Generative AI - Fact Or Fiction!

Generative AI solutions are a potential panacea for the banking and capital market data-analytics driven innovation — let’s run a fact check!

Banks, financial institutions, and consulting firms need to understand Generative AI is not all about ChatGPT or GPT4 (GPT is just one of the LLM methods) — this article outlines the art of possible and its various method, forms, data requirements, and use cases!

#If your data isn’t ready for Generative AI, your business isn’t ready for Generative AI

Banks should first develop a comprehensive data strategy based on the business use cases they plan to explore Generative AI solutions for — the data strategy should include structured, semi-structured, and unstructured data spanning banks internal data, industry third-party data, public records, news feeds, market & economics data, streets & customer’s earning calls data, SEC filings, market research, etc. Additionally, banks should plan for redesigning and preparing their data architecture and intelligent data platform for LLM data (e.g., chunked text and vectors) management, speed of access and computing.

The banking industry is under constant pressure to meet the ever-changing needs of customers while also dealing with strict regulations. This has made it difficult for banks to keep up with the pace of change and adopt new technologies. However, this is starting to change as more and more banks are turning to artificial intelligence (AI) solutions to improve their operations.

One of the most promising AI solutions for the banking industry is the Generative Pre-trained Transformer (GPT). This is a type of AI that is very effective in tasks such as text generation and machine translation. The GPT can be used to generate reports, customer letters, and even marketing materials. The GPT can help banks save time and money, while also providing better customer service. In addition, can help banks comply with regulations and reduce risks.

According to Gartner’s predictions: By 2025, generative AI will be producing 10% of all data (now it’s less than 1%) with 20% of all test data for consumer-facing use cases.

• Generative AI Models Overview
• Application of Various Generative AI Techniques in Banking Industry
• What is the Best Way to Evaluate Generative AI Models?
• Types of Generative AI Models Diffusion Model/Foundation Model Generative Adversarial Networks Transformer-Based Model Variational Auto Encoders (VAEs)
• What Kinds of Output Can a Generative AI Model Produce?
• What Kinds of Problems Can a Generative AI Model Solve?

According to Gartner’s predictions: By 2025, generative AI will be producing 10% of all data (now it’s less than 1%) with 20% of all test data for consumer-facing use cases.

Generative AI Models Overview

Generative artificial intelligence (AI) models are a combination of various AI algorithms used to represent and process content. These models utilize?natural language processing techniques to generate text, transforming raw characters such as large documents corpus (e.g., loan documents and unstructured borrower’s information), words into sentences, entities , and actions, which are then represented as vectors using multiple encoding techniques. Images are also transformed into various visual elements, also expressed as vectors.

However, these techniques can also encode biases, racism, deception, and puffery that exist in the training data. Developers must be cautious of these limitations while representing data.

Once developers select the representation of data, they apply a specific neural network to generate new content in response to a query or prompt. Techniques such as?Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs)?are appropriate for generating realistic human faces, synthetic data for AI training, or facsimiles of particular humans.

Transformers such as Google’s Bidirectional Encoder Representations from Transformers (BERT), OpenAI’s Chat GPT, and Google AlphaFold have advanced progress in generating new content. In addition to encoding language, images, and proteins, these neural networks can also generate new content. Generative Pre-trained Transformers, commonly known as GPT, are?a family of neural network models that uses the transformer architecture?and is a key advancement in artificial intelligence (AI) powering generative AI applications such as ChatGPT.

Application of Various Generative AI Techniques in Banking Industry

Generative Pre-Training Transformer can benefit the banking industry by providing a powerful tool to process and analyze large volumes of banking sector data quickly and accurately. improved customer service and faster processing times

GPT is used to improve customer experience by providing customers with more detailed information about their accounts and transactions.

Automating boring administrative tasks, such as document processing and document summarization with GPT will increases accuracy and reduces potential human errors by automating

Furthermore, GPT can be used to create a more personalized experience for customers by offering tailored advice, product recommendations, and other services.

GPT can also help banks to better manage risk by providing more accurate and timely data. This data can be used to identify potential fraud and other risks, allowing banks to take proactive steps to protect their customers and their assets.

Additionally, GPT can be used to improve customer segmentation, allowing banks to better target their services and products to the right customers.

According to the KPMG data, 76% of financial services executives said they see fraud detection as a major application. More than two-thirds of respondents said compliance and risk will be a top use. For instance, banks could use generative AI to automate regulatory filings and analyze historical data to simulate risk scenarios. Generative AI will also likely be used to power more sophisticated consumer-facing chatbots, per 66% of respondents.

With Large Language Models (LLM) generative AI techniques such as GPT, and other Gen AI methods such as GANs and VAEs, Corporate Investment Banks and financial Institutions can now capture and process large varieties of unstructured data (e.g., public records, news feeds, DOJ & Watchdog agencies, LexisNexis & other third-party data, social media, capital market & earning call data, SEC filings, and others) to solve a large number of use cases ranging from CRM Insights/NBO, Intelligent Loan Intake & document processing, credit research & due-diligence/financial spreading & Pro-Forma, CDD/EDD/AML/KYC scoring, fraud detection & prevention, cybersecurity & data privacy, to loan review for policy and regulatory compliance, and others. We will discuss the application of each individual Generative AI technique in the following sections of this document. ?????

KPMG's June data showed that 42% of financial service executives expected to increase generative AI investments by 50% to 99%, and 41% of respondents expected more than 100% increase.?

Most financial institutions are still in the exploration or ideation phases, researching which practices would be most valuable to them. Major banks like JPMorgan Chase, Wells Fargo and Goldman Sachs, which have been experimenting with the innovative technology for years, are rolling out a number of AI-powered programs.

JPMorgan Chase is beginning to use?large language models to detect fraud, by examining patterns in emails for signs of compromise, among other uses. Goldman Sachs is using generative AI to?assist software engineers?in code development.

Ally Bank has?piloted an AI-powered program?that transcribes and summarizes customer service calls, a job previously done manually by contact center representatives. SouthState Bank in Winter Haven, Florida?trained an enterprise-version of ChatGPT?that allows employees to use it to query on bank policies, draft emails and summarize meetings.

GPT-Based Generative AI in Banking -- Few General Use Cases

• Detect and Prevent Fraud
• Manage Risk and Improve Credit Scoring (e.g., Provide credit ratings, Borrower’s ability-to-repay analysis for wise loan decisions, Early-warning, and Lower default risk)
• Credit Research and Due-Diligence, Financial Spreading, and Pro-Forma Financial Statements
• Personalize Marketing Efforts
• Boost the Customer Experience
• Generate Financial Advice
• Intelligent Loan Intake and Documents Processing

Examining the potential benefits of the Generative Pre-Training Transformer for the banking industry

GPTs could be used to automate tedious customer service tasks, such as responding to customer inquiries. By automating certain processes, banks can reduce the need for manual labor and free up resources to focus on other areas of the business. It could be used to generate highly targeted ads and promotions that would optimize customer engagement and loyalty.

GPTs could be used to automatically generate financial documents, such as loan contracts, statements, and agreements. This would save time and costs associated with manual document creation. In addition, it could be used to automate financial analysis tasks and generate insightful, accurate reports.

Banks can use GPT to Automate customer service interactions; Enhance fraud detection (using GAN model to generate real-life like synthetic fraud data); Streamline loan document processing and automate loan review for risk & regulatory compliance.

What is the Best Way to Evaluate Generative AI Models?

Here are the following things you should take into account while evaluating generative AI models:

• Generative AI Method Selection: Banks should delve into business use cases to analyze whether they demand more data mining or data augmentation and simulation to select the right Generative AI method. For example, GPT-based Generative AI is a good candidate for data mining as the architecture of GPT consists of a stack of identical encoder-decoder Transformer layers. The encoder processes the input text, while the decoder generates the output -- GPT is solely focused on the decoding process, making it particularly effective for mining large corpus. In comparison, the GANs model is the best candidate for data augmentation and simulation. GANs can take real-sample data as input and generate synthetic data close to real-samples with high accuracy.
• Data Strategy and Collection for Generative AI: If your data isn’t ready for Generative AI, your business isn’t ready for Generative AI. Banks should first develop a comprehensive data strategy based on the business use cases they plan to explore Generative AI solutions for — the data strategy should include structured, semi-structured, and unstructured data spanning banks internal data, industry third-party data, public records, news feeds, market & economics data, streets & customer’s earning calls data, SEC filings, market research, etc. Additionally, banks should plan for redesigning and preparing their data architecture and intelligent data platform for LLM data (e.g., chunked text and vectors) management, speed of access and computing.
• Quality: When it comes to applications that interact with users, the quality of the generated output is of utmost importance. In speech generation, for instance, if the quality of the speech is poor, it becomes challenging to understand. Similarly, in image generation, the desired outputs should be visually indistinguishable from natural images to be considered of high quality.
• Diversity: This is another crucial factor to consider. A good generative model should be able to capture the minority modes in its data distribution without sacrificing generation quality. This helps reduce any undesired biases in the learned models.
• Speed: Many interactive applications require fast generation, particularly real-time image editing, to allow for use in content creation workflows. As such, speed is a significant consideration when evaluating generative AI tools.

There are various types of generative AI models, and the combination of their unique features can lead to even more powerful models. Let’s dive deeper into these models:

1. Generative Adversarial Networks
2. GPT or Transformer-Based Model
3. Variational AutoEncoders (VAEs)
4. Diffusion Model/Foundation Model

Application of Diffusion Model/Foundation Model:

Diffusion models, also known as denoising diffusion probabilistic models (DDPMs), are a type of generative model used in artificial intelligence and machine learning. Applications include:

• Decision-Making
• Reaction Time Analysis
• Perceptual Decision Making
• Financial Markets
• Medical Diagnosis

Applications of Generative Adversarial Networks or GANs:

Generative adversarial networks, or GANs for short, are a type of machine learning model that uses deep learning techniques to generate new data based on patterns learned from existing data. GAN is a framework which trains two sub-models, a generator and a discriminator, like a zero-sum game to produce increasingly accurate examples of the target data.?In the adversarial process, the generator can be seen as a simulator to generate the similar data as the real data, while the discriminator plays the role of a filter to distinguish the real data and generated data. They can reach an ideal point that the discriminator is unable to differentiate the two types of data. At this point, the generator can capture the data distributions from this game. See GAN architecture below for the prediction of stock closing price.

Some of the Applications of GAN-based Generative AI solutions for the Banking & Capital Market Include:

• Banking & Capital Market (e.g., Multi-time period asset return with excess Alpha) and prediction model of stock market trading actions using generative adversarial network (GAN) and piecewise linear (PLR) representation approaches. GAN-based framework which combining PLR approach, LSTM, and three GAN models to generate trading actions with high performance. The overall process is illustrated below.

Build your Alpha Universe collecting and processing wide array of data sets:

1. Structured: internal portfolio positions & transactions, security master, price master, reference data, and real-time market data feeds;
2. Vendor data: Bloomberg, FactSet, Refinitiv, S&P, Moody’s, ICE data services, Morningstar, MSCI, and Dow Jones Factiva;
3. Unstructured: SEC filings, investors presentation, street transcripts, market events, analyst feedback, broker research, news feeds, etc.
4. Factor universe (e.g., Fundamental, Momentum, Statistical, ?Macroeconomics, Cyclical, Sector, Geopolitical, and ESG, etc.):
5. Fundamental and statistical factors: Balance sheet, cash flows, income statements, financial rations, R-Squared, correlation, implied volatility, etc.;
6. Style factors: Value, Size, Growth, Low Volatility, Quality, Yield, Momentum, and others;
7. ESG factors for sustainable finance: (E) Carbon Emissions, GHG Emissions, Renewable Energy, Toxic Air Emissions, Water Efficiency, Climate Change, etc.; (S) Benefits, Diversity and Inclusion, Employee Training, Human Capital, Operational Performance, Product Quality and Safety; (G) Board Profile, Board Skillset, Business Ethics, Compensation, Ownership and Control, and Sustainability.

***AWM quants can use Generative Pre-Trained Transformer i.e., GPT-based LLM data miner to mine vectors from the unstructured data to incorporate more factors to the equity portfolio analysis and optimization.

• Generative AI solutions for Asset & Wealth Management (AWM) -- Active Portfolio Management: AI/ML methods and Generative Modeling such as generative adversarial network (GAN) using as Generator & Discriminator, Long short-term memory (LSTM) recurrent neural network (RNN), multi-layer perceptron (MLP), stochastic discount factor (SDF -- pricing kernel), and simulation engine etc. can significantly reduce noise on time-series asset returns and generate much 95% to 99% accurate future returns compare to traditional ARIMA or GARCH models. Asset Managers and Investors that target for excess Alpha outperforming benchmark indices can now use AI/ML new methods for active asset management. With high performance GPU led cloud computing, FIs can process a wide array of data to engineer large number of features & apply AI/ML methods to glean patterns in trade signals for excess alpha.

Portfolio analysis with generative adversarial networks (PA-GAN). PA-GAN directly models the market uncertainty, the main factor driving future price trend, in its complex multidimensional form, such that the non-linear interactions between different assets can be embedded. Consider Using an optimization methodology that utilizes the probability distribution of real market trends learnt from training PA-GAN to determine the best portfolio diversification minimizing the risk and maximizing the expected returns observed over the execution of multiple simulations.

Data preprocessing: The process of generating training samples involves three steps, namely feature generation, trading sequence generation, and training sample generation.

Step 1: Feature generation as input: Create input features for model training including daily opening price, closing price, highest price, lowest price, trading volume, and technical indices.

Step 2: Employ the piecewise linear representation (PLR) approach to generate real output targets such as trading actions by identifying trading points (trading timing) based on the adjusted closing price. Subsequently, transform all trading points to trading actions, such as buying, selling or holding.

Step 3: Merge the features and trading actions to obtain sample data by date.

The generator of the GAN model is designed by LSTM with its strong ability in processing time series data. Let’s choose the daily data in the last 20 years with 7 financial factors to predict the future closing price. The 7 factors of the stock data in one day are High Price, Low Price, Open Price, Close Price, Volume, Turnover Rate and Ma5 (the average of closing price in past 5 days). The 7 factors are valuable and significant in price prediction with the theory of technical analysis, Mean Reversion, or MAR. Therefore, these factors can be used as 7 features of the stock data for the price prediction.

Data collection:

1. Structured: Portfolio positions, active trade, mark-to-market, notional, reference data, etc. Vendor: FactSet, Bloomberg, Refinitiv, S&P, Moody's, IEC, NYSE, etc.
2. Unstructured***: SEC filings, investors presentation, street transcripts, market events, analyst feedback, broker research, news feeds, etc.
3. Alternative****: Management change, M&A, lawsuits, public records, bankruptcy, geopolitical (e.g., Brexit, trade dispute, supply chain disruption), cyber crisis, social media, etc.
4. Fundamental and statistical factors: Balance sheet, cash flows, income statements, financial rations, R-Square, correlation, volatility, etc. Style factors: Value, Size, Growth, Low Volatility, Quality, Yield, Momentum, others
5. ESG factors***: Environmental: Carbon Emissions, GHG Emissions, Renewable Energy, Toxic Air Emissions, Water Efficiency, Waste, Climate Change, etc.; Social: Benefits, Diversity and Inclusion, Employee Training, Human Capital, Operational Performance, Product Quality and Safety; Governance: Board Profile, Board Skillset, Business Ethics, Compensation, Ownership and Control, Sustainability, etc.

***AWM quants can use Generative Pre-Trained Transformer i.e., GPT-based LLM data miner to mine vectors from the unstructured data to incorporate more factors to the equity portfolio analysis and optimization.

See the AWM Generative AI solution framework below for asset return, allocation, and portfolio optimization method:

Outcome of Active Portfolio Management for Asset & Wealth (AWM) Managers Using GAN-Based Generative AI Method as Depicted in the Diagram Above: Explain asset prices for different assets; Design optimal risk-adjusted portfolios; Find mis-priced assets to earn alpha; Use all available information in the market and understand which information is relevant.

• Enhanced FinCrime and Fraud detection

Application of GPT -- Transformer-Based Model:

• Language Translation
• Sentiment Analysis
• Chatbot and Virtual Assistant
• Questioning Answering
• Content Creation
• Detect Trends and Anomalies

Transformer-based models are neural networks that excel at learning context and meaning by closely analyzing relationships in sequential data. For example, technologies such as generative Pre-Trained (GPT) language models can use Banks' internal portfolio data, external industry, news feeds, capital market data including structured, semi-structured, and unstructured data to gain customer insights and suggest next-best-offer (NBOs), model customer segmentation and sentiment analysis, intelligent document intake and processing, automate commercial credit research and due-diligence/financial spreading and develop credit package contents, ALM/KYC scoring, Customer Onboarding CDD/EDD and PEP screening, etc. Applications of GPT include:

Applications of Variational AutoEncoders (VAEs):

Variational AutoEncoders (VAEs): Variational AutoEncoders (VAEs) are a type of generative model, similar to Generative Adversarial Networks (GANs). VAEs are incredibly useful for creating complex generative models of data, especially when working with large datasets. Applications of VAEs include:

• Image and Video Generation
• Anomaly Detection -- VAEs anomaly detection can inadvertently expose sensitive information if anomalous data is privacy-sensitive.
• Data Imputation
• Dimensionality Reduction
• Signal Processing
• Security Analytics (Cybersecurity etc.)

Generative AI for Data Privacy, Security and Governance: Generative AI models have unique applications in data privacy, Cybersecurity, and governance.

While some of these applications focus on improving security measures, others involve potential risks related to data privacy. Let’s explore how each of the previously noted generative AI types is used in these areas:

• Generative Adversarial Networks (GANs): Security & Privacy Use: GANs can be used in security applications to generate realistic synthetic data for training robust models and testing security systems. For example, in cybersecurity, GANs can create realistic network traffic data to test the resilience of intrusion detection systems or to generate realistic malware samples for evaluating antivirus software.
• Variational Autoencoders (VAEs) Security & Privacy Use: VAEs have applications in anomaly detection and security -- Security & Privacy Use. They can learn the normal patterns in data and identify anomalies or potential security breaches. For example, VAEs can detect unusual network activity or fraudulent transactions.
• Transformer-based Models Security & Privacy Use: Transformer-based models, particularly large language models like GPT, can be used in security applications for natural language understanding and processing, helping detect and prevent potential security breaches in textual data.

Generative Pre-Training Transformers (GPTs) are a type of natural language processing (NLP) technology that uses deep learning to generate text. GPTs have become increasingly popular in the past few years, especially for applications in natural language understanding (NLU), such as language generation and comprehension. The banking industry has not yet adopted GPTs on a large scale, but there are many potential benefits the technology could provide.

GPT can provide a range of long-term benefits for the banking industry. The technology can allow for faster processing times and improved accuracy in transaction processing, which can lead to increased customer satisfaction and a stronger reputation for the bank. Additionally, by leveraging GPT for customer service, banks can create a more personalized experience for their customers and improve loyalty. Furthermore, GPT can help banks to gain insights into customer behavior and preferences, allowing them to better tailor their services to meet customer needs.

Fluid AI can assist you to develop your own version of GPT which can help your organization to automate manual tasks to become more efficient and cost-effective, while also improving customer satisfaction. Our dedicated team is ready to answer any questions you may have and help you get started with this Game changing technology.

Challenges with Generative AI:

• Generative AI hallucinations: Generative AI can produce hallucinated result , particularly GPT, if adequate training data is not available or the training data is not representative of context. While leading AI experts aren't entirely sure what causes hallucinations, there are several factors that are often cited as triggers. First, hallucinations can occur if the training data used to develop the model is insufficient or includes large gaps leading to edge cases that are unfamiliar to the model.
• Sample bias refers to the biases that are present in the training data used to build LLMs. Since these models learn from vast human-generated datasets, they tend to absorb the biases present in the text, perpetuating stereotypes and discriminations. Biases pertaining to race, gender, ethnicity and socioeconomic status can inadvertently be perpetuated by the AI system, leading to biased outputs. Other Biases are:
• Availability bias -- stems from the fact that LLM generative AI models are exposed to large amounts of publicly available data. As a result, the model is more likely to favor content that is more readily available while neglecting perspectives and information that are less prevalent online.
• Confirmation bias -- is a psychological tendency in which individuals seek information that confirms their existing beliefs while ignoring evidence that challenges them.
• Selection bias -- emerges when the training data is not representative of the entire population or target audience.
• Group attribution bias -- emerges when the generative AI attributes specific characteristics or behaviors to an entire group based on the actions of a few individuals.
• Contextual bias -- arises when the LLM model struggles to understand or interpret the context of a conversation or prompt accurately.
• Linguistic bias -- occurs when the LLM generative AI favors certain linguistic styles, vocabularies or cultural references over others.
• Anchoring bias -- occurs when an AI model relies too heavily on the initial information it receives.
• Automation bias -- refers to the tendency of humans to blindly trust AI-generated outputs without critically evaluating them.

Challenges and Limitations of Using Generative AI in Banks & Financial Services

Of course, working with Generative AI in the banking sector has its challenges and limitations. It’s not a magic bullet that can do everything. It’s just a tool.

For example, Generative AI should be used cautiously when dealing with sensitive customer data. It also shouldn’t be relied upon to stay compliant with different government regulations, such as the General Data Protection Regulation (GDPR) or the General Data Protection Regulation (CCPA).

Another limitation of Generative AI is that it can produce incorrect results if it’s fed with poor or incomplete data. So you always need to ensure your data is accurate and up to date. Otherwise, it could lead to poor financial decision-making.

As a rule of thumb, you should never let Generative AI have the final say in loan approvals and other important decisions that affect customers. Instead, have it do all the heavy lifting and then let financial professionals make the ultimate decisions.

All that said, Generative AI can still be a powerful banking tool if you know how to use it properly.

LLM generative AI offers transformative potential across industries, yet biases pose significant risks. Adhering to ethical AI development principles is paramount. Biases built into the models can affect the result, emphasizing the need for inclusive datasets, robust governance and vigilant evaluation.

Saroj Das, Banking and Capital Market Business & Technology Transformation Consultant

Saroj led IBM’s business & technology digital transformation practice for financial services sector and has held leadership roles with PwC, KPMG, and EY. He is a seasoned Banking & Capital Market business platform, operations, and risk & compliance practitioner. Over 25 years, he has been advising both US-based regional banks and global corporate investment banks on their business platform ecosystem and risk transformation programs spanning Lending and Trading books of business, Wealth & Asset Management, Custody & Trust, Finance & Treasury, Balance Sheet Management, Risk & Regulatory Compliance. Saroj works with C-Suits and line of business owners to identify high-stake use cases. He drives solutions for these use cases leveraging next-gen digital technologies such as cloud computing, Big-data & enterprise data fabric, advanced analytics e.g., Natural Language Processing (NLP)/AI/ML/deep learning data science, predictive analytics, Generative AI, APIs & Microservices, intelligent workflows & process automation; Vendor solutions evaluation & integration; Metaverse, Blockchain, and Web3 enablement.

Please contact Saroj at: Email: [email protected] ; Cell#: 646-285-3166

#GenerativeAI