst. Manag., vol. 22, no. Suppl 2, pp.
decisions. The capacity to precisely choose assets and adjust 107–132, 2021.
to changing economic conditions may have a big impact on [3] W. Hynes, B. Trump, P. Love, and I. Linkov, “Bouncing forward: a
the performance of investments. Transparency and data- resilience approach to dealing with COVID-19 and future systemic
shocks,” Environ. Syst. Decis., vol. 40, pp. 174–184, 2020.
driven methodologies are introduced by machine learning-
[4] X. Cheng et al., “Combating emerging financial risks in the big data
driven models, giving decision-makers a powerful toolset to era: A perspective review,” Fundam. Res., vol. 1, no. 5, pp. 595–606,
handle the intricacies of the financial markets. Looking to the 2021.
future, opportunities for further research emerge, such as [5] M. C. Pietronudo, G. Croidieu, and F. Schiavone, “A solution looking
enhancing model interpretability, integrating real-world for problems? A systematic literature review of the rationalizing
constraints like transaction costs and liquidity, exploring the influence of artificial intelligence on decision-making in innovation

Authorized licensed use limited to: SRM University Amaravathi. Downloaded on December 12,2024 at 13:50:40 UTC from IEEE Xplore. Restrictions apply.
 management,” Technol. Forecast. Soc. Change, vol. 182, p. 121828, [21] D. Kumar et al., “Smart school selection with supervised machine
2022. learning,” in Sustainable Smart Cities: Theoretical Foundations and
[6] S. K. Sahoo and S. S. Goswami, “A comprehensive review of multiple Practical Considerations, Springer, 2022, pp. 221–235.
criteria decision-making (MCDM) Methods: advancements, [22] P. Duboue, The art of feature engineering: essentials for machine
applications, and future directions,” Decis. Mak. Adv., vol. 1, no. 1, pp. learning. Cambridge University Press, 2020.
25–48, 2023. [23] M. Chandra, T. Shrimali, and A. Gupta, “A survey: Recent
[7] H. Guan and Z. An, “A local adaptive learning system for online development in fractals,” in 2012 Fourth International Conference on
portfolio selection,” Knowledge-Based Syst., vol. 186, p. 104958, Computational Intelligence and Communication Networks, 2012, pp.
2019. 251–256.
[8] G. Tsaousoglou, J. S. Giraldo, and N. G. Paterakis, “Market [24] W. Long, Z. Lu, and L. Cui, “Deep learning-based feature engineering
mechanisms for local electricity markets: A review of models, solution for stock price movement prediction,” Knowledge-Based Syst., vol.
concepts and algorithmic techniques,” Renew. Sustain. Energy Rev., 164, pp. 163–173, 2019.
vol. 156, p. 111890, 2022. [25] A. Gupta and M. K. Gupta, “HIVE-processing structured data in
[9] F. Woebbeking, “Cryptocurrency volatility markets,” Digit. Financ., Hadoop,” Int. J. Sci. \& Eng. Res., vol. 8, no. 6, pp. 45–48, 2017.
vol. 3, no. 3–4, pp. 273–298, 2021. [26] M. Mallikarjuna and R. P. Rao, “Evaluation of forecasting methods
[10] J. Zaidi, “Unlearning for change? Conceptualising the phenomenon of from selected stock market returns,” Financ. Innov., vol. 5, no. 1, pp.
organisational unlearning in different types of organisational change.,” 1–16, 2019.
The University of Auckland, 2023. [27] T. C. Mills, Applied time series analysis: A practical guide to modeling
[11] I. MacNeil and I. Esser, “From a financial to an entity model of ESG,” and forecasting. Academic press, 2019.
Eur. Bus. Organ. Law Rev., vol. 23, no. 1, pp. 9–45, 2022. [28] R. Rodr\’\iguez-Pérez and J. Bajorath, “Evolution of support vector
[12] P. Yu, J. S. Lee, I. Kulyatin, Z. Shi, and S. Dasgupta, “Model-based machine and regression modeling in chemoinformatics and drug
deep reinforcement learning for dynamic portfolio optimization,” discovery,” J. Comput. Aided. Mol. Des., vol. 36, no. 5, pp. 355–362,
arXiv Prepr. arXiv1901.08740, 2019. 2022.
[13] Y. Ma, R. Han, and W. Wang, “Portfolio optimization with return [29] A. Arabameri, M. Yamani, B. Pradhan, A. Melesse, K. Shirani, and D.
prediction using deep learning and machine learning,” Expert Syst. T. Bui, “Novel ensembles of COPRAS multi-criteria decision-making
Appl., vol. 165, p. 113973, 2021. with logistic regression, boosted regression tree, and random forest for
[14] J. Breeden, “A survey of machine learning in credit risk,” J. Credit spatial prediction of gully erosion susceptibility,” Sci. Total Environ.,
Risk, vol. 17, no. 3, 2021. vol. 688, pp. 903–916, 2019.
[15] X. Liu, J. Zhang, and Z. Pei, “Machine learning for high-entropy alloys: [30] F. K. Y. Amevenku, R. K. Asravor, and J. K. M. Kuwornu,
Progress, challenges and opportunities,” Prog. Mater. Sci., vol. 131, p. “Determinants of livelihood strategies of fishing households in the
101018, 2023. volta Basin, Ghana,” Cogent Econ. \& Financ., 2019.
[16] V. Singh, S.-S. Chen, M. Singhania, B. Nanavati, A. Gupta, and others, [31] Verma , P. ., & Tyagi, D. P. . (2021). Credit Card Fraud Detection using
“How are reinforcement learning and deep learning algorithms used for Selective Class Sampling and Random Forest Classifier. SPAST
big data based decision making in financial industries--A review and Abstracts, 1(01). Retrieved from
research agenda,” Int. J. Inf. Manag. Data Insights, vol. 2, no. 2, p. https://spast.org/techrep/article/view/3528
100094, 2022. [32] B. Lim, S. Zohren, and S. Roberts, “Enhancing time series momentum
[17] S. J. Yoo, Y. H. Gu, and others, “Safety AARL: Weight adjustment for strategies using deep neural networks,” arXiv Prepr. arXiv1904.04912,
reinforcement-learning-based safety dynamic asset allocation 2019.
strategies,” Expert Syst. Appl., vol. 227, p. 120297, 2023. [33] Chaudhary, M., Gangele, A., Naved, M., Gaur, L., & Singh, G. (2022,
[18] M. Abbasi et al., “Designing optimized drug candidates with November). The Function of Driver Categorisation in the Ride-Hailing
Generative Adversarial Network,” J. Cheminform., vol. 14, no. 1, p. Industry: A Study on On-Demand Transport. In 2022 3rd International
40, 2022. Conference on Computation, Automation and Knowledge
Management (ICCAKM) (pp. 1-6). IEEE.
[19] S. D. Mohanty, D. Lekan, T. P. McCoy, M. Jenkins, and P. Manda,
“Machine learning for predicting readmission risk among the frail: [34] Sharma, M., Sharma, V., & Kapoor, R. (2022). Study of E-Commerce
Explainable AI for healthcare,” Patterns, vol. 3, no. 1, 2022. and Impact of Machine Learning in E-Commerce. In Empirical
Research for Futuristic E-Commerce Systems: Foundations and
[20] N. Yang, S. Ruan, L. Han, H. Liu, L. Guo, and C. Xiang, Applications (pp. 1-22). IGI Global
“Reinforcement learning-based real-time intelligent energy
management for hybrid electric vehicles in a model predictive control
framework,” Energy, vol. 270, p. 126971, 2023.

Authorized licensed use limited to: SRM University Amaravathi. Downloaded on December 12,2024 at 13:50:40 UTC from IEEE Xplore. Restrictions apply.