Deep learning for enhanced risk management: a novel approach to analyzing financial reports

Motivations

The key motivations of the proposed work are:

Complexity and volume of financial data: Huge volumes of data are presented precisely in MS Financial Statements of the contemporary period which are voluminous, extensive, detailed and compounded with unorganized information that are not easy to process and analyze by conventional techniques of analysis.

Limitations of traditional methods: Traditional approaches to managing risk involve the use of analysis and models which work on heuristics and take time in their processing. Neither of these methods is adequate for handling the massive amount of data that is being produced today and for the higher level of sophistication that is found in today’s financial data.

Advancements in deep learning: Nevertheless, given that deep learning models have the potential to process and analyze big data with good accuracy then the prediction of financial risks can be enhanced. Using a combination of quantitative and qualitative data extracted from financial texts, deep learning models’ risk assessment is more accurate.

Improved risk mitigation: Although the new approach, the Hybrid Financial Risk Predictor (HFRP) model based on CNN and LSTM networks has provided a better solution for accurate risk prediction. With the help of this model, all kinds of financial risks including credit risk, liquidity risk, market risk, and operational risk can be minimized to a great extent.

Efficiency and accuracy: The objective of the research is to improve the effectiveness of the risk management process by simplifying and automating it, as well as to minimize the time needed to analyze the companies’ financial reports to obtain reliable results about the existence and characteristics of the potential risks. This would create a knowledge that enables the financial institutions to make right decisions and also mitigate risks effectively.

The rest of the article is organized as follows to cover all the aspects of deep learning in improving risk management with the help of financial report analysis. The Introduction draws the background that raises the extent of utilization of effective risk management mechanisms in financial markets and problems associated with the use of traditional approaches, the statement of the problem, problem formulation, and research questions and objectives. The Literature Review of this study plots the current literature on risk management, deep learning, and natural language processing to uncover research voids that this work seeks to fill. Our research’s Methodology presents an overview of the steps taken to build, train, and incorporate the deep learning framework in use as well as the integration of NLP methods. In the Results and Discussions section, we report and compare the results obtained in the experiments with the traditional methods and analyze the significance of the results. Last, the Conclusions contain an overview of the main findings of the research, its implications, and possible future studies in the sphere of financial risk management.

Knowledge gap and contribution

Existing financial risk models predominantly employ statistical methods or single-modality neural networks, such as using only CNN for text analysis or LSTM for numerical data. These methods have inherent limitations:

• 1)

  Inadequate text analysis: Traditional models often fail to extract semantic information from complex, unstructured text data present in financial disclosures.

• 2)

  Lack of temporal awareness: Models using only numerical data typically miss out on the contextual nuances captured in textual information, leading to suboptimal risk assessments.

• 3)

  Scalability issues: Many models are designed for specific financial contexts and may not generalize well across different industries or economic conditions.

The HFRP model addresses these issues by:

• Combining the strengths of CNN and LSTM to process both textual and numerical data simultaneously.

• Utilizing advanced NLP techniques for feature extraction from financial disclosures, enhancing the model’s ability to interpret qualitative risk factors.

• Implementing a scalable framework that adapts to diverse financial datasets, improving the generalizability of the predictions.

This integrated approach fills a critical gap in the current state of financial risk analysis by providing a unified model capable of capturing intricate relationships within multimodal data. It aims to set a new standard for predictive modeling in financial risk management, with broader implications for decision-making processes in financial institutions.

Computing infrastructure

The system configuration includes Ubuntu 20.04 LTS as the operating system, providing a stable and efficient platform for various tasks. The hardware is equipped with an Intel Core i7-9700K CPU running at 3.60 GHz, complemented by 32 GB of RAM, which ensures smooth multitasking and high performance. For graphical processing, it utilizes an NVIDIA GeForce RTX 2070. On the software side, the setup includes Python 3.8 for programming, Jupyter Notebook for interactive computing, and Anaconda Distribution for managing packages and environments.

Model description

The proposed model is based CNN and LSTM networks.

Convolutional neural networks (CNN): CNN are primarily used for feature extraction from text data. They are effective in identifying and learning patterns within the data through convolutional layers. The CNN part of the model consists of multiple convolutional layers followed by pooling layers, which help in reducing the dimensionality of the data while preserving important features.

Long short-term memory (LSTM) networks: LSTM are utilized for capturing temporal dependencies in the data. They are particularly useful for sequential data and can effectively handle long-term dependencies. The LSTM network includes multiple LSTM cells, each capable of maintaining information over long periods, which is essential for understanding the context in financial texts.

Integration of CNN and LSTM: The integration of CNN and LSTM networks leverages the strengths of both models. CNN are highly effective for feature extraction from textual data, while LSTM excel at capturing temporal dependencies. The combined model, referred to as HFRP, is specifically designed to process both quantitative and qualitative aspects of financial data, providing a more comprehensive risk assessment.

Evaluation: The performance of the model is evaluated using standard metrics such as accuracy, precision, recall, and F1-score to ensure its reliability. Additionally, the model is validated on a separate dataset to fine-tune parameters and prevent overfitting.

Figure 1 shows the workflow begins with two distinct input streams: textual data and numerical data. The textual data undergoes preprocessing, which includes tokenization, stop word removal, stemming, and lemmatization to prepare it for analysis. This preprocessed text is then vectorized and passed through CNN layers to extract meaningful semantic features. Simultaneously, the numerical data, typically structured financial metrics, is fed into LSTM networks, which capture temporal dependencies and trends inherent in the data. The features extracted from the CNN and LSTM layers are then combined into a unified feature representation, enabling the model to leverage insights from both textual and numerical data. This joint representation is processed through a series of dense layers for further refinement and decision-making. To enhance model robustness and mitigate data scarcity, GANs are incorporated, generating synthetic data to improve training quality. Additionally, RL is applied to adapt the model dynamically based on evolving market conditions, ensuring continuous improvement in predictions. Finally, the refined features are used to generate risk metrics, culminating in accurate and comprehensive financial risk predictions. This integrated approach allows for a holistic analysis of financial risks, leveraging the strengths of multiple deep learning techniques.

Comprehensive risk prediction in financial reports using hybrid deep learning models

Introduce a deep learning system that integrates attributes of CNN and LSTM to dental with textual and numerical data from financial statements. The objective is to formally define a model that is capable of identifying financial risks and how different variables may depend on one another within the data. The proposed Hybrid Financial Risk Predictor (HFRP) model integrates CNN and LSTM networks to process both textual and numerical data from financial statements. The input data representation involves two components: $X$, which is a matrix capturing numerical features of size $n\times m$, and $T$, a tensor of dimensions $n\times p\times q$ for the textual data extracted from financial reports. The CNN is utilized for feature extraction from the textual data. It applies convolution operations using filter weights $(W\mathrm{c}\mathrm{o}\mathrm{n}\mathrm{v})$ and biases $(b\mathrm{c}\mathrm{o}\mathrm{n}\mathrm{v})$, followed by an activation function ( $\sigma$), resulting in feature maps ( $F_i$ CNN). This process captures essential patterns and semantic information from the text data, which are critical for identifying potential risk indicators.

1. Input data representation:

(1) $$\mathbf{X}\in {\mathbb{R}}^{n\times m},\mathbf{T}\in {\mathbb{R}}^{n\times p\times q}$$

2. Feature extraction:

(2) $${\mathbf{F}}_i^{\mathrm{C}\mathrm{N}\mathrm{N}}=\sigma \left({\mathbf{W}}_{\mathrm{c}\mathrm{o}\mathrm{n}\mathrm{v}}\ast {\mathbf{T}}_i+{\mathbf{b}}_{\mathrm{c}\mathrm{o}\mathrm{n}\mathrm{v}}\right)$$

(3) $${\mathbf{h}}_t=\varphi \left({\mathbf{W}}_h\cdot {\mathbf{x}}_t+{\mathbf{U}}_h\cdot {\mathbf{h}}_{t-1}+{\mathbf{b}}_h\right)$$

3. Combined feature representation:

(4) $${\mathbf{F}}_i=\left[{\mathbf{F}}_i^{\mathrm{C}\mathrm{N}\mathrm{N}},{\mathbf{h}}_T\right]$$

4. Objective function (Eq. (5)): This represents the objective function used in the HFRP model. Its components ensure that the model balances minimizing prediction error while avoiding overfitting and improving model sparsity:

(5) $$\underset{\theta }{\mathrm{m}\mathrm{i}\mathrm{n}}\mathcal{L}={\displaystyle \frac{1}{n}\sum _{i=1}^n\left[\alpha {\left(y_i-{\hat{y}}_i\right)}^2+\beta \left(\sum _{j=1}^k{\theta }_j^2\right)+\gamma \sum _{j=1}^m\left|{\theta }_j\right|\right]}$$

5. Prediction function:

(6) $${\hat{y}}_i=f\left({\mathbf{F}}_i\right)={\mathbf{W}}_{\mathrm{o}\mathrm{u}\mathrm{t}}\cdot {\mathbf{F}}_i+{\mathbf{b}}_{\mathrm{o}\mathrm{u}\mathrm{t}}$$

This formulation jointly uses CNN for text and LSTM for numerical data thus formulating a joint feature representation. The objective function also consists of mean squared error, L2 regularization, and L1 regularization to increase the optimization’s complexity.

Dynamic financial risk forecasting using reinforcement learning and deep learning integration

Propose an advanced risk prediction model that may combine deep reinforcement learning (DRL) with traditional deep learning to improve the forecast with new data. The model will adjust techniques of risk management about a dynamic market environment.

1. State representation:

(7) $$s_t=\left[{\mathbf{X}}_t,{\mathbf{T}}_t\right]$$

2. Action space:

(8) $$a_t\in \left\{\mathrm{b}\mathrm{u}\mathrm{y},\mathrm{s}\mathrm{e}\mathrm{l}\mathrm{l},\mathrm{h}\mathrm{o}\mathrm{l}\mathrm{d}\right\}$$

3. Reward function:

(9) $$r_t={\omega }_1\cdot \mathrm{A}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y}\left({\hat{y}}_t,y_t\right)+{\omega }_2\cdot \mathrm{R}\mathrm{e}\mathrm{t}\mathrm{u}\mathrm{r}\mathrm{n}\left(\mathrm{P}\mathrm{o}\mathrm{r}\mathrm{t}\mathrm{f}\mathrm{o}\mathrm{l}\mathrm{i}{\mathrm{o}}_t\right)$$

4. Policy network:

(10) $$\pi \left(a_t|s_t;\theta \right)=\mathrm{s}\mathrm{o}\mathrm{f}\mathrm{t}\mathrm{m}\mathrm{a}\mathrm{x}\left({\mathbf{W}}_{\pi }\cdot s_t+{\mathbf{b}}_{\pi }\right)$$

5. Value function:

(11) $$V\left(s_t;\varphi \right)={\mathbf{W}}_V\cdot s_t+{\mathbf{b}}_V$$

6. Objective function:

(12) $$\underset{\theta }{\mathrm{m}\mathrm{i}\mathrm{n}}{\mathcal{L}}_{\pi }=-\mathbb{E}\left[\mathrm{l}\mathrm{o}\mathrm{g}\pi \left(a_t|s_t;\theta \right)\cdot \left(r_t+\gamma V\left(s_{t+1}\right)-V\left(s_t\right)\right)\right]$$

(13) $$\underset{\varphi }{\mathrm{m}\mathrm{i}\mathrm{n}}{\mathcal{L}}_V={\displaystyle \frac{1}{2}\mathbb{E}\left[{\left(r_t+\gamma V\left(s_{t+1};\varphi \right)-V\left(s_t;\varphi \right)\right)}^2\right]}$$

This problem formulation is a coupling of DRL with deep learning that makes it a dynamic model that can learn from new financial data a current feed. That is, the policy and value networks are trained to acquire the best risk management solutions at a level of accuracy desired while being a good investment.

For numerical data, the LSTM network captures temporal dependencies and trends using its hidden states. The hidden state at time t, denoted as $h_t$, is computed using the input data ( $x_t$), previous hidden state $\left(\mathit{h}\mathit{t}-1\right)$, and respective weight matrices ( $W_h$ and $U_h$). This mechanism allows the model to retain relevant information over long sequences, making it effective for analyzing time series data like financial metrics. The extracted features from both CNN and LSTM are then combined into a unified feature vector ( $F\mathrm{\_}i$), representing a comprehensive joint feature set that integrates insights from both text and numerical data.

The objective function for the model is designed to minimize prediction error while also incorporating regularization terms to prevent overfitting. The loss function includes three components: the mean squared error between the actual $\left(yi\right)$ and predicted values $\left(y^i\right)$, an L2 regularization term ( $\beta {\sum }_{\theta }^{j2}j$) to penalize large weight values, and an L1 regularization term (γ∑θj) to encourage sparsity in the model parameters. Finally, the prediction function applies a linear transformation using output weights (W_out) and biases (bout), generating the final risk prediction ( $y^i$) based on the combined feature representation.

Multi-modal financial risk analysis using GANs

Propose multi-modal risk assessment financial risk GAN for generating synthetic data for increasing the accuracy of the forecasts. The model is expected to solve problems of data deficit and at the same time improve the efficiency of risk modelling.

1. Generator network:

(14) $$\mathbf{G}\left(z;{\theta }_G\right)=\sigma \left({\mathbf{W}}_G\cdot z+{\mathbf{b}}_G\right)$$

2. Discriminator network:

(15) $$D\left(\mathbf{x};{\theta }_D\right)=\sigma \left({\mathbf{W}}_D\cdot \mathbf{x}+{\mathbf{b}}_D\right)$$

3. Adversarial loss:

(16) $$\underset{{\theta }_G}{\mathrm{m}\mathrm{i}\mathrm{n}}\underset{{\theta }_D}{\mathrm{m}\mathrm{a}\mathrm{x}}{\mathcal{L}}_{\mathrm{G}\mathrm{A}\mathrm{N}}={\mathbb{E}}_{\mathbf{x}\sim p_{\mathrm{d}\mathrm{a}\mathrm{t}\mathrm{a}}}\left[\mathrm{l}\mathrm{o}\mathrm{g}D\left(\mathbf{x};{\theta }_D\right)\right]+{\mathbb{E}}_{z\sim p_z}\left[\mathrm{l}\mathrm{o}\mathrm{g}\left(1-D\left(\mathbf{G}\left(z;{\theta }_G\right);{\theta }_D\right)\right)\right]$$

4. Risk prediction network:

(17) $$\hat{y}=f\left(\mathbf{x};{\theta }_R\right)=\sigma \left({\mathbf{W}}_R\cdot \mathbf{x}+{\mathbf{b}}_R\right)$$

5. Objective Function for risk prediction:

(18) $$\underset{{\theta }_R}{\mathrm{m}\mathrm{i}\mathrm{n}}{\mathcal{L}}_R={\displaystyle \frac{1}{n}\sum _{i=1}^n\left[\alpha {\left(y_i-{\hat{y}}_i\right)}^2+\beta {\left(y_i-{\hat{y}}_{i,\mathrm{s}\mathrm{y}\mathrm{n}\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{t}\mathrm{i}\mathrm{c}}\right)}^2+\lambda \left(\sum _{j=1}^k{\theta }_j^2\right)\right]}$$

The model incorporates a GAN framework, which consists of a generator network and a discriminator network. The generator (refer to Eq. (14)) receives a random noise input vector and transforms it using a series of weight matrices, biases, and activation functions. This process produces synthetic data that aims to resemble real financial data. On the other side, the discriminator (refer to Eq. (15)) evaluates both real and generated data. It applies a similar transformation process with its own weights and biases, outputting a probability score that indicates whether the input is real or synthetic. The training process for the GAN involves an adversarial loss function (refer to Eq. (16)). Here, the generator attempts to minimize its loss by creating realistic data, while the discriminator tries to maximize its ability to correctly distinguish between real and synthetic data. This adversarial setup ensures that the generator continually improves the quality of the synthetic data. For predicting financial risk, the model includes a risk prediction network (refer to Eq. (17)), which applies a linear transformation to the combined input features using a set of weights and biases, followed by an activation function. The network generates the predicted risk values based on these transformations. The model’s overall objective function (refer to Eq. (18)) minimizes the prediction error, incorporating regularization terms to prevent overfitting. This loss function includes three components: the mean squared error between actual and predicted values, an error term for discrepancies in synthetic data predictions, and regularization penalties to ensure model stability. Together, these elements contribute to an effective risk prediction and analysis framework.

This formulation employs the use of GANs to synthesise financial data and reduces the problem of data deficiency as well as enhances the accuracy of risk prediction. The adversarial loss guarantees the synthesized data to be realistic while the combined loss improves the stability of the risk prediction network. Additionally, the framework includes a GAN component for synthesizing financial data, which addresses issues of data scarcity and enhances model robustness. The GAN consists of a generator network (G), which produces synthetic data, and a discriminator network ( $D$ D), which differentiates between real and generated data. The adversarial loss function optimizes both networks, ensuring that the generated data closely mimics the real data. The risk prediction network leverages the synthetic data alongside real data, minimizing a combined loss function that includes both the error in predicting real data and the discrepancy between real and synthetic predictions. This multi-modal approach allows the model to generalize better and provide more accurate financial risk assessments, even when faced with limited or incomplete data. In summary, the HFRP model uses a hybrid architecture combining CNN and LSTM to extract and analyze features from both textual and numerical data. The dynamic integration of reinforcement learning adapts to market changes, while the GAN module enhances data quality and model reliability. This comprehensive system aims to provide a robust solution for financial risk prediction, leveraging the strengths of DL techniques to offer a nuanced and effective approach to risk management.

Dataset description

This work’s data set is made up of a large number of documents containing financial reports of different companies. The text and numerical financial information are also provided in the disclosures. Table 2 contains a detailed description of the dataset:

The dataset spans over ten years, encompassing various economic conditions, which allows for robust model training and validation.

In Fig. 2 (bottom-right scatter plot), the increased number of data points per company likely reflects the presence of multiple financial entries over a period of time. Instead of representing a single point for each company, the scatter plot aggregates data across different fiscal years or quarterly reports, capturing variations in total assets and total liabilities for each firm. This approach helps visualize the financial trends and stability of the companies over time, providing a more comprehensive overview of their financial positions rather than a static snapshot. The recurrent entries for the same companies enhance the depth of analysis by highlighting temporal changes in their financial metrics.

Figures 2 and 3 illustrating data distribution, such as histograms and scatter plots, are not merely for showing basic statistics; they are essential for uncovering the underlying financial trends. For example, the histograms of revenue, net income, earnings per share (EPS), and total assets reveal important insights into the central tendency, skewness, and outliers in the financial data. The presence of outliers can indicate firms with extraordinary performance or extreme financial distress, both of which are critical for assessing risk. Additionally, scatter plots showing the relationship between total assets and total liabilities help identify potential liquidity issues. A linear trend suggests stable growth, while deviations from this trend can highlight companies at risk of financial imbalance. By delving into these patterns, we gain a deeper understanding of the financial health of the dataset, allowing for better model calibration.

In Fig. 2, their respective revenue, net income, EPS, as well as the distribution of total assets against the total liability of the identified companies in the dataset. The histograms give information about the distribution and measure of the central tendency of these financial metrics and the scatter plot shows how these firms’ total assets fare against their total liabilities. Figure 3 shows bar plots of operating income, cash from operating activities, total assets, and total liabilities. Such plots are useful for analyzing patterns and dispersion of the data which is essential for constructing and assessing the predictive models applied in this research.

Figure 4 shows the most important feature distribution. Instead of simply listing the financial metrics with their importance scores, this section should connect these scores to practical implications. The feature importance analysis, particularly using permutation importance, identifies key predictors for financial risk, such as net income, total assets, and total liabilities. These features rank highest due to their strong correlation with a company’s ability to meet its financial obligations and withstand market fluctuations. The emphasis on net income reflects its role as a direct measure of profitability, while total assets and liabilities are critical indicators of a company’s financial structure and risk exposure. Understanding these key features helps refine the model’s focus, improving its predictive performance and aligning it with financial realities.

Data preprocessing

The data is preprocessed to extract meaningful features. The following steps are carried out during data preprocessing:

Text preprocessing

Text preprocessing is crucial for extracting meaningful features from the textual disclosures in financial reports. Figure 5 shows the analysis of texts. The following steps are implemented:

• 1)

  Tokenization: Splitting the text into individual words or tokens.

• 2)

  Stop word removal: Eliminating common words that do not contribute to the analysis, such as “and”, “the”, etc.

• 3)

  Stemming and lemmatization: Reducing words to their base or root form to ensure uniformity.

• 4)

  Vectorization: Converting the text into numerical vectors using techniques like TF-IDF or Word2Vec.

Combining features

In this stage of data preprocessing, we’ve done the integration of the features extracted from both the textual and numerical data for a comprehensive analysis.

The textual disclosures’ cumulative occurrence of the top 20 words is illustrated in Fig. 6. This helps in comprehending some of the frequently used terms that appear in financial reports with the aid of which it may be important for the next steps of analysis and modeling.

Correlation analysis

It is necessary to remember that every single financial characteristic has its relationships with the other ones, and thus to determine the interconnections between various analytics is crucial for constructing effective math models. Figure 7 shows the results of the correlation test between the financial measures adopted in this study. The correlation matrices of all the financial ratios have been depicted in Fig. 7 for the convenience of understanding. It is also useful in feature selection and model building as it assists in the determination of the relations and interdependencies of the metrics.

The correlation analysis reveals critical interdependencies among financial metrics that are vital for achieving the research objectives. For instance, a strong positive correlation between revenue and net income highlights how revenue growth consistently translates into profitability, underlining its importance as a key feature in risk assessment. Similarly, a linear relationship between total assets and total liabilities indicates proportional scaling, which is beneficial for financial stability but raises liquidity concerns if liabilities grow disproportionately. Negative correlations, such as between operating income and total liabilities, signal potential red flags for firms with high debt struggling to maintain operational profitability. These insights directly support the design of the HFRP model by guiding feature selection and prioritization, ensuring the model focuses on metrics that significantly influence financial stability and risk. Additionally, identifying weak or anomalous correlations helps the model flag firms with unbalanced financial structures, enhancing its predictive accuracy. Correlation analysis also validates the use of LSTM networks for temporal dependencies and supports the interpretability of the model, providing stakeholders with clear, data-driven justifications for predictions. This alignment between data characteristics and model design ensures the framework’s robustness and reliability in real-world risk management scenarios.

Feature importance metrics

To establish the relevance of the various financial metrics in the computation of the risk we used permutation importance. Figure 8 exhibits the details of the feature importance of the HFRP model. The relative importance of each predictor can be seen in Fig. 8: The most important test has Net Income as the measure, while total assets and total liabilities come next. They have the highest importance in the model, which overall suggests that they are probably the most relevant features in assessing financial risk.

Principal component analysis and clustering

Below is the cluster analysis based on principal component analysis (PCA) on the following financial metrics after processing the data collected, the following is the conclusion. It also becomes easy to find clusters of companies with characteristics that differentiate them financially. Figure 9 presents the clustering of the financial metrics in which they are distinguished in terms of color at each cluster. The companies in each cluster are listed below:

Figure 9 shows the clustering of financial metrics, with each color representing a different cluster. The initial description of PCA lacked depth and did not justify its usage. PCA is employed to reduce the dimensionality of the financial data, making it more manageable for clustering without losing significant information. By transforming the original features into a set of uncorrelated principal components, PCA captures the variance and underlying structure in the data. The first few principal components usually explain a substantial portion of the variance, highlighting the key drivers of financial performance. In this study, PCA not only aids in visualization but also helps identify patterns related to financial stability and risk, such as distinguishing between firms with strong asset growth vs those relying heavily on liabilities. This dimensionality reduction is critical for effective clustering and subsequent analysis. The discussion on PCA and clustering should emphasize how the identified clusters represent distinct financial risk profiles and inform targeted mitigation strategies. Each cluster, based on key financial metrics like net income, total assets, and liabilities, reflects groups of firms with shared financial characteristics. For example, a cluster with low net income and high liabilities could indicate firms at high risk of financial distress, necessitating strategies focused on debt reduction and liquidity management. Conversely, clusters with high revenue and assets may represent stable firms with minimal risk, requiring less aggressive mitigation. By analyzing these differences, the clustering approach provides actionable insights, enabling tailored interventions for each risk group. This targeted strategy not only optimizes resource allocation but also strengthens the predictive accuracy of the HFRP model by aligning risk assessment with the unique financial dynamics of each cluster. Explaining these distinctions and linking them to specific mitigation plans would enhance the clarity and practical relevance of this section. The companies in each cluster are listed in Table 3.

Proposed model: Hybrid Financial Risk Predictor (HFRP)

We developed a new model named the Hybrid Financial Risk Predictor (HFRP) which includes CNN and LSTM for dealing with text and numerical data of financial reports. To address the issues of risk prediction, the interdependencies within the data are modelled based on a complex structure of the architecture.

There is a general framework of HFRP which is a model of financial architecture that can effectively handle financial data and evaluate the level of risk. In Fig. 10, it is illustrated that the model comprises of input layer, three densified layers, and last layer is the output layer. Preprocessing is important in the model because it prepares the data for analysis and reduces the input data’s variability before it is fed into the model. It’s common to arrange the dense layers in such a way that each layer performs a sequence of transformations on the input data to enable the model to infer complex features. The training procedure includes adapting various parameters of the model to obtain the minimum of the loss function and then the model will be tested to compare with some indicators. The obtained results inform about the performance of the proposed model for risk estimation and prediction. Clustering financial data helps uncover hidden patterns and groups of companies with similar risk profiles. This process is not arbitrary; it is designed to identify homogeneous groups based on their financial metrics, which can be highly valuable for risk management. For instance, companies within a high-risk cluster might exhibit low net income, high leverage, and frequent cash flow issues, indicating potential distress. Identifying such clusters enables financial analysts to focus on specific groups that require more attention and tailored risk assessment strategies. Furthermore, clustering allows the model to refine its predictions by adjusting risk parameters based on the characteristics of each group, enhancing the overall accuracy and interpretability of the risk predictions.

Model architecture

1. Input layer: The model accepts two types of inputs: textual data and numerical data.

(19) $$\mathbf{X}\in {\mathbb{R}}^{n\times m},\mathbf{T}\in {\mathbb{R}}^{n\times p\times q}$$

2. CNN for textual data: The textual data is processed using a CNN to extract relevant features.

(20) $${\mathbf{F}}_i^{\mathrm{C}\mathrm{N}\mathrm{N}}=\sigma \left({\mathbf{W}}_{\mathrm{c}\mathrm{o}\mathrm{n}\mathrm{v}}\ast {\mathbf{T}}_i+{\mathbf{b}}_{\mathrm{c}\mathrm{o}\mathrm{n}\mathrm{v}}\right)$$

3. LSTM for numerical data: The numerical data is processed using an LSTM network to capture temporal dependencies.

(21) $${\mathbf{h}}_t=\varphi \left({\mathbf{W}}_h\cdot {\mathbf{x}}_t+{\mathbf{U}}_h\cdot {\mathbf{h}}_{t-1}+{\mathbf{b}}_h\right)$$

4. Feature combination: The features extracted from both the CNN and LSTM are combined.

(22) $${\mathbf{F}}_i=\left[{\mathbf{F}}_i^{\mathrm{C}\mathrm{N}\mathrm{N}},{\mathbf{h}}_T\right]$$

5. Output layer: The combined features are fed into a fully connected layer to predict financial risk.

(23) $${\hat{y}}_i=f\left({\mathbf{F}}_i\right)={\mathbf{W}}_{\mathrm{o}\mathrm{u}\mathrm{t}}\cdot {\mathbf{F}}_i+{\mathbf{b}}_{\mathrm{o}\mathrm{u}\mathrm{t}}$$

Loss function

The loss function combines mean squared error with regularization terms to prevent overfitting:

(24) $$\underset{\theta }{\mathrm{m}\mathrm{i}\mathrm{n}}\mathcal{L}={\displaystyle \frac{1}{n}\sum _{i=1}^n\left[\alpha {\left(y_i-{\hat{y}}_i\right)}^2+\beta \left(\sum _{j=1}^k{\theta }_j^2\right)+\gamma \sum _{j=1}^m\left|{\theta }_j\right|\right]}$$

Training process

The model is trained using historical financial data, including both numerical metrics and textual disclosures. The training process involves:

• 1)

  Data splitting: Dividing the dataset into training, validation, and test sets.

• 2)

  Parameter optimization: Using techniques like gradient descent to optimize model parameters.

• 3)

  Validation: Evaluate the model on the validation set to tune hyperparameters and prevent overfitting.

This novel approach leverages the strengths of both CNN and LSTM, providing a robust framework for financial risk prediction. The integration of textual and numerical data ensures comprehensive analysis, leading to more accurate and reliable risk assessments.

Evaluation metrics

To evaluate the performance of the proposed HFRP model, we utilize a comprehensive set of metrics that focus on both risk management and regression accuracy. These metrics provide a holistic view of the model’s effectiveness in predicting financial risks and accurately forecasting financial metrics. The evaluation metrics for the HFRP model are shows in Table 4.

Definitions

• Risk score (R): The average risk score assigned to each instance, indicating the overall risk level.

• Credit risk (CR): The average score indicating the risk of default on credit obligations.

• Market risk (MR): The average score indicates the potential for financial loss due to market fluctuations.

• Operational risk (OR): The average score indicating the risk of loss resulting from inadequate or failed internal processes, people, and systems.

• Liquidity Risk (LR): The average score indicating the risk of a company’s inability to meet its short-term financial obligations.

• Mean squared error (MSE): A measure of the average squared differences between actual and predicted values.

• Root mean squared error (RMSE): The square root of the mean squared error, providing a measure of the average magnitude of errors.

• Mean absolute error (MAE): The average of the absolute differences between actual and predicted values.

• R-squared ( ${\mathit{R}}^2$): A statistical measure representing the proportion of variance in the dependent variable that is predictable from the independent variables.

Training and testing loss

Figure 11 presents the training and testing loss of the HFRP model over 50 epochs.

Figure 11 shows a steady decline in both training and testing loss, indicating that the model is learning effectively and generalizing well to unseen data. The final training loss is 0.0013, and the testing loss is 0.003, which demonstrates the robustness of the model. The training and testing losses across different epochs are also shown in Table 5.

Prediction accuracy

The accuracy of the HFRP model was assessed by comparing the predicted values of various financial metrics with the actual values. Figure 12 presents the comparison of several key metrics.

Figure 12 and also Table 6 illustrate the close alignment between the actual and predicted values for metrics such as revenue, net income, EPS, total assets, total liabilities, operating income, and cash flow from operations. The near-perfect alignment along the ideal line (red) indicates the high accuracy of the model’s predictions.

Risk levels before and after mitigation

To evaluate the risk mitigation effectiveness of the HFRP model, we compared the risk levels before and after applying the model. Figure 13 shows the risk levels for different risk types.

Figure 13 shows a significant reduction in risk levels across all types of risks, including credit risk, liquidity risk, market risk, and operational risk. The HFRP model effectively mitigates these risks, demonstrating its potential to enhance financial stability.

In Table 7 the “before” risk levels were determined using traditional risk assessment methods commonly employed in financial analysis. Specifically, we used a standard statistical model based on a combination of financial ratios, including debt-to-equity ratio, current ratio, and return on assets (ROA). These metrics provide a conventional evaluation of a company’s financial health and risk exposure. The baseline risk scores were calculated by applying these traditional metrics to the dataset without incorporating advanced DL methods. This approach reflects the typical risk assessment performed by financial analysts, relying on historical averages and basic trend analysis.

Risk score distribution

The distribution of risk scores, credit risk, market risk, operational risk, and liquidity risk for the companies in the dataset is shown below.

Figure 14 displays the histograms of different risk scores. These distributions provide insights into the overall risk landscape and the effectiveness of the model in managing and mitigating these risks. The different risk score distribution is also shown in Table 8.

Average risk scores by cluster

The average risk scores for different clusters of companies provide further insights into the effectiveness of the HFRP model.

Figure 15 shows the average risk scores for different clusters, indicating how risk levels vary across different groups of companies. This analysis helps in understanding the risk profile of different clusters and the targeted risk mitigation strategies. Additionally, Table 9 shows the average risk score by cluster in tabular form.

Risk reduction percentage by risk type

The percentage reduction in risk levels for different risk types is shown below.

Figure 16 and Table 10 provide the percentage reduction of risk levels when using the HFRP model on the investment. The model also successfully attains a large decrease in all the risk categories, with credit risk having the maximum degree of decrease. That can be concluded from the results that the presented HFRP model-based CNN-LSTM is more efficient at predicting financial indicators and eliminating multiple types of financial risks. The gradual decline of the training and testing loss demonstrates that the model had strong learning and even better generalization ability. The correlation coefficients between actual and predicted values demonstrate a strong fit for most of the financial variables analysed, as manifested by the nearness of the actual and predicted values.

Training and testing loss analysis

The gradual decrease of training and testing loss within 50 epochs with the values 0. 0013 and 0. Organization 2 with a score of 003 and Organization 3 with a score of 003 demonstrate the model’s strong learning function. The values of the minimal loss confirm that the relations have been replicated well by the model and thus it can successfully act in new conditions and with new inputs. This robustness is crucial for the banking industry, where the accuracy of the predictions has a significant and tangible effect on the organization’s decisions and assessment of risks.

Prediction accuracy

The precision of the HFRP model is evident from the manner in which the actual and the resultant predicted values are close to each other for all the financial ratios. For instance, the forecasted amount of annual sales of $9,950,000 was close to $10,000,000 of actual sales. Similar precision was observed in other metrics such as net income (predicted: In terms of total sales, it had set a target of $1,980,000 while the actual figure recorded was $2,000,000. On the aspect of EPS, an average of $3.48 had been projected while the actual average EPS was $3.50. Such accuracy is especially important to financial analysts because it increases the credibility of the predicted financial statement results which in turn assist in strategic planning and management of resources.

Risk mitigation effectiveness

Thus, one of the most important outcomes of this work is to note the approximately five-fold decrease in risk levels provided by the HFRP model. Credit risk was down by 0.75 to 0.20: liquidity risk from 0.70 to 0.25, and market risk from 0. The company’s beta coefficient is 1: The beta coefficient is an index that shows how much a change in the overall market index will affect the firm’s stock price 0.65 to 0.30, while the operational risk mean value is at 0.80 to 0.35. This significant level of risk mitigation on all the risk types shows that the model is very useful in risk management for financial risks. Some of the findings are striking; the reduction of credit risk down to as little as 30% for instance can be considered quite dramatic since it suggests a step change in the institution’s ability to handle debt and hence liquidity.

Risk score distribution and clustering analysis

To be specific, the comparison of risk score distribution offers an understanding of the risk profile by dollars at risk. The distribution of different risk scores is presented in the histogram to show how risks are distributed in the companies and use that as a way to address the issue. Also, the results for the clustering analysis suggested average risk scores of 0.40 for Cluster 0, 0.35 for Cluster 1, and 0 for the other. 30 of Cluster 2 assists in the identification and analysis of various degrees of risk related to groups of companies. This information can prove to be very useful in the formulation of risk management solutions that are unique to each cluster.

Implications for financial institutions

The success of the HFRP model in the aspect of giving good predictions and managing risks carries several significant implications for financial institutions. First, financial accuracy or, specifically, the high level of forecasts’ accuracy in terms of financial performance allows for improving the strategies and the company’s efficiency in general. Second, increased claimants’ risk levels improve financial stability, thereby decreasing the chances of financial failure. Third, the specific delineation of risks and their location helps to be more efficient in risk management and choice of priorities.

In summary, it is apparent that the employment of advanced deep learning paradigms including CNN and LSTM in financial risk management facilitates a groundbreaking way of dealing with intricate financial information. The study has also sheltered the HFRP model in estimating and decreasing the risks involved thereby revolving the ways various models are used in analysing the business turmoil and managing the risks involved. Another area for research may include examining the application of this model with other kinds of financial risks and enhancing its accuracy to guarantee steady usefulness in such a highly volatile field as the financial one.

Limitations/Validity

The study addresses limited application of deep learning in creating model to mitigate contemporary risks. Although the research presents a rather encouraging way, it however concedes that additional work needs to be done in order to determine the worthiness and usefulness of deep learning models in other financial environment, especially in real time risk monitoring and appropriate decision support systems. Moreover, the quality of the input data is one of the preconditions for the success of the HFRP modeling. If the model relies on the use of historical financial data but which is either incomplete, old or biased, then the forecasts generated by such a model will tend to be erroneous and less dependable. Another possible limitation of this study is that the model development incorporates certain assumptions regarding the data and the underlying market behaviours and these assumptions do not always bring true in reality. This may constrain its use in other financial markets or such as when the actual market is very different from the one assumed when the model was designed.

The validation of the proposed model is as follows:

Extensive training and validation: It is based on a huge amount of historical data. Large-scale model training helps to learn various financial patterns under conditions. The use of validation sets on separate datasets fine-tunes the model parameters avoiding overfitting which in turn has a broader validity for these results.

Use of standard evaluation metrics: Many standard metrics accuracy, precision, recall, and F1-score were used to see how well our model performed. Such metrics are popular in the field and they make sure that whatever model we are being trained, predictions from it can not only be accurate but be reliable.

Risk mitigation: The article demonstrates that the model is quite good at reducing risk, by showing declines in credit, liquidity, market, and operational risks from before to after firms are placed into its system.