AI

Article

Cybersecure XAI Algorithm for Generating Recommendations

Based on Financial Fundamentals Using DeepSeek
Iván García-Magariño 1,2, * , Javier Bravo-Agapito 1 and Raquel Lacuesta 3

1 Department of Software Engineering and Artificial Intelligence, Complutense University of Madrid,

28040 Madrid, Spain; [Link]@[Link]
2 Instituto de Tecnología del Conocimiento, Complutense University of Madrid, 28040 Madrid, Spain
3 Department of Computer Science and Engineering of Systems, University of Zaragoza, 44003 Teruel, Spain;
lacuesta@[Link]
* Correspondence: igarciam@[Link]; Tel.: +34-913947500

Abstract: Background: Investment decisions in stocks are one of the most complex tasks
due to the uncertainty of which stocks will increase or decrease in their values. A diversified
portfolio statistically reduces the risk; however, stock choice still substantially influences the
profitability. Methods: This work proposes a methodology to automate investment decision
recommendations with clear explanations. It utilizes generative AI, guided by prompt
engineering, to interpret price predictions derived from neural networks. The methodology
also includes the Artificial Intelligence Trust, Risk, and Security Management (AI TRiSM)
model to provide robust security recommendations for the system. The proposed system
provides long-term investment recommendations based on the financial fundamentals
of companies, such as the price-to-earnings ratio (PER) and the net margin of profits
over the total revenue. The proposed explainable artificial intelligence (XAI) system
uses DeepSeek for describing recommendations and suggested companies, as well as
several charts based on Shapley additive explanation (SHAP) values and local-interpretable
model-agnostic explanations (LIMEs) for showing feature importance. Results: In the
experiments, we compared the profitability of the proposed portfolios, ranging from 8 to
28 stock values, with the maximum expected price increases for 4 years in the NASDAQ-
100 and S&P-500, where both bull and bear markets were, respectively, considered before
Academic Editor: Xianrong and after the custom duties increases in international trade by the USA in April 2025. The
(Shawn) Zheng
proposed system achieved an average profitability of 56.62% while considering 120 different
Received: 26 March 2025 portfolio recommendations. Conclusions: A t-Student test confirmed that the difference in
Revised: 22 April 2025
profitability compared to the index was statistically significant. A user study revealed that
Accepted: 29 April 2025
the participants agreed that the portfolio explanations were useful for trusting the system,
Published: 2 May 2025
with an average score of 6.14 in a 7-point Likert scale.
Citation: García-Magariño, I.;
Bravo-Agapito, J.; Lacuesta, R.
Keywords: DeepSeek; explainable artificial intelligence; financial fundamentals; investment
Cybersecure XAI Algorithm for
Generating Recommendations
recommendation; stock price prediction; XAI security
Based on Financial Fundamentals
Using DeepSeek. AI 2025, 6, 95.
[Link]

Copyright: © 2025 by the authors.

1. Introduction
Licensee MDPI, Basel, Switzerland. Although fixed-income deposits offer security and stability, their profitability is signifi-
This article is an open access article cantly lower compared to alternatives, such as variable incomes based on stock investments.
distributed under the terms and
However, stock investments are inherently volatile, with potential for both gains and losses,
conditions of the Creative Commons
particularly in the short term. The careful selection of stock assets substantially influences
Attribution (CC BY) license
([Link]
the overall profitability of investment portfolios [1]. In this scenario, investors highly appre-
licenses/by/4.0/). ciate proper investment recommendations, which may stem from either human expertise or

AI 2025, 6, 95 [Link]
AI 2025, 6, 95 2 of 27

automated systems, such as robo-advisors. Indeed, numerous studies have confirmed that
computerized methods, e.g., investments in passive Exchange-Traded Fund (ETF) tracking
benchmark indices such as Standard & Poor’s 500 (S&P 500) and NASDAQ-100, typically
outperform actively managed ETFs overseen by human experts [2]. These passive ETFs are
managed using straightforward rules based on market capitalization.
In the contemporary era of advanced artificial intelligence (AI), researchers have
explored innovative approaches for recommending stock investments. For example,
Pelster et al. [3] investigated whether the intelligent chat bot ChatGPT could assist in stock
selection. Their findings indicate that ChatGPT-4, when equipped with internet access,
successfully provided useful recommendations that yielded interesting dividend returns.
It is worth noting that dividend yields are relatively predictable based on historical data
(e.g., king stocks have consistently provided increasing dividend values over decades).
These developments reflect a growing interest in using modern AI technologies, such as
generative AI-based chat systems, for effective stock selection in investment strategies.
Furthermore, stock price prediction is widely regarded as one of the most challeng-
ing tasks in financial analysis due to its inherent complexity and significant uncertainty.
Nevertheless, markets such as S&P 500 have demonstrated consistent average growth over
decades [4]. Thus, investors recognize that the probability of achieving positive returns
favors long-term investment, though they also know that successful outcomes depend on
the careful selection of stocks and optimal timing.
Novice and experienced investors not only need a proposed stock investment portfolio
with some expectations, but they also need to understand the analysis behind and the most
relevant information of each of the companies. Our approach provides such explanations
with some of the well-known charts from XAI and also a proper description of each
company, providing an explained portfolio with a technique that reduces individual stock
investment risk with a diversified portfolio.
In this study, we propose an automated approach that integrates a traditional machine
learning (ML) method; specifically, it is a multi-layer perceptron (MLP) neural network,
where explainable artificial intelligence (XAI) is used to generate investment portfolio
recommendations. Additionally, we incorporate DeepSeek, a tool of generative artificial
intelligence (GAI), to enhance the interpretability of our system’s recommendations. This is
achieved through prompt engineering and template-based text generation, which provide
detailed explanations of the outputs. In addition, this work evaluates the application of
GAI within the framework of XAI. Moreover, the XAI recommendations must be protected
to guarantee the security of the information proposed by the system.
The main objective of this research is to provide a portfolio accompanied by compre-
hensive information, thereby fostering investor trust in these recommendations. Therefore,
our system provides transparent and interpretable insights into the decision-making pro-
cess behind each recommendation. This transparency not only enhances user confidence in
the AI-driven suggestions, but it also enables investors to better understand the rationale
underpinning the proposed strategies, thereby supporting informed decision making in
stock investment.
The contributions of this research can be summarized as follows: (1) providing a
dataset containing key financial metrics of the last five years; (2) providing an explainable
investment recommendation system to recommend stock portfolios, utilizing both XAI and
GAI tools; (3) bridging the gap between XAI and GAI tools and showcasing their combined
potential in the context of the financial decision-making process; and (4) introducing a
model based on the Artificial Intelligence Trust, Risk, and Security Management (AI TRiSM)
framework to provide an integrated approach to trust, risk, and security management.
AI 2025, 6, 95 3 of 27

Thus, our system delivers clear, transparent, and secure explanations, enhancing user
understanding and trust in the decision-making process.
This paper is structured as follows: Section 2 provides the research context, Section 3
details the methodology, Section 4 introduces the evaluation method, Section 5 presents the
results, and Sections 6 and 7 offer a discussion of the findings and the main conclusions,
respectively.

2. Related Work
AI is frequently employed for stock recommendation systems. For instance, Kang
et al. [5] introduced an AI-based approach for rebalancing highly volatile financial portfo-
lios. In particular, ML techniques are commonly applied to stock market prediction, but
there are still challenges related to non-linear methods [6]. Despite these advancements,
only a limited number of studies address the use of GAI for stock market forecasting. One
of the objectives of our work is to provide insights into the application of GAI for stock
market forecasting.
The integration of AI into financial markets has significantly transformed the land-
scape of stock market forecasting, especially in the area of generative modeling and XAI.
In this section, we review the literature on the application of GAI for stock investment
recommendations and the use of XAI techniques to enhance the explanation of predictions.
Moreover, in the financial markets, trust, risk, and security management need to be
considered when working with XAI systems. Optimizing AI system security requires a
comprehensive approach that integrates the proper security strategies. In this case, some
models, such the TRiSM framework, can be applied [7].
By examining these two research areas, this study seeks to integrate the innovative
potential of GAI with the transparency offered by XAI, applying the TRiSM framework
to propose a methodology that guarantees trust, risk, and security management in these
scenarios, thereby contributing to the development of more robust and reliable stock
investment recommendation systems.

2.1. GAI in Stock Investment Recommendation

In recent years, several researchers have explored the use of AI, especially GAI, to
directly construct, manage, or optimize investment portfolios [3,8–10]. Others have con-
centrated their efforts on developing methods to generate synthetic financial data [11],
enhance financial advisory services [12], conduct sentiment analysis of unstructured finan-
cial data [13], or improve operational capabilities [14].
Aiche et al. [8] employed an innovative AI-driven approach to build and manage a
portfolio of cybersecurity stocks from 2018 to 2024. They collected structured financial data,
such as historical stock prices, financial ratios, and quarterly earnings, from sources such as
Bloomberg. They used financial metrics, such as the trailing P/E ratio (also known as PER),
price-to-earnings-growth (PEG), return on equity (ROE), and forward PER, to prioritize
stocks. Additionally, they incorporated unstructured data from financial news (Yahoo
Finance, [Link], and MarketWatch) and social media (X, Reddit, and StockTwits).
They used sentiment analysis to identify the companies with high investor confidence and
market perception. The selection process filters by market capitalization and applies criteria,
such as revenue growth, sentiment score, forward PER, PEG ratio, and ROE. Subsequently,
they employed ML models (Random Forest and SVM) and fundamental analysis to predict
the direction of the stocks. Predictions are generated through a weighted combination: 40%
from the Random Forest and SVM forecasting models, 25% from sentiment analysis, 20%
from fundamental metrics, and 15% from revenue growth. They created the final portfolio
using mean-variance optimization (MVO) and rebalanced quarterly with assistance from
AI 2025, 6, 95 4 of 27

GAI (ChatGPT-40). They evaluated the performance of the portfolio using the Sharpe
ratio and maximum drawdown. Their results outperformed the returns of traditional
cybersecurity investments.
Advancing the use of GAI for portfolio recommendations, Kim [9] examined the
integration of ChatGPT into a quantitative investment framework, acting as a quant asset
manager. They collected 10 years of data from sources such as Federal Reserve Economic
Data and Yahoo Finance. ChatGPT is used to construct the investment portfolio by rec-
ommending asset classes based on economic conditions, replacing human asset managers.
While they use ChatGPT to understand macroeconomic relationships between economic
conditions and asset class movements, they do not employ it to assess current market
conditions. ChatGPT recommendations are evaluated using equally weighted, global
minimum-variance, and risk-parity portfolio strategies, demonstrating improved portfolio
efficiency compared to traditional models.
Similarly, Pelster et al. [3] investigated whether ChatGPT-4, enriched with additional
information from a web search engine, can provide valuable investment advice by pre-
dicting earnings surprises and evaluating the attractiveness of stocks. They proposed two
prompts: one for predicting earnings and another for assessing the relative attractiveness
of each S&P 500 company. They developed an application that automates a conversation
with ChatGPT via the WebChatGPT Chrome extension, entering prompts about earnings
surprises or stock attractiveness. Recommendations are evaluated using the mean return
and Sharpe ratio, finding a positive correlation between ChatGPT ratings and future stock
returns. This suggests that ChatGPT can successfully identify stocks that outperform.
In the same vein, Huang et al. [10] explored stock selection and portfolio optimization
using a structured prompt framework with GAI (ChatGPT) and a novel bio-inspired
algorithm. They proposed a structured framework composed of strategy (basic guidelines
for the stock selection process), goals (investment objectives), and constraints (investors’
restrictions). They evaluated their results using risk, expected return (ER), and the Sharpe
ratio. Although their findings are promising, they have pointed out that a weakness of AI
recommendations is that they do not have real-time market analysis.
While the aforementioned studies focus on direct portfolio management, other re-
search has utilized AI, particularly GAI, as an analytical tool or to generate synthetic data
that enhances financial understanding. For instance, Ante et al. [15] applied Natural
Language Processing (NLP) to analyze the 10-K filings of NASDAQ companies (2010–2022),
classifying AI exposure using binary and weighted scores based on the frequency of
AI-related terms. Their approach generated four indices, though without AI-driven opti-
mization. Their methodology prioritizes objective classification over predictive portfolio
construction. Nevertheless, their findings reveal a rising trend in AI mentions over time
and significant positive abnormal returns for companies with high AI exposure following
the emergence of ChatGPT. In contrast, Takahashi et al. [11] used GAI to generate synthetic
financial time series for modeling market dynamics, rather than focusing on proposing
investment portfolios.
Other studies have assessed AI capabilities in financial sentiment analysis and advisory
roles, offering broader insights into its potential. Muhammad et al. [13] evaluated the
effectiveness of traditional sentiment analysis techniques compared to ChatGPT and Gemini
in the financial domain. They created a dataset of 1476 news headlines from the Bloomberg
financial news service, which were enriched with manual sentiment score annotations.
They evaluated the effectiveness of each of the conventional sentiment analysis models,
ChatGPT and Gemini, using classification metrics, such as precision, recall, F1-measure, and
accuracy. Their results showed that ChatGPT slightly outperformed the highly optimized
conventional method.
AI 2025, 6, 95 5 of 27

Huang et al. [12] investigated the role of GAI in improving robo-advisers (RA) for
financial advisory services, concluding that AI-driven financial advisors can analyze market
conditions and provide insights into future trends.
Finally, Jackson et al. [14] explored the transformative role of AI and GAI in supply
chain and operations management. These authors used GAI to improve operational capa-
bilities (e.g., synthetic data for demand forecasting), but not to create investment portfolios.
Our research differs from the previous related works by integrating XAI and GAI tools
to create a transparent and explainable investment recommendation system. While studies
such as [3,8,9] utilized GAI for portfolio management, they did not incorporate XAI to
enhance user understanding and trust. Additionally, our approach provides a dataset with
key financial metrics, which was not a primary focus in the aforementioned studies.

2.2. XAI in Stock Investment Recommendation

The integration of XAI in stock investment recommendations to enhance both the
accuracy and interpretability of predictions has been explored. For instance, Ferrara
et al. [16] investigated the combination of ML and XAI techniques to improve the accuracy
and interpretability of stock price predictions. They focused on companies with strong
environmental sustainability commitments, such as Tesla, NextEra Energy, Vestas, and Enel,
using data from Yahoo Finance. Their approach involves training short-term stock price
prediction models using Random Forest and XGBoost with variables like Open, High, Low,
and Volume reflecting market behavior. To understand the contribution of each variable,
they employ XAI techniques, such as Shapley additive explanations (SHAPs) and local
interpretable model-agnostic explanations (LIMEs). Notably, SHAP analysis revealed that
the high and open variables significantly impacted Random Forest predictions, while the
high variables were dominant in XGBoost.
In contrast, Wang et al. [17] developed a deep learning model based on genera-
tive adversarial networks to predict stock market returns and to construct a portfolio.
Although they did not use XAI techniques like SHAP and LIME, their model provides
insights into key predictors, such as momentum, liquidity, and fundamentals. To test their
model, they collected data on stocks in the Chinese market from the Wind platform. Their
model outperformed traditional and ML methods and portfolio performance (achieving
a 23.52% annualized return and a Sharpe ratio of 1.29). Their work contributes to both
asset pricing literature and explainable AI, offering practical insights for investors, analysts,
and regulators.
Our research differs from existing XAI studies by combining XAI tools with GAI to
create a transparent and explainable investment recommendation system. While works
like [16] use XAI to enhance the interpretability of stock price predictions, they do not
integrate GAI to broaden its scope in the financial decision-making process. Additionally,
our approach uses the fundamental financial information of companies, as shown in our
dataset with key financial metrics, which was not a primary focus in the aforementioned
works. This allows for a more comprehensive understanding of the investment process.
The novelty of the proposed system lies in its ability to not only provide portfolio
recommendations, but also deliver explanations of these recommendations.

3. Methodology
The proposed methodology includes the use of ML techniques to provide an invest-
ment portfolio recommendation. In addition, these recommendations are enriched with
XAI and the proper explanations provided by DeepSeek. Our predictions were evaluated
using the t-Student test, and the generated AI explanations were evaluated using GAI tools,
such as ChatGPT, Gemini, Grok, and Qwen. We explain, in more detail, each step of our
AI 2025, 6, 95 6 of 27

proposed methodology in the following lines and later on when we present the evaluation
method in Section 4.
• Step 1. Data collection: This step comprises collecting the financial fundamentals
PER and Net Profit Margin (NPM) from technological companies using the Yahoo
Finance service for the last five years. We implemented an ad hoc Python API to
retrieve these data from Yahoo Finance. More details about the data collection are
explained in Section 3.2. As a result of this step, we provide the generated dataset,
which can be downloaded from the Mendeley dataset (“Dataset about stock funda-
mentals and later stock price increases in a four-year period”, Mendeley Data, V1,
DOI: 10.17632/5jk4bm7x5v.1).
• Step 2. Prediction of an investment portfolio: This step is composed of the typical
steps of knowledge discovery, i.e., preprocessing, modeling, and evaluation. In
preprocessing, the dataset is separated into input data (X) and output data (Y), where
X contains two columns, PER and NPM, and Y contains a percentage of the increased
value. Descriptive statistical analysis is necessary to check whether transformations
(e.g., normalization) are necessary. Afterward, the dataset is split into a training
and a testing set. Modeling includes generating an initial model with the training
dataset. In this stage, some of the most common ML approaches, such as MLP, k-
Nearest Neighbors, and Support Vector Machine (SVM) are tested. An MLP neural
network was chosen due to its flexibility and potential prediction feasibility. A k-
fold cross-validation strategy is used to predict the optimal hyperparameters of MLP.
Once this step is completed, an MLP is generated by taking into account the optimal
hyperparameters, and it is then trained with the training dataset. We used different
numbers of neurons, ranging from 25 to 300 neurons and from 2 to 3 layers, and
we picked the combination of 2 layers of 150 neurons each as the one that avoided
overfitting and provided an average profitability over the average of the index.
• Step 3. Explanations of an investment portfolio: This step includes the use of GAI tools
to provide more information about the selected companies in the investment portfolio
recommendation. In this research, we use the DeepSeek tool and prompt engineering.
• Step 4. Design of the security protocols: This step uses the AI TRiSM model for
managing security and risk. This step has two sub-phases concerning, respectively,
the design of the security protocol and the simulation of potential attacks.
Each of the most relevant aspects of the methodology are described in each of the
following subsections.

3.1. System Architecture

The proposed methodology recommends developing a system with a certain architec-
ture. Figure 1 shows an overview of the system architecture. Green boxes represent external
API or libraries, blue boxes denote internal system modules, and pink boxes indicate system
outputs. On the left, the Data Collector module retrieves data from the Yahoo Finance
API, extracting key fundamental financial values for companies with a given index. This
module stores the data into a dedicated financial data repository. The ML Model module
contains a ML model trained on this repository to forecast company stock prices four years
ahead. The Portfolio Recommender module integrates the outputs from the ML model, the
XAI module, and the Prompt Generator. Within this module, the ML model is used for
picking up a portfolio of a given number of companies. The Prompt Generator is used for
generating prompts for the purpose of obtaining a description of the most relevant aspects
of each company, which are obtained thanks to DeepSeek. The XAI Security module uses
AI TRiSM to ensure security. The XAI module is used for providing SHAP charts using
the SHAP technique and local interpretable model-agnostic explanation (LIME) charts to
AI 2025, 6, 95 7 of 27

describe the relevance of each fundamental financial value in the decisions. Together, these
components enable the Portfolio Recommender to deliver a recommended portfolio with a
description of the portfolio and explanations about how the decision was reached when
suggesting such a portfolio.

Figure 1. Overview of the system architecture.

3.2. Dataset
The dataset was collected from the source of the Yahoo Finance service. This dataset
serves as a valuable resource, offering a solid and reliable foundation for conducting
financial analysis and supporting data-driven decision making.
The most relevant financial metric of stock fundamentals is PER, which is the pro-
portion of the price of the stock divided by the total earnings. This shows undervalued
stocks. However, being undervalued might also mean that there might be a reason for a
real-problem that makes the market undervalue a specific stock.
Another relevant aspect is NPM, which represents the ratio of net profits divided by
all the incomes of the company. This margin is important to know how much of the sales
are profit. This also depends on the sector, since, for example, technology and consumer
goods companies usually have completely different margins.
Dividend yield shows the percentage of the stock price that is provided in dividends.
Normally, dividends show that the company is so profitable that it can distribute its profits.
This is very important to consider with the payout ratio, which represents the percentage
of the earnings that are distributed in dividends. Another interesting aspect is the dividend
growth rate, which is used to see how much the dividend is growing per year (normally
calculated within the last five years).
Cashflow indicates how much liquid money is generated in the company for a year
per action. It can distinguish between operative cashflow (not considering financing costs),
investment cashflow (that considers investments of a company in financial services, real
estates, and other properties), and financing cashflow (considering financing investments
and payments).
AI 2025, 6, 95 8 of 27

Other relevant metrics are return on investment (ROI) measures as the NPM is divided
by the cost of investment.
All these financial metrics are calculated from the last five years, and it is also important
to see them in perspective.
Regarding the management of missing values, we only collected the stock values from
which all the information was collected to only recommend stock investments based on
stocks with full fundamental financial information. In the data collection, if this happened,
our program informed which stock values were discarded for this reason. We also consid-
ered the possibility of replacing the missing values with the average values of their sectors
(notice that some fundamental values depend very much on the sectors). However, the
automatic classification of stock values into a sector was not a straightforward problem
since some companies belonged to several sectors and the result of this classification could
alter the final price predictions.
The dataset also included the price of the stock four years before the experiment
(February 2021) and the price during the experiment (February 2025), as well as the price
increase. The first published version of the dataset only included PER and NPM, both
of the mentioned prices, and the percentage increase in the four-year period since these
were the values that were actually used in our first experiment. Later, we included more
data concerning the other financial ratios about dividend yield, payout ratio, cashflow
and ROI. We also included information from April 2021 to April 2025, where the latter
was understood as a different scenario due to the impact of custom USA duties on the
global economy. The financial ratios were obtained by obtaining information through the
Yahoo Finance library in Python and applying the required equations for calculating the
corresponding ratios.

3.3. Analysis About the Financial Ratios for Automatic Recommendations

This approach focuses on diversified value investment based on the known data from
companies that are usually useful in long-term investments (i.e., a period of several years).
For this purpose, our automated system is based on the fundamental financial aspects
of companies.
To begin with, this approach considers the PER and NPM. Intuitively, PER represents
the number of years such that the companies earn as much as their initial costs, where
both the price and this ratio are frozen over time. From a basic and abstract viewpoint,
a low PER should represent buying opportunities for its inherent profitability. However,
this assumption is naive and far from the truth, as the last four years of history shows in
Figure 2. This figure presents the percentage of increase ratio in the stock price month of
our initial experiments (February 2025) over the prices that were applicable during the four
years before the experiment, as well as those related to the PER values four years ago, for
the companies in the S&P 500. As one can observe, there was no clear relation between
these two concepts, at least when considering the below chart.
The proposed approach also considers the NPM, which represents the percentage of
profits out of total revenue. Figure 3 shows the scatter plot of all the S&P 500 companies
regarding their 4-year price increases in relation to the NPM at the beginning of this period.
This figure does not show a clear relationship between these two variables.
Moreover, this work also considered dividend yield, payout ratio, ROI, and operating
cashflow per share as an expansion for considering the most relevant features.
AI 2025, 6, 95 9 of 27

Figure 2. The relation of PER with the next 4 years of price increases.

Figure 3. Relation of the net margin with the next 4 years of prices.

Figure 4 shows the correlation among the different features. One can observe that PER
was inversely slightly correlated with the dividend yield, showing that low PER is related
to high “dividend yield”. This correlation is relatively low in absolute value, although
any correlation is difficult to obtain in stock price predictions given the high levels of
uncertainty in this domain. The NPM was significantly correlated with the “dividend
yield”, “ROI”, and “operating cashflow”, so it provided different information from the
PER. Price increase was significantly correlated with dividend yield (inversely), operating
cashflow (directly), and price now (directly).
We created a feature selection mechanism from these variables. For this purpose, we
applied a standard scalar to the input fundamental variables, and we then later applied
a variance threshold selection mechanism. We observed that, with a threshold of 0.25,
all six of the variables were selected, so we used all of them for our recommendations
and explanations.
Figure 5 shows the distribution of companies when considering the PER, and it shows
whether these had either positive or negative price increases over the next four years. Most
companies had a PER between 10% and 40%.
AI 2025, 6, 95 10 of 27

Figure 4. Correlation matrix heatmap of the fundamentals of stocks (* significant with .sig < 0.05)

Figure 5. Distribution of the S&P 500 of the PER when distinguishing between the 4-year positive
and negative increases.

Figure 6 shows the relative positive increase (red) in comparison to the amount of each
interval. It was observed that the intervals with 100% of price increases over decreases (i.e.,
0–5, 40–45, 75–80, and 85–90) had very few cases, so these did not seem to be relevant. In
fact, intervals 70–75, 80–85, and 90–95 had the exact opposite result. Thus, it seems safer to
invest in companies with a score from 0 to 29 than in other companies.
Firstly, we decided that the most appropriate feature normalization was to subtract
the mean and divide the result by the standard deviation, but some stock values had so
many different values that they altered the total results. The common way of proceeding
would have been to consider these stock values as outliers and to keep them out of our
analysis. However, these outliers included companies as prestigious as NVIDIA, who
are recommended by many technical experts, so we decided to avoid leaving the outliers
out. Therefore, our presented results use the original values without normalization for the
aforementioned reasons.
AI 2025, 6, 95 11 of 27

Figure 6. The relative distribution of the S&P 500 of the PER when distinguishing between 4-year
positive and negative increases.

Since negative PER and NPM revealed losses, and as we believed that recommending
companies with losses is very risky and could hinder the trust of our users in our recom-
mender system, we decided to filter out all of the possible recommendations with either a
negative PER or negative NPM.

3.4. XAI System for Recommendation

DeepSeek is competing against ChatGPT in GAI. One of their advantages is that
it is freely open, although its maturity is lower than ChatGPT given their duration in
the market.
The proposed XAI system is based on enhancing the information about companies
inside the investment portfolio recommendation. This information is enhanced using
prompt engineering with DeepSeek. This task includes the implementation of the Python
code of some prompts that were used in DeepSeek to provide meaningful descriptions.
Regarding prompt engineering, the system calculated the recommendations using a
combination of XAI and prompt engineering. It was used for the following steps:
• Providing a description of the company.
• Indicating which financial data were used.
• Described the most relevant factors taken into account based on SHAP outputs.
Note that the system first generates the following description:
This system provides investment recommendations based on a neural network system
that has been trained with the information of the last four years of the financial balances
of companies. This recommendation is intended for value investors, and, consequently,
these recommendations are for long-term investments (e.g., around four years). More
concretely, this system uses PER (price-to-earning ratio), NPM (net profit margin),
dividend yield, ROI (return of investment), and operating cashflow per share. NPM is
a financial metric that represents the percentage of a company’s revenue that remains
as profit after all expenses, taxes, and costs have been deducted. Up to the date, there
is no way to predict the value of stocks with certainty. The predicted price increases
are only raw estimations based on ML techniques that should be regarded as simply
uncertain indicators that show probable increases. The portfolios use the principle of
AI 2025, 6, 95 12 of 27

diversification, in which picking up several stocks reduces the impact of the risk where a
company decreases its price due to very particular circumstances.
Below, one can observe an example of a prompt used in DeepSeek for describing a
financial ratio:
Describe in one sentence the meaning of “net margin” of a company
The system also introduces the reference index on which the portfolio is based
as follows:
This system provides recommendations based on specific sectors since financial ratios
are usually comparable in the same field. In particular, the following portfolio rec-
ommendation is based on the NASDAQ-100 index. The NASDAQ-100 is a market-
capitalization-weighted index of the 100 largest non-financial companies listed on the
NASDAQ stock exchange, primarily featuring technology and growth-oriented firms. The
NASDAQ-100 is a market-capitalization-weighted index of the 100 largest non-financial
companies listed on the NASDAQ stock exchange, primarily featuring technology and
growth-oriented firms.
An example of a prompt when generating the introduction of an index:
Describe in one sentence the NASDAQ-100
In the particular case of indexes, we tested both ChatGPT and DeepSeek, and, in this
specific index, we used the ChatGPT description. The approach used human supervision
since it can be selected before users use the system, and it remains the same for each
particular index.
Figure 7 shows the recommended investment portfolio by the system. It provides
the ticker as each stock-valued identifier. For example, “CHTR” is the ticker of “Charter
Communications, Inc.” company. It also presents the fundamental financial information of
the PER and NPM used for performing the estimations. Finally, it presents the predicted
increase percentages over 4 years, which are most likely different from the actual price
increases but which help conform a portfolio with a high profitability on average.
Figures 8 and 9 present the description of the most relevant information of the compa-
nies of the proposed investment portfolio, and these were generated utilizing GAI through
DeepSeek with the prompt generated by our system for the given portfolio.
To complete the explanation, our system provides an explanation according to SHAP
values of the given index used as a reference for the fundamental financial features in the
trained ML model. It is worth mentioning that obtaining a global SHAP chart is especially
useful in linear ML models with feature independence, but, in other cases such as neural
networks, its usefulness is arguable, and many researchers would find it useless for not
fulfilling the basic assumptions of SHAP. However, a recent work [18] has shown that
SHAP information can be useful in neural networks. In particular, its authors proved
how they used SHAP interpretations to effectively select features when applying deep
learning, specifically in convolutional neural networks (CNNs). Being aware of the known
limitations of applying SHAP to ML models with dependent features [19], Figure 10 shows
the SHAP explanation for the globally trained MLP regressor model when using 70% of
the S&P-500 for training. One can observe that, in the general system, the NPM, ROI, and
PER are probably having (in that particular order) high influences on the predicted system.
AI 2025, 6, 95 13 of 27

Figure 7. A portfolio recommended by the system.

Figure 8. Descriptions of the first five companies of the portfolio that was recommended by the
system when assisted by DeepSeek.
AI 2025, 6, 95 14 of 27

Figure 9. Descriptions of the last five companies of the portfolio that was recommended by the
system when assisted by DeepSeek.

Figure 10. The SHAP bar plot for the global learned MLP model.

The proposed system also provides an explanation for each recommendation, as one
can observe in the explanation SHAP bar plot for the Apple company shown in Figure 11.
This example shows that a specific case may have a completely different influence from
the general model. More concretely, in the particular case of Apple, its dividend yield was
found to be the most influential aspect.
The system also provides a distribution of local explanations with a SHAP Beeswarm
plot, as one can observe, as an example, for the same MLP regressor model on the S&P-500
in Figure 12. This plot is useful to see the distribution of individual influences and the
relation with the value of each feature. Generally, positive net margins contribute to high
SHAP influence values. In the case of the PER feature, low positive values make high
AI 2025, 6, 95 15 of 27

SHAP values, but high PER values can also have either a positive or negative impact. In
the years shown, for instance, technological companies had high PER values, and, even
though they obtained high price increases, medium and low dividend yields are usually
related to positive price increases.

Figure 11. SHAP bar plot for the particular case of Apple.

Figure 12. SHAP beeswarm plot showing the distribution of the SHAP values.

We also applied a LIME technique with the LIME Tabular Explainer in Python 3.11
Figure 13 shows the result of the explanation of a particular prediction about the feature
importance of PER (Feature 0), NPM (Feature 1), dividend yield (Feature 2), ROI (Feature 3),
payout ratio (Feature 4), and operating cashflow per share (Feature 5), respectively. One can
observe that this LIME chart explains similar facts to the previously presented individual
SHAP explanation, such as the negative impact of NPM. Providing both LIME and SHAP
explanations can help users to cross-validate findings.

Figure 13. LIME explanation.

3.5. Cybersecurity and Encryption on the Recommender Website

Recommender systems based on AI in the financial domain, especially those accessible
through web platforms, require a comprehensive strategy to ensure security, user trust,
AI 2025, 6, 95 16 of 27

and proper risk management. The exposure to external threats, the sensitivity of the data,
and the system’s potential to influence economic decisions make a robust, ethical, and
traceable approach essential. This section presents cybersecurity methodology for making
this system safe.
This methodology proposes using advanced encryption mechanisms for the web
interface of the system. The cybersecurity in this kind of system is important to avoid
potential hackers trying to alter the recommendations to influence the market to their own
benefit. Firstly, common website attacks are recommended to be controlled by sanitizing
the inputs. Secondly, the website server needs to be secure, so either a commercial web
server with proper security is used or the website has common cybersecurity mechanisms.
This methodology includes a model proposal for risk, security, and trust analysis. We
propose an advanced model for risk, security, and trust management that is specifically
tailored to a system that utilizes DeepSeek for long-term investment recommendations
based on financial fundamentals, such as PER and NPM. This proposal is grounded in the
AI TRiSM framework, which offers a comprehensive methodology to address the challenges
in AI systems. AI TRiSM emphasizes transparency, accountability, fairness, reliability, and
compliance with ethical standards, making it a strong foundation for ensuring trust and
security in financial AI applications combined with XAI.
Our proposed model is designed to address vulnerabilities specific to systems like
DeepSeek by creating a robust and transparent environment. In terms of risk management,
we recommend adversarial perturbation tests to evaluate the resilience of the DeepSeek
model when data inputs are manipulated. This ensures that key financial indicators, such as
PER and NPM, remain dependable predictors, even under adversarial conditions. Similarly,
black box testing can simulate external manipulations in historical data or APIs, helping
the system detect and manage unreliable data without requiring direct access to its internal
structure. White box testing is proposed to examine and reinforce the gradients within
DeepSeek, preventing potential vulnerabilities from being exploited. We also suggest
simulating failures in historical financial data intentionally to analyze how the system
identifies and rectifies anomalies, thus ensuring high data quality standards. Furthermore,
sandbox environments are recommended to simulate real attack scenarios, such as API
disruptions, and to validate the system stability during high-stress conditions. Finally,
resilience validation procedures are critical to ensuring that the system can recover swiftly
after adversarial events, thus preserving security and performance integrity.
For security management, we propose measures to safeguard the system from unau-
thorized access by employing multi-factor authentication (MFA) and encrypting sensitive
data during storage and transmission. Adversarial training is integrated into the system
development to prepare DeepSeek for secure operation under adverse conditions. Addition-
ally, ongoing audits are recommended to uphold compliance with international regulations,
such as GDPR, ensuring the system adheres to ethical and legal standards. This can be
supported by recent studies on privacy-preserving AI, such as in distributed matrix compu-
tations [20], where minimizing information leakage is fundamental to adversarial resilience.
This methodology proposes the use of homomorphic encryption that enables computations
on encrypted data without decrypting it. In particular, this encryption is especially useful
if federated learning is used to apply the ML model, as shown in [21]. To guarantee that
the executing code is actually the one proposed by the authors, we propose to authenticate
the code through our approach based on the principles of the blockchain [22].
Trust management plays a vital role in our proposal. Interactive dashboards are
suggested to provide clear and comprehensive explanations of how DeepSeek utilizes
financial indicators in its decision-making process. This transparency fosters user trust by
making complex AI-driven recommendations understandable. Accountability is supported
AI 2025, 6, 95 17 of 27

by maintaining detailed records of all decisions made by DeepSeek, which would facilitate
audits and reinforce ethical oversight. Real-time dashboards are recommended to allow
users to visualize the impact of data on recommendations, promoting user understanding
while enhancing system transparency.
Continuous monitoring is fundamental to our proposed model. Advanced platforms
such as Datadog or Splunk are recommended to track data inputs, system metrics, and
anomalies in real time. Key Risk Indicators (KRIs) are established to automatically alert
to unusual changes in metrics, like PER, NPM, ROI, dividend yield, and operating cash-
flow, thereby ensuring the prompt identification of critical risks. Predictive analysis tools,
including algorithms like Isolation Forest, would detect anomalies and classify risks based
on severity, facilitating proactive interventions. Enhanced dashboards could visualize
historical trends, current metrics, and alert statuses, providing comprehensive oversight of
DeepSeek’s operation. Finally, regular audits are proposed to ensure compliance with qual-
ity and security standards, aligning with AI TRiSM’s principles of continuous evaluation
and improvement.
This proposed model addresses vulnerabilities specific to AI systems in finance while
integrating ethical governance to ensure responsible deployment. By leveraging the AI
TRiSM framework and incorporating XAI principles, DeepSeek would operate transpar-
ently and reliably, mitigating risks while fostering user trust and ethical compliance. These
recommendations position DeepSeek as a dependable solution for long-term investment
decision making as it has achieved the highest standards in AI-driven financial systems.
Table 1 summarizes the key elements of the proposed model, emphasizing its align-
ment with AI TRiSM principles and its application within DeepSeek and XAI frameworks
for optimal security, trust, and risk management.

Table 1. Summary of the AI TRiSM-Based risk, security, and trust management model for XAI and
DeepSeek.

Key Proposed Objectives Specific to XAI

Aspect
Mechanisms and DeepSeek
Adversarial testing, black
Ensure robust predictions
box testing, white box
from DeepSeek while
testing, simulation of data
Risk Management maintaining the reliability
failures, sandbox attack
of financial indicators
scenarios, and resilience
under XAI principles.
validation
Safeguard sensitive data
Multi-factor authentication
within DeepSeek, ensuring
(MFA), encryption,
Security Management secure and ethical
adversarial training, and
operation consistent with
GDPR-compliant audits
AI TRiSM.
Enhance explainability in
DeepSeek’s
Interactive dashboards,
recommendations by
Trust Management detailed decision logs, and
showcasing XAI-driven
real-time data visualization
transparency and fostering
user trust.
Facilitate proactive
Advanced platforms
anomaly detection and
(e.g., Datadog and Splunk),
real-time risk management,
Continuous Monitoring KRIs, predictive analysis,
ensuring DeepSeek’s
enhanced dashboards, and
reliability and alignment
regular audits
with AI TRiSM and XAI.
AI 2025, 6, 95 18 of 27

4. Evaluation Method
This work was evaluated both in a computerized way based on statistical analyses
and GAI and with a user study involving participants.
The computerized evaluation of this work involved both the portfolio prediction and
the explanations provided by GAI tools. Figure 14 outlines the processes of two distinct
evaluations: (1) the evaluation of the recommended portfolio, and (2) the evaluation of
explanations generated by DeepSeek.

Figure 14. Overview of the evaluation phase based on statistical analyses and GAI.

For the evaluation of the recommended portfolio, our model generates predictions
using the test dataset. These predictions are subsequently assessed through two methods:
statistical comparison and GAI evaluation. In the statistical comparison, the portfolio
prediction was compared to the NASDAQ-100 prediction, and a Student t-test was then
conducted to determine whether the differences between them were statistically significant.
Additionally, the predictions were evaluated using other GAI tools, including ChatGPT,
Gemini, Grok, and Qwen, in a process referred to as GAI evaluation.
For the evaluation of the explanations, which are depicted in DeepSeek explanations,
they were assessed by other GAI tools using evaluation prompts specifically designed
through prompt engineering. The evaluation score provided by each GAI tool is based
on two factors: the quality of the recommended portfolio (evaluation of recommended
portfolio), and the quality of the explanations.
In order to ensure the consistency in the evaluation results from different GAI tools,
we used the exact same prompt for all GAI tools in the evaluation.
We constructed this prompt based on the two aforementioned factors about the portfo-
lio and its explanations. In this prompt, we also explicitly included the range of the scores
so that the evaluations were unified in scale across the different GAI tools. The constructed
prompt for evaluation was literally the following one:
“Please, evaluate from 0 to 100 each of the following aspects of the following recommended
portfolio: (1) Quality of the portfolio, and (2) Explanation of the portfolio”
Furthermore, we also performed similar comparisons with the S&P 500 in a completely
different scenario (7 April 2025), in which there was an international world-wide crisis due
to the custom duties for regulating foreign trade.
For the user study, we presented our created system to volunteering participants,
and we then surveyed them about the most relevant aspects of our presented system.
AI 2025, 6, 95 19 of 27

Section 5.3 describes the sample of participants, introduces the research questions of the
survey, and discusses the results.

5. Results
In this section, we explain the main results of applying the proposed methodology
and the evaluation method when considering the NASDAQ-100 and S&P 500 indexes. In
the first step, we obtained the generated dataset, which is composed of eight attributes
(PER, NPM, dividend yield, ROI, operating cashflow per share, price 4 years ago, current
price, and price increase) and 600 instances.
In Step 2, we preprocessed the dataset, dividing it into X and Y. Then, we preprocessed
X, obtaining the following descriptive statistics for the PER (mean = 67.61, std = 189.50,
max = 1236.83, and min = −157.59) and NPM (mean = 10.985, std = 40.947, max = 63.717,
and min = −313.50). These results indicate that the scales of PER and NPM were not similar;
as such, it was necessary to normalize the dataset. Then, we split the dataset into training
(50%) and test (50%) datasets. As a result, the size was 300 and 300 for the training dataset
and test dataset, respectively. Then, the MLP hyperparameters (the number of neurons of
the first and second layers) were tested using a k-fold cross-validation strategy.
The experiments used k = 4. As a result, the distribution of 150 neurons for the first
layer and 150 for the second layer seemed to be an optimal neural network configuration.
We tested the optimizers limited-memory Broyden–Fletcher–Goldfarb–Shanno (L-BFGS),
adaptive moment estimation (ADAM), and stochastic gradient descent (SGD) with a
learning rate of 0.0001 and a ReLU activation function. Our results indicated that the
L-BFGS optimizer obtained better results than ADAM and SGD. Next, we generated an
MLP model using an L-BFGS optimizer with a learning rate of 0.0001 and a ReLu activation
function, and we then trained this model with the training dataset.
We applied a cross-validation of four random balanced splits following these criteria.
Since the input data size was small, MLP was an appropriate choice for this kind of data, as
recommended in the literature [23].
In Step 3, our system provided explanations of the obtained investment portfolio when
using the DeepSeek tool and the following prompt.
Provide brief descriptions of about 100 words for each of the companies with the following
tickers: CHTR, PCAR, VRSK, MSFT, MNST, ADBE, SWKS, AVGO, LRCX, QCOM.
For example, Figure 8 shows the explanations offered by the DeepSeek tool for the
first five companies in the recommended portfolio. For instance, DeepSeek informed that
“Charter Communications is a leading broadband connectivity and cable operator in the
U.S.” This information enhances the feasibility of the investment portfolio recommendation
provided by our system.
Following our evaluation method, we compared the effectiveness of the recommen-
dations of our approach and the DeepSeek baseline. For example, Figure 15 shows the
profitability of the proposed system for the recommendations of a 10-stock portfolio. In
this particular case, one can observe the profitability of the proposed system of 110.02%
over four years compared to the NASDAQ profitability of 50.71% in the same period.
AI 2025, 6, 95 20 of 27

Figure 15. Selected stocks based on the proposed system after filtering positive PER and NPM using
the NASDAQ-100.

It is worth mentioning that the NASDAQ-100 mainly includes technological and

growth companies, but it also utilizes different subfields, and each subfield may have
been identified by the PER and NPM. This might be the reason why most companies
in the proposed portfolio might be from the technological, semiconductors, software, or
telecommunication fields. Additionally, the proposed portfolio includes sectors beyond
technology, such as trucks and beverages. This inclusion provides a degree of diversifica-
tion, particularly given the NASDAQ-100’s primary focus on technology stocks.
Moreover, Figure 16 presents an example of portfolio recommendation with the S&P-
500. In this particular portfolio, the profitability was 45.32% compared to the 40.96% of the
reference index.

Figure 16. Selected stocks based on the proposed system after filtering the positive PER and NPM
using the S&P-500.
AI 2025, 6, 95 21 of 27

The next subsections provide the statistical analysis of the profitability of the proposed
investment portfolio recommendations and evaluations of our system explanations, which
were partially generated by DeepSeek.

5.1. Statistical Analysis About the Profitability of the Proposed Stock Investment Recommendations
over the Passive Index Investment
When investing, it is important to conduct a statistical analysis on the results. This
ensures that any improvements are not due to chance and that the proposed portfolio
recommendations offer a statistically significant improvement over the index.
We generated portfolios by splitting them into training and test datasets in four differ-
ent random divisions. For each possible split, we generated 20 different sizes (comprising
portfolios from 8 to 27 companies). In this way, we had 120 different proposed portfolios.
We compared the profitability of these portfolios with the profitabilities of the reference
indexes NASDAQ-100 and S&P-500. We also compared the results using t-Student for the
paired groups, as these portfolios were in the same four-year period. Both groups were in
the same period of years, so the macroeconomic aspects were the same in both the control
group (reference index) and the experimental group (proposed portfolio).
Figure 17 shows the results of the t-Student test. One can observe that the average
profitability was greater than the reference index within four years, achieving 56.62% of
profits on average rather than 47.46%, which was the average of the companies in the
reference indexes. The results were statistically significant, with a significance of 0.000
in total and a two-tail significance of also 0.000. Therefore, we concluded that there are
statistically significant differences between the recommendation portfolio profit and the
reference–index profit.

Figure 17. The t-Student results when comparing the profits of the 120 portfolios with the NASDAQ-
100 and S&P-500 indexes.

5.2. Evaluation of the Explanations of Proposed Portfolios

To assess the usefulness of the portfolio proposed in Section 5, we asked the following
prompt for GAI tools, which are different from the one used for generating company
descriptions:
Please evaluate from 0 to 100 each of the following aspects of the following recommended
portfolio: (1) Quality of the portfolio, and (2) Explanation of the portfolio
The ChatGPT (model used: ChatGPT-40) evaluation showed the strengths and weak-
nesses of the proposed portfolio, providing a high score in both dimensions: 85 and 90,
respectively. The Grok (model used: Grok3) evaluation also showed the strengths and
weaknesses, but it provided scores below the ChatGPT scores: 75 and 65. In the same vein
was the evaluation provided by Qwen (model used: Qwen2.5-Max). It showed the strengths
AI 2025, 6, 95 22 of 27

and weaknesses of the proposed portfolio but it provided scores below ChatGPT: 60 and 65.
Nevertheless, the Gemini (model used: Gemini 2.0 Flash Thinking Experimental) evaluation
was different than the previous ones. Its evaluation is based on different aspects, such as
diversification, risk, and target investors, for the quality dimension. Gemini evaluated the
explanation dimension in the following aspects: clarity and simplicity, logical rationale,
transparency, and context. The Gemini scores were 58 and 45. Taking into account the four
evaluations provided by the GAI tools, the average score was 69.5 and 66.25, respectively.
As an example, the ChatGPT evaluation is shown below.

5.2.1. Quality of the Portfolio (Score: 85/100)

• Strengths: The portfolio consists of well-established companies across various in-
dustries, including technology (MSFT, ADBE, AVGO, QCOM, SWKS, and LRCX),
telecommunications (CHTR), automotive (PCAR), analytics (VRSK), and consumer
goods (MNST). Many of these companies have strong financials, high profit margins,
and solid market positions.
• Diversification: While the portfolio was diversified across different sectors, it leaned
heavily toward tech and semiconductors, which could expose it to sector-specific risks.
• Valuation and Growth Potential: The predicted price increases seemed ambitious, es-
pecially for CHTR (+360%), but the presence of profitable and innovative companies
added credibility to the selections.

5.2.2. Explanation of the Portfolio (Score: 90/100)

• Company Descriptions: The explanations for each company were clear, concise, and
informative, giving a strong overview of their business models and market relevance.
• Financial Data: The inclusion of P/E ratios, net margins, and predicted price increases
added a quantitative dimension, although more context on how these metrics were
derived would be helpful.
• Suggested Improvement: It would be beneficial to provide the reasoning behind the
selections, such as the industry trends, competitive advantages, or specific catalysts
driving the predicted price increases.
As the last step, we evaluated a human expert portfolio recommendation with an
explanation as the baseline. On 17 April 2025, we searched on Google with the keywords
“portfolio recommendation with explanation”. We explored the results of Google following
the ranking order of the first web page, and the first entry that actually contained investment
portfolio examples with explanations was the White Coat Investor website ([Link]
[Link]/150-portfolios-better-than-yours/ (last accessed on 17 April 2025)).
From the given recommendation examples of this website, we picked a diversified portfolio
with a reasonable explanation (this recommended portfolio was referred to as Portfolio
26 on this website). We evaluated this portfolio with the exact same mechanism using
ChatGPT, and the result was a score of 85/100 in terms of quality of the portfolio and a
score of 75/100 for the explanation of the portfolio. As one can observe, our approach
obtained a better result in the explanation (90 instead of 75, i.e., out of 100) according to
this GAI evaluation. Regarding the quality of the portfolios, both of them obtained the
same score. Thus, our approach improved the explainability when compared to a common
human expert recommendation that was quickly retrieved from the Internet, according to
ChatGPT evaluation. In the comparison with the baseline, we only considered this tool
since we believed that the other GAI tools were not properly considering SHAP outputs.
Consequently, there were no significant improvements in the scores over human expert
recommendations. This comparison with the baseline was limited as the human expert
review was obtained freely online with a simple search, and only a GAI tool was used,
AI 2025, 6, 95 23 of 27

although it was probably the most mature. Our approach may be useful compared with
this kind of human recommendation, but a further future comparison will be necessary to
reliably corroborate such statement.

5.3. User Study About Portfolio Recommendations and Their Explanations

For this study, we asked for volunteers who participated in the evaluation of a tool
for recommending stock portfolios based on XAI with explanations. In this study, seven
people participated. This study included seven adult participants, with a balanced gender
distribution (four women and three men). Six participants were from Spain, and one
was from the United States. Three participants had a background in financial and stock
investment, while four did not.
For evaluating our XAI-based portfolio recommender, we designed a questionnaire
based on the key factors that we were interested in evaluating. More concretely, we asked
participants to evaluate the following sentences in a 7-point Likert scale: (Q1) “I understand
the portfolio recommendation”, (Q2) “I understand the descriptions of the companies in
the recommendation”, (Q3) “I think this application is useful”, and (Q4) “I think having
these explanations are useful for trusting portfolio recommendations”.
Figure 18 presents the results of the survey about our XAI-based portfolio recom-
mender. It is worth mentioning that the highest value was obtained in Q2 about the
understandability of company descriptions (ranked as 6.57 out of 7), which were generated
through GAI by means of DeepSeek, so the result of incorporating this technology was
really appreciated by the participants. Q3 also obtained a high value (6.14 out of 7), showing
that the users found our recommender application useful.

Figure 18. Survey results about our XAI-based portfolio recommender.

The most important finding was that the novelty of the explanations generated through
the XAI-based techniques was considered useful when it came to trusting portfolio recom-
mendations (Question Q4), and it also had a high value (6.14 out of 7). This corroborates
the importance and utility of providing explanations in such systems. Figure 19 shows the
distribution of replies to Q4, showing that most participants (57.1% of the sample) agreed
(6 out 7) with the usefulness of these XAI-based explanations for gaining trust, and another
AI 2025, 6, 95 24 of 27

representative group of participants (28.6% of the sample) strongly agreed (7 out of 7) with
this usefulness.

Figure 19. Distribution of replies to Q4 about the usefulness of explanations for trusting portfolio
recommendations.

6. Discussion
The presented 4-year profitability percentages were considered in two 4-year periods
(from February 2021 to February 2025, and from April 2021 to April 2025). Consequently,
these were affected by the macroeconomic aspects of these two periods, and they were
subsequently not comparable to either of the other past or future periods. However, what
is relevant was that, in the same periods, our proposed approach outperformed the passive
indexes of the reference in a statistically significant way. It is worth mentioning that the
two periods used included different scenarios since the ending dates were before and after
the USA custom duties for international trade. Thus, the first period was a bull market
with index prices near maximums, while the latter period included market crashes with
price decreases of up to 20% in several days with very high volatility.
Financial indicator ratios depend on the field. Different fields are usually associated
with different ranges of PER, NPM, ROI, dividend yields, cashflow, and payout ratios. This
approach has been successfully when applied to a reference index that overweights certain
sectors and companies with certain features; more specifically, we used NASDAQ-100
companies for this purpose. Thus, our approach might also be useful in this kind of index
when it is focused on certain sectors or companies with shared features. The outcomes may
not be extendable when using more diversified indexes, either in fields or geographically.
Even though we also analyzed the S&P 500 for our study with diversified sectors in the
USA, all of the results together of the NASDAQ-100 and S&P 500 still saw statistically
significant improvements. Thus, our approach might be on the right path toward a more
generic approach for recommending stock investments in different sectors and regions
with the proper training.
Every trimester, each company provides the financial fundamentals that provide
insight information that reveals new information about the value of the company to the
public. However, almost immediately, the price of the company changes with this new
information. Normally, the reports are released out of market hours. The price changes
in the pre-market without the possibility of either buying or selling. Thus, the opening
stock prices now incorporate the new information. Thus, this approach is competing with
everyone else, who also have exactly the same information, and it uses the fundamental
value based on the financials for long-term investments, like in the long-term investments
of the well-known investor Warren Buffett [24].
It is worth mentioning that this work provided ML attribution-based explanations,
which focus on ranking the importance value of input features. However, as mentioned
AI 2025, 6, 95 25 of 27

in [25], there are two other types of ML explanations. Model-based explanations focus
on describing the learned ML, like the decision trees’ depth and the number of non-zero
weights in linear models. In our short-term future work, these could help in understanding
the complexity of ML models necessary to outperform market profitability and whether
there is any relation between complexity/simplicity and profitability. Another kind of
ML explanation are example-based explanations. This can be useful for understanding
the recommendations based on similarity or their difference (i.e., counterfactual exam-
ples) with some other stock values. All these improvements could enrich the portfolio
recommendations of our system and their interpretation.
The SHAP assumes feature independence and model additivity, but these are not
met in neural networks with non-linear dependencies. SHAP charts are only illustrative
explanations, and users need to be aware of their potential limitations on non-linear models,
such as the ones described in [26]. As a possible mitigation strategy, the SHAP outcomes
can be enhanced using the SHAP proxy proposed by [19]. This XAI proxy approach has
been successfully applied in the context of image processing in cardiac magnetic resonance
images and brain images (specifically MRI) in the health domain. However, the application
of this XAI proxy is not straightforward, as the underlying software is not public and its
usefulness has only been proven in image processing, which is not exactly our context. It is
worth mentioning that LIME provides deeper explanation insights without this limitation.
However, LIME explanations are performed on each individual prediction [27], so it has
the limitation that users cannot analyze the complete learned model as a whole.
The results of the user study are promising, although it has some limitations. First, the
small sample size of seven participants limits the generalizability of the findings. Moreover,
this study lacked domain knowledge control and did not require any specific economic or
XAI knowledge. Most participants lacked experience in stock investment or XAI, which
may have constrained their ability to provide nuanced feedback. If the users had experience
in financial analysis, they may have had either corroborated or refuted recommendations
with more insight into financial knowledge. Consequently, this study would have been
more meaningful for users with such backgrounds. We plan to address these limitations
in future iterations of the user study by expanding the sample size and incorporating
domain-specific expertise.

7. Conclusions and Future Work

This work has presented a methodology for recommending stock investment portfolios
in a given reference index based on the fundamentals of companies that are reflected in
the financial information reported on their trimestral reports. The proposed methodology
has been able to provide 120 different portfolios based on the NASDAQ-100 and S&P-
500 lists of companies, with different portfolio sizes between 8 and 28 companies based
on the financial information obtained from the four years before the experiment. The
profitability, on average, of these recommended portfolios increased compared to the
average profitability of the lists of reference indexes in the same period. The t-Student
analysis of this comparison demonstrated a statistically significant improvement (p < 0.001).
The portfolios were explained with XAI and GAI tools using SHAP, LIME, and DeepSeek as
the underlying technologies. The recommended portfolios and descriptions were positively
assessed according to our GAI-based evaluation and our user study.
Our work can assist novice investors who struggle to understand portfolio recommen-
dations. The proposed system not only provides a portfolio recommendation, but it also
offers clear and useful explanations for each stock in the portfolio. Additionally, the pro-
posed system can benefit expert investors by suggesting a potential portfolio. Experts can
analyze the fundamentals, evaluate the provided explanations about these fundamentals,
AI 2025, 6, 95 26 of 27

and integrate them with external factors, such as current macroeconomic aspects. Further-
more, they can gain insights into the sector-specific implications tied to the fundamentals.
This work was planned to be extended by considering more fundamental values, i.e.,
in terms of more fundamental financial ratios and the sequence of the evolution of these
financial ratios before the time the predictions would be made. In this line, our work will
consider recurrent neural networks, like long-short term memory (LSTM) networks or deep
learning approaches (such as convolutional neural networks (CNN)). In an alternative line
of work, we will also consider the possibility of providing diversified portfolios departing
from several sectoral indexes based on a separate training for each of them.

Author Contributions: Conceptualization, I.G.-M.; methodology, I.G.-M., J.B.-A. and R.L.; software,
I.G.-M.; validation, J.B.-A. and I.G.-M.; formal analysis, I.G.-M. and J.B.-A.; investigation, I.G.-M., J.B.-
A. and R.L.; resources, J.B.-A.; data curation, I.G.-M. and J.B.-A.; writing—original draft preparation,
I.G.-M. and J.B.-A.; writing—review and editing, J.B.-A., I.G.-M. and R.L.; visualization, I.G.-M.;
supervision, I.G.-M. and R.L.; project administration, I.G.-M. and R.L.; funding acquisition, R.L. All
authors have read and agreed to the published version of the manuscript.

Funding: This research was funded by the project “Proyecto Estratégico de Ciberseguridad Desarrol-
lado en Zonas Escasamente Pobladas”, which was funded by the Instituto Nacional de Ciberseguridad.
This work was developed in the context of “Plan de Recuperación, Transformación y Resiliencia”,
which is funded by the European Union (Next Generation). This work was also partially supported
by the Spanish Ministry of Science [grant number: PID2023-149341OB-I00]. The APC was funded by
the first project.

Institutional Review Board Statement: Not applicable.

Informed Consent Statement: Not applicable.

Data Availability Statement: The dataset of this research is available at [Link]

datasets/5jk4bm7x5v/1 (accessed on 1 May 2025).

Conflicts of Interest: The authors declare no conflicts of interest.

References
1. Sun, P.W.; Liao, W.J.; Lin, W. Analyzing the nature of fund selection measures: Stock picking or trading skill? Q. Rev. Econ. Financ.
2024, 97, 101883. [CrossRef]
2. Joshi, G.; Dash, R.K. Exchange-traded funds and the future of passive investments: A bibliometric review and future research
agenda. Future Bus. J. 2024, 10, 17. [CrossRef]
3. Pelster, M.; Val, J. Can ChatGPT assist in picking stocks? Financ. Res. Lett. 2024, 59, 104786. [CrossRef]
4. Khuwaja, P.; Khowaja, S.A.; Khoso, I.; Lashari, I.A. Prediction of stock movement using phase space reconstruction and extreme
learning machines. J. Exp. Theor. Artif. Intell. 2020, 32, 59–79. [CrossRef]
5. Kang, M.; Templeton, G.F.; Kwak, D.H.; Um, S. Development of an AI Framework Using Neural Process Continuous Reinforce-
ment Learning to Optimize Highly Volatile Financial Portfolios. Knowl.-Based Syst. 2024, 300, 112017. [CrossRef]
6. Gandhmal, D.P.; Kumar, K. Systematic analysis and review of stock market prediction techniques. Comput. Sci. Rev. 2019,
34, 100190. [CrossRef]
7. Habbal, A.; Ali, M.K.; Abuzaraida, M.A. Artificial Intelligence Trust, risk and security management (AI TRiSM): Frameworks,
applications, challenges and future research directions. Expert Syst. Appl. 2024, 240, 122442. [CrossRef]
8. Aiche, A.; Winer, Z.; Cohen, G. Constructing Cybersecurity Stocks Portfolio Using AI. Forecasting 2024, 6, 1065–1077. [CrossRef]
9. Kim, J.H. What if ChatGPT were a quant asset manager. Financ. Res. Lett. 2023, 58, 104580. [CrossRef]
10. Huang, Z.; Zhang, Z.; Hua, C.; Liao, B.; Li, S. Leveraging enhanced egret swarm optimization algorithm and artificial intelligence-
driven prompt strategies for portfolio selection. Sci. Rep. 2024, 14, 26681. [CrossRef]
11. Takahashi, S.; Chen, Y.; Tanaka-Ishii, K. Modeling financial time-series with generative adversarial networks. Phys. A Stat. Mech.
Its Appl. 2019, 527, 121261. [CrossRef]
12. Huang, Z.; Che, C.; Zheng, H.; Li, C. Research on Generative Artificial Intelligence for Virtual Financial Robo-Advisor. Acad. J.
Sci. Technol. 2024, 10, 74–80. [CrossRef]
AI 2025, 6, 95 27 of 27

13. Muhammad, I.; Rospocher, M. On Assessing the Performance of LLMs for Target-Level Sentiment Analysis in Financial News
Headlines. Algorithms 2025, 18, 46. [CrossRef]
14. Jackson, I.; Ivanov, D.; Dolgui, A.; Namdar, J. Generative artificial intelligence in supply chain and operations management:
A capability-based framework for analysis and implementation. Int. J. Prod. Res. 2024, 62, 6120–6145. [CrossRef]
15. Ante, L.; Saggu, A. Quantifying a firm’s AI engagement: Constructing objective, data-driven, AI stock indices using 10-K filings.
Technol. Forecast. Soc. Change 2025, 212, 123965. [CrossRef]
16. Ferrara, M.; Ciano, T. Improving the Interpretability of Asset Pricing Models by Explainable AI: A Machine Learning-based
Approach. Econ. Comput. Econ. Cybern. Stud. Res. 2024, 58, 5–19. [CrossRef]
17. Wang, J.; Chen, Z. Factor-GAN: Enhancing stock price prediction and factor investment with Generative Adversarial Networks.
PLoS ONE 2024, 19, e0306094. [CrossRef]
18. Yang, C.; Guan, X.; Xu, Q.; Xing, W.; Chen, X.; Chen, J.; Jia, P. How can SHAP (SHapley Additive exPlanations) interpretations
improve deep learning based urban cellular automata model? Comput. Environ. Urban Syst. 2024, 111, 102133. [CrossRef]
19. Salih, A.M.; Galazzo, I.B.; Raisi-Estabragh, Z.; Petersen, S.E.; Menegaz, G.; Radeva, P. Characterizing the contribution of
dependent features in XAI methods. IEEE J. Biomed. Health Inform. 2024, 28, 6466–6473. [CrossRef]
20. Morteza, A.; Chou, R.A. Distributed Matrix Multiplication: Download Rate, Randomness and Privacy Trade-Offs. In Proceedings
of the 2024 60th Annual Allerton Conference on Communication, Control, and Computing, Urbana, IL, USA, 24–27 September
2024; pp. 1–7.
21. Kumbhar, H.R.; Rao, S.S. Federated learning enabled multi-key homomorphic encryption. Expert Syst. Appl. 2025, 268, 126197.
[CrossRef]
22. García-Magariño, I.; Lacuesta, R.; Rajarajan, M.; Lloret, J. Security in networks of unmanned aerial vehicles for surveillance with
an agent-based approach inspired by the principles of blockchain. Ad Hoc Netw. 2019, 86, 72–82. [CrossRef]
23. Rana, A.; Rawat, A.S.; Bijalwan, A.; Bahuguna, H. Application of multi layer (perceptron) artificial neural network in the
diagnosis system: A systematic review. In Proceedings of the 2018 International Conference on Research in Intelligent and
Computing in Engineering (RICE), San Salvador, El Salvador, 22–24 August 2018; pp. 1–6.
24. Henriques-de Brito, M. Warren Buffett’s Bet and Berkshire Hathaway’s Investment Skill. Procedia Comput. Sci. 2015, 55, 1193–1202.
[CrossRef]
25. Lopes, P.; Silva, E.; Braga, C.; Oliveira, T.; Rosado, L. XAI systems evaluation: A review of human and computer-centred methods.
Appl. Sci. 2022, 12, 9423. [CrossRef]
26. Sabbatini, G.; Manganaro, L. On potential limitations of differential expression analysis with non-linear machine learning models.
EMBnet J. 2022, 28, 1035. [CrossRef]
27. Salih, A.M.; Raisi-Estabragh, Z.; Galazzo, I.B.; Radeva, P.; Petersen, S.E.; Lekadir, K.; Menegaz, G. A perspective on explainable
artificial intelligence methods: SHAP and LIME. Adv. Intell. Syst. 2025, 7, 2400304. [CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.