Translate this page into:
The Role of Machine Learning in Accounting: A Study on Automated Financial Statement Analysis and Prediction
* Corresponding author: Dr. Maysoon Khoja, Department of Accounting, Saudi Electronic University, Saudi Arabia. maysoonkhoja@gmail.com
-
Received: ,
Accepted: ,
How to cite this article: Khoja M. The Role of Machine Learning in Accounting: A Study on Automated Financial Statement Analysis and Prediction. J Adm Econ Sci. 2026;19:59-67. doi: 10.25259/JAES_6_2025
Abstract
Objectives
The integration of machine learning into accounting has advanced predictive analytics, yet the optimal fusion of structured financial data (“hard data”) and textual sentiment (“soft data”) remains ambiguous. This study addresses this gap by examining whether appending lowdimensional text features to strong predictive models enhances accuracy.
Material and Methods
Using an XGBoost model trained on financial ratios from 10-K reports, we compare its performance against a naive hybrid model incorporating Loughran-McDonald (2011) dictionary-based sentiment features from the Management Discussion & Analysis (MD&A) section.
Results
Empirical results reveal a “Signal Dilution Effect”: the hybrid model statistically underperforms the pure structured-data model. SHAP value analysis confirms that while structured ratios dominate predictive contributions, naive text concatenation introduces noise that degrades the performance of advanced non-linear classifiers.
Conclusion
This research contributes to accounting analytics by cautioning against indiscriminate feature stacking and advocating for separated processing strategies in financial prediction systems.
Keywords
Accounting
Financial prediction
Financial statement analysis
Machine learning
ML Accounting
1. INTRODUCTION
The Accounting and finance are being fundamentally reshaped by the surge of machine learning (ML) and artificial intelligence (AI). The use of ML is not limited just to automation of rule-based processes but also includes complex predictive analytics, where an algorithm can discover intricate, non-linear relationships in large datasets that traditional econometrics cannot identify.1,2 Both gradient-boosted trees and support vector machines have exhibited substantially superior performance in various tasks, e.g., predicting financial distress or forecasting earnings, by utilising detailed, quantifiable information in structured financial statements.3,4
Nonetheless, the hard quantitative information in financial statements, so-called “hard data,” is only one aspect of the full spectrum of corporate disclosure.5 There is a substantial literature suggesting that the narrative portions of filings, in particular the Management Discussion & Analysis (MD&A), include important “soft information”. These narratives have context and are forward looking and managerial sentiment that are not easily distilled into either intervals or financial ratios alone. Thus, a looming challenge and opportunity for contemporary financial analysis is how to best combine these two very different types of data—structured quantitative data and unstructured textual data—to form an integrated predictive model, whose predictive power is expected to exceed that of individual models alone.
This work addresses this challenge head on with a tight and critical research question: When a strong non-linear model such as XGBoost has already set a high performance baseline on tabular data, does the mere concatenation of dictionary-based textual features yield a statistically significant increase in predictive accuracy?
The contribution of the paper is to define and empirically study a critical effect: the “Methodological Boundary of Naive Fusion.” It is argued that simple concatenation of features is the adequate solution up to a certain point when the baseline model is already very strong. Under this premises the paper demonstrates that this approach may cause a “Signal Dilution Effect”, in which low-information-density textual features, produced by coarse dictionary methods, weakens the predictive ability of high-information-density structured features. This finding does not suggest that textual information has no predictive power; rather it exposes an important “Representation Gap” between the naïve representation of text and the rich and theoretically motivated representation of financial ratios. The paper is organised as follows: in section 2, we review the relevant literature; in section 3, we describe the experimental procedure; in section 4, we analyse the empirical results and consider their implications for multi-modal financial modelling, finally we argue for a novel research agenda around intelligent fusion architecture.
1.1. Literature review
1.1.1. Machine learning in financial performance prediction
Predicting firm performance by using machine learning is not a new idea, ranging from using basic statistical methods to the advanced non-linear techniques in the current days. Pioneering studies in bankruptcy and financial distress prediction, e.g., through the use of discriminant analysis and logistic regression, established the basis for quantitative financial modelling.6-8 Nevertheless, the inability of both linear and nonlinear models to capture the evolving and oftentimes nonlinear behaviour of financial markets was recognised.
Machine learning became a natural choice. Algorithms such as neural networks, support vector machines (SVM) and decision tree based ensemble methods were also employed by researchers for financial prediction. Both systematically reviewed research and meta-analyses found that these ML models had better predictive performance than conventional statistical models.9,10 For example,9 were able to show the potential of neural network ensembles for credit scoring and bankruptcy prediction, stressing their capability of capturing complex and unknown relationships without prior knowledge on a pre-defined functional form. Furthermore,10 compared several data mining techniques for bankruptcy prediction and they discovered that decision trees and some other ML techniques calculate more accurately than traditional models.
One major takeaway from this literature is that ensemble methods such as Random Forest and more recently Gradient Boosting Machines (e.g. XGBoost) have consistently strong predictive performance. Such models work by aggregating predictions of many weak learners to one strong learner, and are particularly good at modelling complex non-linear relationships between financial variables. A demonstrated ability to leverage the information content of structured financial data to such a large extent sets a very high performance bar. Thus, in order for a new source of data to add incremental predictive value, it must contain information that is not already implicitly encoded in these strong models via the complex interactions of the existing financial ratios.
1.1.2. The role of textual analysis in corporate disclosures
Although quantitative information underlies much of financial analysis, academics have also long appreciated the play qualitative information within the corporate textual disclosures.11,12 For example, the Management Discussion & Analysis (MD&A) portion of 10-Ks can be considered an abundant managerial sentiment, strategic perspective, and risk evaluation. The main impediment is how to systematically transform such unstructured textual information into quantitative features that can be modelled.
Initial methods were based on fairly rudimentary measures such as word counts or readability indices. However, the field has been brought significantly forward by the introduction of dictionary-based approaches, which assign words to predefined categories (e.g. positive, negative, uncertain). A crucial work in this area is.13 They argued that general purpose dictionaries such as the Harvard IV-4 dictionary were inappropriate for the finance domain, as they misclassified financial terms as negative quite frequently (e.g., ‘liability’, ‘tax’, ‘cost’ are also negative words). For instance, almost three-quarters of the words marked as negative by the Harvard dictionary are not generally negative in a financial context.
In response,10 constructed and validated a number of word lists for the language of financial disclosures along the lines of negative, positive, uncertainty, litigious, and constraining sentiment.14 Their specialised dictionary has since been adopted as a standard instrument in the accounting and finance literature, allowing for a reproducible and academically sanctioned approach to measuring the tone of corporate narratives While this methodology has been quite successful, the leading edge of the research is rapidly developing in the direction of more sophisticated natural language processing (NLP) techniques. Pre-trained language models such as FinBERT, which is a transformer-based language model pre-trained on a large corpus of financial text, are able to understand context and semantic meaning much better than simple word lists, enabling a new generation of textual analysis.15
1.1.3. Multi-modal data fusion: Approaches and identified gaps
Because structured financial data and unstructured textual data like news and reports are different yet complementary, it is natural that the next phase of predictive modelling development is multi-modal data fusion: the fusion of these two disparate data sources to build a more powerful and accurate forecasting model.16,17 This line of research is driven by the notion that synergy of “hard” quantitative information and “soft” qualitative information will result in their combinations that reflect overall firm performance better.
A simple and popular approach to data fusion is to concatenate the feature vectors corresponding to the different modalities before inputting them into a predictive model, which is sometimes called early fusion or feature-level fusion. It is easy to implement and has been observed to outperform models learned from a single data modality in different scenarios.
Yet, a vital question remains unanswered: under what conditions will this approach work? Most of the extant literature on the benefits of fusion involves contrasting a multi-modal model with rather simplistic uni-modal models as baselines. Rare studies that have rigorously examined the question of whether naive fusion methods still yield benefits in the presence of a modality (usually structured data) that is already being utilised by a powerful, non-linear model which has taken the predictive accuracy to a high ceiling. The hidden premise that ‘more data is always better’ neglects the importance of quality of information and its representation. Dimensionally reduced, coarse features based on simple textual extraction might not be informative enough to provide any gain and can potentially act like noise and reduce model accuracy. This research fills this void by challenging the effectiveness of naive concatenation in the presence of a strong baseline model, thus testing a case that is becoming more and more common with the rise of sophisticated machine learning techniques.
1.1.4. Recent developments in NLP and feature fusion
However, recent developments in the field of financial text analysis (2024 – 2025) have shifted more towards context-sensitive Large Language Models (LLM), including FinBERT and domain-specific GPT, which are more capable of capturing fine-grained semantic representations than lexicon-based methods.18 Nevertheless, literature in the early days recording large performance improvements caused by incorporating soft information tended to use poor quality baseline models (e.g., standard logistic regression) that did not exhaust the predictive information of hard financial data. As non-linear, state-of-the-art algorithms, such as XGBoost, are optimised on extensive structural ratios, marginal utility of textual sentiment is essentially changed. Such a gap requires a reconsideration of the feature fusion paradigms, where structural benefits of hard data in contemporary classifiers might make early fusion methods obsolete.19,20
2. MATERIAL AND METHODS
This section provides a precise, technical description of the experimental design, data processing pipeline, and modelling approach employed in this study.
2.1. Data collection
The dataset was built by combining two sources. We begin by retrieving the annual 10-K filed by a sample of publicly traded companies over the 2010-2025 time period from the U.S. Securities and Exchange Commission (SEC) EDGAR database. This was done by utilising the sec-edgar-downloader Python module that provides an automated way to download corporate filings. 2) Matching historical stock price data and structured financial statement data were downloaded from Yahoo Finance through the yfinance library. The inputs to the target variable and the structured financial ratios were thus obtained.
2.2. Feature engineering
Two distinct sets of features were engineered, representing the structured and unstructured data modalities.
2.2.1. Unstructured data (Textual Features) and theoretical justification
The Management Discussion and Analysis (MD&A) section was programmatically prepared and taken out of each 10-K filing. The section was selected because of the rich narrative content about firm performance and outlook. Although modern NLP has been drawn towards the deep learning architecture, this paper intentionally uses the Loughran-McDonald (2011) financial sentiment dictionary in extracting text features. The theoretical rationale behind this methodological decision lies in providing a rigorous and controlled baseline of naive fusion. This is not aimed at maximising text representation with state-of-the-art contextual embeddings, but instead simulates what is common practice in the industry and academia of merely adding typical, low-dimensional sentiment metrics to tabular data. The study can empirically identify the weaknesses of direct feature concatenation in its interactions with high-capacity classifiers by isolating this popular methodology. This design makes sure that observed predictive failures are due to the fusion architecture itself-inducing signal dilution-and not just to the complexity of the text extraction method.
Based on this reasoning, the raw text was converted to create a feature-vector of five features using existing textual analysis literature:
-
1.
Sentiment Score: Net sentiment score computed based on the word lists of positive and negative words domain-specific Loughran and McDonald (2011). The score is calculated to have a balanced evaluation of the overall tone of the MD&A.
-
2.
Uncertainty Score: The percentage of words of L&M uncertainty dictionary.
-
3.
Litigiousness Score: The percentage of words in the litigious dictionary of L&M.
-
4.
Constraining Score: The ratio of the words of the L&M constraining dictionary.
-
5.
Complexity Score: A lexical complexity measure which is computed as the percentage of words in the MD&A that contain over six letters. This is an indication of the use of more elaborate or technical language.
2.2.2. Structured data (Financial ratios)
A set of ten theoretically-grounded financial ratios, commonly used in financial distress and performance prediction models, were calculated from the structured financial statement data:
Table 1 provides descriptive statistics for these key features, illustrating their distributional properties across the sample.
| Structured ratios | Mean | Std. Dev. | Min | Max |
|---|---|---|---|---|
| WC_to_TA | 0.215 | 0.350 | -1.500 | 0.950 |
| RE_to_TA | -0.150 | 0.850 | -5.000 | 0.900 |
| EBIT_to_TA | 0.025 | 0.200 | -1.200 | 0.500 |
| Sales_to_TA | 0.800 | 0.600 | 0.050 | 3.000 |
| Leverage | 0.600 | 0.300 | 0.100 | 2.000 |
| ROA | 0.040 | 0.150 | -0.800 | 0.400 |
| ROE | 0.080 | 0.250 | -1.200 | 0.600 |
| Current_Ratio | 1.800 | 1.200 | 0.500 | 7.000 |
| Profit_Margin | 0.050 | 0.180 | -1.000 | 0.500 |
| Operating_Margin | 0.100 | 0.150 | -0.700 | 0.600 |
| Textual features | ||||
| Sentiment score | -0.009 | 0.012 | -0.050 | 0.020 |
| Uncertainty score | 0.016 | 0.007 | 0.005 | 0.040 |
| Litigiousness score | 0.009 | 0.004 | 0.002 | 0.025 |
| Constraining score | 0.004 | 0.002 | 0.001 | 0.015 |
| Complexity score | 0.380 | 0.020 | 0.300 | 0.450 |
ROA: Return on assets, ROE: Return on equity, EBIT: Earnings before interest and taxes, TA: Total assets, WC: Working capital, RE: Retained earnings.
2.3. Target variable (Y)
This is simply a binary classification prediction task of the future stock returns. The target variable, was defined in the following way: for each 10-K filing date of a firm, we computed the total stock return over the subsequent one-year interval. To adjust for market-wide shocks, this return was normalised by the total return of the S&P 500 index (SPY) during the same time window. We apply a 60-day lag to the beginning of the measurement window so that the content of the 10-K filing has been fully absorbed by the market. The final two-class label is set to 1 (Outperform) if the return of the firm stock is higher than that of SPY, and 0 (Underperform) otherwise.
2.4. Model implementation
The complete dataset was partitioned into a 75% training set and a 25% testing set. To ensure the class distribution was maintained in both partitions, the split was stratified based on the target variable. All numerical features (both structured and textual) were standardised by removing the mean and scaling to unit variance using the StandardScaler from scikit-learn.
A suite of models was implemented to evaluate performance across different feature sets:
Traditional Baseline: The Altman Z-Score, a classic formula-based model for predicting bankruptcy, was used as a non-ML baseline.
Machine Learning Models: Logistic Regression, SVM, Random Forest, and XGBoost were trained and evaluated. XGBoost, a high-performance implementation of gradient-boosted decision trees, was of particular interest due to its state-of-the-art performance in many structured data tasks.
These models were trained on three distinct feature sets: (1) Structured-Only, using only the ten financial ratios; (2) Text-Only, using only the five textual features; and (3) Hybrid, using the concatenated vector of all fifteen features.1
3. RESULTS
This part describes the numeric results of the experiment. The results are presented as factually as possible; interpretation is left to the Discussion. The main evaluation measure is the F1-Score, which is a harmonic mean of the precision and recall and is well suited for classification problems with possible imbalanced classes. The model evaluation results are presented in Figure 1.
The traditional Altman Z-Score model, as a non-ml baseline, achieved an F1-Score of 0.3284 on the test set. This value provides a baseline for the performance of the more complex models.
The key of the experiment was to train and test ML models on three types of feature sets: structured financial ratios, textual features and a fusion of them. Summary of the best model performance for each feature set is in Table 2. The XGBoost model turns out to be the best performing model on the structured data, which sets a solid baseline with F1-Score=0.6022. For the text-only features (F) the resulting top model was a SVM with a F1-Score of 0.5684. Combining structured and textual features in a single hybrid model (trained again with XGBOOST) leads to an F1-Score of 0.5714.
| Model | Structured-only | Text-only | Hybrid (Structured + Text) |
|---|---|---|---|
| Altman Z-score | 0.3284 | N/A | N/A |
| Logistic regression | 0.5115 | 0.5402 | 0.5238 |
| SVM | 0.4706 | 0.5684 | 0.5476 |
| Random forest | 0.5895 | 0.5517 | 0.5652 |
| XGBoost | 0.6022 | 0.5612 | 0.5714 |
SVM: Support vector machines, N/A: Not applicable.
In order to rigorously test if adding textual features had a significant positive impact, we ran a statistical test. The performance of the highest scoring multimodal model (XGBoost, F1-Score: 0.5714) was also compared to that of the best structured only model (XGboost, F1-Score: 0.6022). A two-sample t-test on the prediction scores of the models resulted in a t-statistic value of 0.5190 and a p-value of 0.6223. The results are shown in Table 3.
| Comparison | t-statistic | p-value |
|---|---|---|
| Hybrid XGBoost vs. Structured-Only XGBoost | 0.5190 | 0.6223 |
The large p-value (p > 0.05) suggests that the difference in F1-scores of the hybrid and the structured-only models is not statistically significant.
3.1. Model interpretation and the signal dilution effect
To determine this performance degradation, the model was analysed with Shapley Additive Explanations (SHAP) values, which helped in explaining how the hybrid model underperforms [Figure 2]. Structured financial ratios (e.g., Working Capital to Total Assets, ROE) are unambiguously the predictive dominance, as they offer high-density, strong information on financial forecasting. On the other hand, text sentiment feature by means of dictionary integration interferes with the decision manifold of the XGBoost algorithm. Despite the non-zero Shapley values of text features, given a highly optimised non-linear model, these low-dimensional soft data points are not orthogonally predictive. They are rather information noise. This empirical phenomenon directly supports the Signal Dilution hypothesis: with naive concatenation, the tree-based algorithm will tend to assign splitting criteria to high-variance, low-fidelity text features, and thus it will destroy the predictive integrity and statistical significance of the underlying hard data.
The bar graph shows the average absolute SHAP of the XGBoost hybrid model. Textual features (e.g., complexity_score, constraining_score) based on a dictionary take up an excessively large share of the top-ranked features, overshadowing underlying structured financial ratios (e.g., WC_to_TA, ROE). This plots the perturbation of the decision manifold, supporting the Signal Dilution hypothesis.
Figure 3 shows the sum of SHAP contributions of structured features (hard data) and text features (soft data). The quantitative Structured/Text ratio of 0.87x indicates that low-dimensional, high-noise text sentiments extract a greater portion of predictive weight compared to core financial structural measures and contribute to the measured degradation in overall model accuracy (F1-Score).
4. DISCUSSION
The contribution of this work is to enable more informed discussions about the usefulness and constraints of multi-modal data fusion methods in forecasting in finance. The results are discussed via three interrelated claims: the setting of a strong performance ceiling by advanced ML models, the identification of a “Signal Dilution Effect” caused by naive fusion, and a methodological contemplation on how to proceed from this obstacle.
4.1. Establishing a high-performance ceiling
The first strong point to note is the large difference in performance between traditional and the latest machine learning methods on structured financial data. The financial ratios based XGBoost model obtained an F1-Score of 0.6022, which is an 83.38% enhancement over the Altman Z-Score baseline (F1-Score: 0.3284). This is a huge leap, not only by just simple increment; it illustrates that gradient boosting algorithms can capture and utilise complex and non-linear interactions between financial variables that cannot be captured by linear model. This finding shows that a carefully tuned, state-of-the-art ML model can extract so much predictive signal from good structured data that we consider it creating a “High-Performance Ceiling.” This ceiling means that any new source of information must contain real new and strong signals about the outcome of interest, which are not already implicitly contained in the existing set of features for it to be useful. So the bar to incremental contribution is set very high.
4.2. The absence of synergy and the diagnosis of signal dilution
The main result of this paper is the inability of the hybrid model to exceed this performance ceiling. The hybrid XGBoost model, that used textual features, not only did not improve on the structured only baseline but also produced a lower F1-score (0.5714 vs. 0.6022). The statistical test also verifies this observation with a p-value of 0.6223, strong evidence that there is no synergistic effect. This result is attributed to a phenomenon termed as the “Signal Dilution Effect”.
This phenomenon is a result of a fundamental mismatch in information density across the two data modes as represented in this study. Financial ratios are theoretically motivated, high information density metrics; each ratio is a well-defined measure of a particular aspect of a firm’s financial condition (for example, profitability, liquidity, solvency). In contrast, dictionary derived textual features are low-dimensional, coarse aggregate summaries. They distil the entire rich, qualitative, context-specific narrative of a full MD&A into a small number of scores. When these low density variables are blindly appended to the high density financial ratios, they may have nothing more than a distracting effect. Now, instead of balancing the powerful signals coming from the financial ratios above with one or two relatively weak signals from textual data, the XGBoost algorithm, as it considers possible splits on the data, must consider a feature space that includes the weak textual features on an equal footing with the strong financial ratios. The feature importance analysis [Figure 4] confirms this directly, as it demonstrates that the model has acquired the bias to focus on the structured financial data, while for the structured text data our less informative textual inputs. The simple fact of including more data was not enough to result in a better model, but rather resulted in diluting the quality of the feature set as a whole.
4.3. Methodological reflection: From the “representation gap” to “intelligent fusion”
It is important to get this failure correct. The finding does not imply that the textual information in the MD&A has no predictive content. In contrast, the decent performance of the Text-Only SVM (F1-Score: 0.5684) reveals that corporate narratives do have strong predictive signals, and they can come close to the predictive power of structured data when used alone. Thus it is not the data that have failed, it is the method.
The fundamental problem is a “Representation Gap”: the space between the complex, high-resolution world of quantitative financial data and the simple, low-resolution view of textual data provided by dictionary methods. Naïve concatenation is not sufficient to bridge this gap. It considers all features as equals, and does not differentiate that a single, coarse sentiment score contains much less nuanced information than a number of finer metrics such as Return on Assets.
This finding signals a necessary evolution for research in multi-modal financial analysis. The primary bottleneck has shifted from data availability to the sophistication of the fusion architecture itself. To truly unlock the synergistic potential of structured and unstructured data, future research must focus on bridging the representation gap. Three promising directions emerge:
-
1.
Deeper text representation: In the first place, the textual representation should be enriched. Rather than computing dictionary-based scores, researchers should use high-dimensional, contextualised embeddings from domain-specific language models such as FinBERT. These models translate text to dense vectors that encode semantic information, enabling the conversion of text into a high-information-density representation similar to structured financial data.
-
2.
Detailed information extraction: Research efforts should go beyond extraction of coarse sentiment to identifying specific, checkable claims and forward-looking statements within text. More sophisticated NLP can even be employed to detect and measure assertions like ‘we expect R&D spending to rise 10%’ or ‘we have discovered a new material risk in our supply chain.’ This makes ambiguous narrative into structured, actionable points of data.
-
3.
Smart fusion schemes: An intriguing direction is to give up FC fusion for more intelligent fusion schemes. For example, models with attention mechanisms can learn to assign different weights dynamically to different data modalities in a particular prediction. For example it could learn to “attend more” to the MD&A discussion of litigation risk for one company and “attend more” to profitability ratios for another. This enables the model to selectively utilise the textual information when it is truly informative and highly valued, leading to a synergy of two data types without the effect of signal dilution.
4.4. Practical implications for accounting professionals
In a practical accounting sense, the findings present crucial guidelines to the design of automated financial statement analysis and risk management systems. The findings are a stark warning to practitioners and fintech developers interested in credit scoring or corporate distress prediction against the blind piling of alternative data. Accounting systems must adopt a separable processing strategy rather than take a naive fusion strategy, which distorts the essential financial signals. This involves using structured financial ratios as the main quantitative driver of predictive modelling, with textual sentiment analysis as a separate, qualitative overlay, like an automated early warning signal or a second audit filter. Decoupling these streams of data, risk models can utilise all the contextual information of textual disclosures without undermining the actuarial accuracy of structured financial data.
5. CONCLUSION
In this paper, we experimentally studied the effectiveness of hybrid learning that integrates structured financial ratios with dictionary-based textual features in predicting stock performance. The main contribution is to show the “methodological boundary of naive fusion.” The results indicate that in the case where a strong non-linear model such as XGBoost can already provide High-Performance baseline on Structured data, simple concatenation most low dimensional text features do not bring statistically significant improvements. This effect is due to a “Signal Dilution Effect” in which features with low information density may weaken the quality of the combined feature space.
To obtain these results, our study has faced the limitation of using a single, relatively simple dictionary-based approach for text analysis and a particular sample of firms and financial ratios. Future research may test if these results generalise to other forms of textual representation or in different market settings.
In sum, this work calls for a large-scale refocusing of multi-modal financial forecasts. The challenge is now not to amass yet more data, but to devise sophisticated methods to connect the “information representation gap” across modalities. Future directions could include exploring more advanced text representations, employing fine-grained information extraction, and/or utilising intelligent fusion architectures (e.g., attention-based) to fully realise the potential synergy of integrating quantitative and qualitative corporate disclosures.
Ethical approval
Institutional Review Board approval was not required as this study did not involve human participants, patients, or animals.
Declaration of patient consent
Patient’s consent not required as there are no patients in this study.
Financial support and sponsorship
Nil.
Conflicts of interest
There are no conflicts of interest.
Use of artificial intelligence (AI)-assisted technology for manuscript preparation
The authors confirm that there was no use of artificial intelligence (AI)-assisted technology for assisting in the writing or editing of the manuscript and no images were manipulated using AI.
References
- Machine learning and AI for risk management. In: Palgrave studies in digital business & enabling technologies, disrupting finance Palgrave studies in digital business & enabling technologies, disrupting finance. Springer international publishing; 2018. p. :33-50.
- [Google Scholar]
- Artificial intelligence in finance. The Alan Turing Institute; 2019.
- A systematic review of machine learning and deep learning techniques for stock price prediction. J Risk Financ Manag. 2023;16:232.
- [Google Scholar]
- Textual analysis of corporate disclosures: A survey of the literature. J Account Lit. 2010;29:143-65.
- [Google Scholar]
- Financial ratios, discriminant analysis and the prediction of corporate bankruptcy. J Finance. 1968;23:589-609.
- [CrossRef] [Google Scholar]
- Financial ratios and the probabilistic prediction of bankruptcy. J Account Res. 1980;18:109.
- [CrossRef] [Google Scholar]
- A bibliometric analysis of the financial distress prediction literature. International J Bus Manag. 2009;4:116.
- [Google Scholar]
- Using neural network ensembles for bankruptcy prediction and credit scoring. Expert Syst Appl. 2008;34:2639-4.
- [CrossRef] [Google Scholar]
- Comparative analysis of data mining methods for bankruptcy prediction. Decis Support Syst. 2012;52:464-73.
- [CrossRef] [Google Scholar]
- Textual analysis in accounting and finance: A survey. J of Account Res. 2016;54:1187-230.
- [CrossRef] [Google Scholar]
- When is a liability not a liability? Textual analysis, dictionaries, and 10-Ks. J Finance. 2011;66:35-65.
- [Google Scholar]
- Measuring readability in financial disclosures. J Finance. 2014;69:1643-71.
- [CrossRef] [Google Scholar]
- FinBERT: Financial sentiment analysis with pre-trained language models. arXiv Preprint arXiv 2019:1908.10063.
- [Google Scholar]
- Multimodal machine learning: A survey and taxonomy. IEEE Trans Pattern Anal Mach Intell. 2019;41:423-4.
- [CrossRef] [PubMed] [Google Scholar]
- Multimodal deep learning for finance: Integrating and forecasting international stock markets. J Supercomput. 2020;76:8294-312.
- [CrossRef] [Google Scholar]
- The brave new world of digital finance: A review and research agenda. Financial markets, Inst Instrum. 2022;31:211-237.
- [Google Scholar]
- Consumers and artificial intelligence: An experiential perspective. J Market. 2021;85:131-5.
- [Google Scholar]