Financial AI: From Fraud Detection and Credit Scoring to Financial LLMs

Why it was needed: finance has always been about judging from data

Banks and card companies make countless decisions every day: whether to approve this payment, whether to lend money to this person, whether this transfer is money laundering. In the past these were filtered with rules written by people (check if the amount is above a certain level, reject if there is a history of late payments), but rules struggle to keep up with new schemes and conflict with each other as they multiply. That is why machine learning, which learns patterns from past data, entered finance early.

The XGBoost paper paper also gives the example that 'fraud detection systems protect banks from malicious attackers' in its introduction when explaining why machine learning matters. In a 2024 survey source by the Bank of England (BoE) and the Financial Conduct Authority (FCA), financial firms named data and analytics insights, anti-money laundering (AML) and fraud prevention, and cybersecurity as the largest current benefits of AI.

Recently, handling text, not just tables of numbers, has also become a job for AI. BloombergGPT paper lists sentiment analysis, named entity recognition (finding company and person names in sentences), news classification and question answering as natural language processing tasks in finance, and explains that financial terminology and context are complex enough to require systems tailored to finance. So financial AI is easiest to understand along two axes: 'tabular data models' and 'financial LLMs'.

Core idea: many trees for tabular data

The data for fraud detection and credit scoring is usually structured data. Each row is one transaction or one customer, and the columns hold values such as amount, time, merchant, income and number of late payments. A long-standing method for this kind of data is gradient boosting, which stacks many decision trees.

An analogy is a 'notebook of wrong answers'. If the first tree makes a rough prediction, the second tree learns to correct what the first got wrong, and the third reduces the remaining error further. As the XGBoost paper paper describes, an example follows each tree's rules down to a leaf, and the sum of the scores written on those leaves is the final prediction. To keep trees from becoming too complex, the objective function also includes regularization that penalizes the number of leaves and the size of leaf scores.

The paper shows how widely this method was used through competition results. Of the 29 winning solutions published on the Kaggle blog in 2015, 17 used XGBoost, while deep neural networks, the second most used, appeared in 11. At KDDCup 2015, the paper states, every one of the top 10 teams used XGBoost. However, the paper also adds that domain analysis and feature engineering played an important role in these problems.

How it works: the fraud detection and credit scoring pipeline

In practice, the pipeline looks roughly like this. First, collect raw data such as transaction records or loan applications. Second, use feature engineering to create columns for decision-making, such as 'number of payments in the last hour' or 'distance from the usual payment location'. Third, label transactions confirmed as fraud in the past or loans that actually went delinquent as the correct answers (labels) and run supervised learning. Fourth, the trained model produces a risk score for each new transaction or application. Fifth, if the score exceeds the threshold, the case is blocked, put on hold or passed to human review.

Financial data has its own difficulties. Fraud is very rare among all transactions, so class imbalance is severe, and there are many missing values. XGBoost paper proposed a sparsity-aware algorithm for sparse data with many missing values, and a structure that reads data too large for memory from disk in chunks. The paper states that it runs more than 10 times faster than existing popular solutions on a single machine and scales to billions of examples in distributed settings.

The key point is that the model produces a 'score', not a 'decision'. At what score to block, who makes the final call, and what to explain to a rejected customer are operational design choices that people and organizations must make.

Development: two branches of finance-specific LLMs

If tabular data models handle numbers, financial LLMs handle text such as disclosures, news and reports. The first major example is BloombergGPT paper. In building a 50-billion-parameter model, Bloomberg created a 363-billion-token financial dataset from financial documents collected and curated over nearly 40 years, and mixed in 345 billion tokens of general data. The key is this 'mixed training', which neither used only financial data nor used a general model as is. The authors reported that this approach significantly outperformed existing models on financial tasks while maintaining performance on general benchmarks. Their careful tracking of data sources and usage rights is also worth noting for the financial sector.

FinGPT paper started from the opposite side. This work argues that proprietary models like BloombergGPT rely on exceptional data accumulation, and that an open-source alternative anyone can use is needed. So instead of training a huge model from scratch, it took a data-centric approach, emphasizing a pipeline that automatically collects and cleans financial data from the internet, and lightweight LoRA (low-rank adaptation) fine-tuning. Example applications include robo-advisors, algorithmic trading and low-code development.

The two branches do not so much replace each other as handle different data. Tabular data models remain at the center of decisions such as loan approval, while LLMs are expanding their use in reading and summarizing documents and answering questions.

Where things are heading

First, research on adding external knowledge to credit scoring models. Launching a new loan product creates a 'cold start' problem with almost no delinquency data. Ranking Prior Alignment paper is a method that distills ranking information assigned by experts, teacher models or LLMs into the scoring model. It can be attached in the same way not only to neural networks but also to XGBoost's custom objective function, and no external model is needed at prediction time. It was tested on industrial data from more than 1.5 million merchants, and the authors report that AUC increased on all 5 folds of the public Amex data (average +0.041).

Second, research on verifying the 'evidence' of financial LLMs. FinRegQA-EU paper built a regulatory question-answering dataset from the official Q&As of the European Banking Authority (EBA) and the European Securities and Markets Authority (ESMA). The model trained with SFT (supervised fine-tuning) received the highest LLM-judge scores but the lowest citation accuracy (Citation F1). It added three times as many citations, most of them unsupported. DPO and GRPO showed a trade-off that was more concise with higher citation accuracy, and the researchers pointed out that LLM-as-a-Judge assumes that many citations mean the answer is grounded and scores it accordingly, so it fails to catch fabricated citations.

Third, fairness. The missing-financial-information study paper fed 96,600 prompts into Llama-3.1-8B-Instruct and measured how much the recommended stock allocation changed when the financial conditions stayed the same and only the investor's identity was changed. The average gap between two identities grew from 4.78 percentage points when all financial information was given to 10.34 percentage points when none was given. Without the information, the model cited income, debt and savings it had never been told about, and even with full information, small gender-based differences remained.

Regulation and accountability: credit scoring is high-risk AI

Annex III of the EU AI Act source, under the high-risk area of 'access to essential private and public services', explicitly lists AI that evaluates individuals' creditworthiness or establishes their credit score as high-risk. However, AI used to detect financial fraud is excluded from this item. AI that assesses individual risk and sets premiums in life and health insurance is also in the same item. In effect, the same gradient boosting model has a different regulatory status depending on whether it is used for fraud detection or credit scoring.

The UK BoE and FCA survey source(118 firms responding) shows that 75% of respondents were already using AI (58% in 2022). 55% of use cases involved some degree of automated decision-making, but only 2% were fully autonomous decisions. Meanwhile, 46% of firms said they had 'only partial understanding' of the AI technologies they use, mainly because of third-party models. 84% had a person accountable for their AI framework. Four of the top five current risks named were data-related: data privacy, data quality, data security, and data bias and representativeness.

The Financial Stability Board (FSB) report source looks beyond individual firms at the financial system as a whole. It identified AI-related vulnerabilities that could increase systemic risk: third-party dependencies and service provider concentration, market correlations (the same models moving in the same direction), cyber risk, and model risk, data quality and governance. It also noted that generative AI raises the potential for financial fraud and market disinformation, and asked authorities to strengthen monitoring of AI developments and review their regulatory frameworks.

All of this points to the same conclusion. Loan rejections and account freezes directly affect people's lives. The model is a tool that supports decisions, and responsibility for the decision and its consequences lies with the financial firm and the people in it.

Limitations and issues: explanation, fairness, model risk

Explainability: a single decision tree's rules can be read directly, but an ensemble summing the scores of hundreds of trees does not show at a glance 'why this person was rejected'. LLMs are even harder. As FinRegQA-EU paper shows, there is no guarantee that a plausible citation is real evidence, and in the missing-financial-information study paper, the model gave reasons based on facts it had never been told. The explanations a model produces must themselves be verified.

Fairness: models learn the biases accumulated in historical data as they are. The BoE and FCA survey source also named data bias and representativeness among the top risks. The fact that advice changes when only identity changes paper is why LLMs used for financial advice should be designed to answer only after receiving sufficient information.

Model risk management: when fraud schemes and economic conditions change, model performance drops due to data drift. Relying on third-party models lowers understanding source, and when many firms concentrate on the same provider, systemic risk grows source. Keep in mind as well that all performance figures in this article come from research settings and do not guarantee the same performance in real operations.

Further study

To start with tabular data models, Section 2 of XGBoost paper, 'Tree Boosting in a NutShell', is a good choice. It briefly lays out the structure of summing the scores of many trees and the regularized objective function. Reading Ranking Prior Alignment paper next shows a recent approach that adds external knowledge to the same XGBoost to handle cold-start credit scoring.

For financial LLMs, reading the introduction and data composition sections of BloombergGPT paper together with FinGPT paper, which represents the open-source view, makes the difference between the two branches clear. For evaluation and risk, FinRegQA-EU paper and the missing-financial-information study paper are good.

For regulation and the state of the industry, reading item 5 of Annex III of the EU AI Act source, the summary of the BoE and FCA 2024 survey source and the summary of the FSB report source, in that order, widens the view from individual models → firms → the financial system.