3 Rise of Deep Generative Models
GenAI represents a significant advancement in AI technology, increasing its utility for financial institutions that have been quick to apply it to a wide range of applications. However, there are hazards inherent in AI technology and its implementation in the financial sector, such as embedded bias, privacy concerns, outcome opacity, performance robustness, unique cyber threats, and the possibility for new sources and transmission channels of systemic problemsYe and Yue (2023). GenAI could exacerbate some of these issues while also introducing new forms of vulnerabilities, particularly to financial sector stability. This study offers preliminary insights into the inherent dangers and solutions of GenAI and their potential influence on the financial sector. The adoption of generative AI in banking functions, like that of other technologies, will most likely follow an Exponential pattern. Finance teams are currently investigating how technology may supplement conventional procedures by writing text conducting research and creating preliminary draughts for jobs that are text-heavy or require little analysis, such as contract drafting and credit review supplementation. Accounting and financial reporting are two terms used interchangeablyHoude et al. (2020). Providing preliminary insights to assist in later versions of financial statements during month-end closes, or assisting in the construction of audit trails for reclassification memorandum. Budgeting and performance management are critical components of financial planning. Conducting ad hoc variance studies using the company’s structured or unstructured data sets, such as comparing actual to forecasts, and producing reports to explain the financial performance of various business units to stakeholders. The adoption of AI solutions in the banking sector will be accelerated by the market growth rate. Competitive forces have fuelled the financial sector’s rapid adoption of machine Learning in recent years by facilitating gains in efficiency and cost reductions, redefining client interfaces, improving forecasting accuracy, and improving risk management and complianceStrasser (2023).In the financial sector, generative models have proven useful for some useful tasks as these organizations can improve their ability to foresee market trends and simulate various economic situations by using generative models, allowing them to make more educated investment decisions. Some of the used models are discussed in this section Fig 1.
3.1 Variational Autoencoders (VAEs)
These are generative AI models frequently employed in the banking industry. VAEsKingma et al. (2019) are intended to understand the underlying structure of the input data and generate new samples that are similar to the original data distribution. Within the financial domain, these models operate through a process of dimensional reduction, whereby input financial data is mapped into a lower-dimensional latent space representation. This latent encoding adeptly captures the intrinsic attributes and latent patterns inherent to the dataset. The encoded data is subsequently subjected to a reverse transformation, enabling a reconstitution of the original data in its native space, thereby facilitating the restoration of the initial input data. The training entails optimizing two goals: reconstruction loss and Kullback-Leibler (KL) Bu et al. (2018)divergence. The reconstruction loss serves as a metric to assess the dissonance between the input and reconstructed data, thus incentivizing the model to furnish precise representations. Simultaneously, the KL divergence imposes constraints upon the latent space, coercing it to conform to a predefined distribution, typically a standard normal distribution, thereby imparting a regularization effect. This regularisation encourages the creation of varied and meaningful samples. VAEs are used in a Xu et al. (2019) variety of sectors in finance, including Portfolio optimization in which using historical market data, VAEs may learn the underlying structure and build new investment portfolios. Then in Anomaly detection VAEs are capable of detecting abnormal patterns in financial transactions or market behaviour. Risk assessment and modeling are another kind of field in which an Autoencoder can be used to simulate and assess hazards. Sophisticated CrossConvolutional Autoencoder can aid in the detection of fraudulent financial transactions. Variational Autoencoders (VAEs) exhibit the capability to synthesize financial data synthetically, thereby addressing limitations inherent in authentic real-world datasets. Notably, they find extensive application within options trading, where they aid in the creation of synthetic volatility surfaces, thereby augmenting the precision of option pricing and affording refined methodologies for trading and risk evaluation.
3.2 Generative Adversarial Networks (GANs)
In finance, generative adversarial networks (GANs)Goodfellow (2016) are utilized for tasks such as synthetic data generation, market simulation, and risk modeling improvement. These are generative AI models that have two components: a generator and a discriminator. They have acquired enormous attraction and adoption within the sphere of finance due to their ability to generate synthetic data and boost various financial procedures. The training procedure takes the form of an enthralling antagonistic dance between the generator and the discriminator. The generator’s goal is to deceive the discriminator by providing specimens that gradually resemble legitimate data, whereas the discriminator’s goal is to improve its ability to distinguish between the genuine and the fake. As training progresses, the generator learns to provide progressively very similar financial data, while the discriminator hones its expertise in the fine art of distinguishing between veracity and deception.
GAN applications in finance include that it can produce synthetic financial data, and addressing difficulties such as limited or biased datasets. Risk modeling, algorithmic trading, and portfolio optimization can all benefit from this information. Financial fraud detection: GANs can help distinguish between real and fraudulent transactions, improving financial fraud detection.GANs can generate artificial market data, assisting in understanding market dynamics, anticipating price changes, and evaluating the impact of various factors on financial markets. GANs can find anomalies in financial data by detecting unexpected patterns or outliersNgwenduna and Mbuvha (2021). CTAB-GANZhao et al. (2021), a conditional GAN-based tabular data generator, to produce synthetic data for credit card transactions, outperforms earlier methods. To detect fraud in imbalanced credit card transactions, a new model named Generative Adversarial Fusion Network (IGAFN)Lei et al. (2020) is used. Integration of diverse credit data was expertly choreographed, alleviating the problem of data asymmetry and exhibiting superior prowess when compared to existing credit scoring algorithms. These findings eloquently demonstrate the effectiveness of Generative Adversarial Networks (GANs) in uncovering the threat of credit card theft, with the tantalizing promise of ushering in advances in risk assessment within the financial arena.
3.3 Autoregressive Models
Temporal chronomancers, also known as time series models, rule supreme in finance, providing important instruments for diving into the complexities of analysis and prediction. These sophisticated models can decipher the delicate threads of temporal links and patterns within sequential data, whether it’s the cryptic undulations of stock prices, the rhythmic cadence of interest rates, or the ebbs and flows of economic indicators.
Autoregressive modelsBond-Taylor et al. () operate on the premise that the value of a variable at a given point in time is determined by its prior values. High dimensional models, such as autoregressive moving average (ARMA) and autoregressive integrated moving average (ARIMA), analyze the relationship between an observation and a lag set of observations. At its essence, the fundamental concept revolves around the notion that the state of a variable in a specific moment can be foretold by summoning forth a symphony of its past incarnations, garnished with a dash of stochastic serendipity. The term autoregressive elegantly alludes to this dynamic reliance on antecedent renditions of the variable. This model adroitly assigns significance to these temporal echoes, orchestrating their harmonious influence to orchestrate a compelling forecast of the present state Doering et al. (2019). In the case of ARMA models, the moving average element alludes to the model’s reliance on previous forecast errors or residuals. These models are often estimated using historical data to minimize the discrepancy between the observed and anticipated values.
The financial deployment of autoregressive models creates a landscape brimming with promise, where future financial variables are created from echoes of their past. These models summon their prophetic prowess in this mystical realm to foretell the dance of stock market prices, the flux of interest rates, the enigmatic waltz of currency exchange rates, and the harmonious rhythms of various financial indices, creating a symphony of insight into the financial future. These models help in risk assessment and portfolio optimization by modeling the volatility and correlations of asset returns. On the other hand, models assume stationarity, which means that the statistical featuresMohapatra et al. (2020) of the data remain constant throughout time. As a result, it is critical to evaluate the data’s stationarity and, if necessary, apply transformations or investigate more advanced models, such as ARIMA, which integrates differences to solve non-stationarity.
3.4 Transformers
A transformerVaswani et al. (2017) is a sort of neural network architecture that has gained popularity due to its capacity to handle sequential input more efficiently, such as text. Since transformer models can understand long-term dependencies and deal with the complex maze of sequential data, they have gained an important place in the financial domain. Their uses are numerous and include sentiment analysis, orchestration of document classification, and creation of financial text compositions. At the core of a transformer model is the attention mechanism, a vital component in the computational gear. This technique effectively allows the model the ability to generate representations while distributing different weights or significance throughout the input sequence’s structure. It turns into a tool for the model to focus its perceptive attention on pertinent sections, reliably capturing the subtleties of inter-element relationships. Transformer models, with their exceptional ability, untangle the intricate emotional fabric inherent in financial news, social media communications, and other textual transmissions. They expertly gather contextual cues and analyze word inter-dependencies, revealing illuminating insights into market emotion and providing investors with discernible tools for educated decision-making. Transformer models have a wide range of applications, including document classification. They take on the task of categorizing a variety of financial documents, research articles, and other textual expositions, producing a symphony of organization amid the textual cacophony. López-Ruiz et al. (2022). This aids in the organization and filtering of enormous amounts of financial data. Transformer models act as alchemists in the world of textual finance, conjuring up fictitious financial treatises, market exegesis, and relevant literary creations. Their art is a data-driven linguistic creative dance, informed by the intricate patterns and structures hidden inside the immense tapestry of financial data. This command of the language allows for the orchestration of automated report production and the seamless generation of content, ushering in a new era of textual creation built from algorithmic enchantments.
Models Restricted Boltzmann Machine (RBM) Variational Autoencoder (VAE) Autoregressive Models (LSTM , Transformers) Generative Adversarial Network (GAN) Abstraction Yes Yes No No Generation Yes Yes Yes Yes Probability Computation Intractible Intractible Tractible No Sampling Speed Markov Chain Monte Carlo Fast Slow Fast Types of Graphical Models Undirected Directed Directed Directed Loss Function KL divergence KL divergence KL divergence JS divergence Samples Very Bad OK Good Best
Table 1: A Comparison of Generative Models based on Architecture
Table 1 shows the power of different generative models and it is suggested that while creating models, these aspects should be taken into consideration to maintain the sanity of output products.