AI Finance & Financial LLMs 2026 — BloombergGPT, FinGPT, FinRobot, AlphaSense, Tegus, Daloopa, Bridgewater, Crunchbase AI Deep Dive

"Financial data is the richest domain data, but also the most expensive and the noisiest. Domain-specialized LLMs are no longer an option — they are infrastructure." — Gilberto Titericz, ex-Bloomberg AI

Finance is, as of 2026, the domain where LLM adaptation has advanced fastest alongside healthcare. When Bloomberg released its 50B-parameter BloombergGPT in March 2023, "general-purpose LLMs cannot handle financial vocabulary" became industry consensus. JPMorgan, Goldman Sachs, Morgan Stanley, BlackRock, Bridgewater, AlphaSense, Tegus, and Daloopa have all followed with their own models and platforms.

This article surveys the AI in finance ecosystem as of May 2026 — domain LLMs, research AI, hedge fund AI, asset manager AI, bank/insurance AI, and Korean/Japanese finance AI. Rather than a flat product catalog, the focus is "what data was it trained on, what workflow does it replace, what are its limits."

1. The 2026 Finance AI Map — Domain LLM, Research, Quant, Risk

Finance AI splits cleanly into five boxes based on role.

The five categories differ fundamentally in data source and user. Domain LLMs are infrastructure others build on. Research AI automates analyst information gathering. Quant focuses on alpha generation. Asset Manager / Bank targets operational efficiency and advisor productivity. Risk / Credit / Insurance handles automated decisioning.

Each category has different barriers to entry shaped by data access. Bloomberg and Refinitiv hold the core asset for Domain LLMs, while general LLMs from OpenAI, Anthropic, and Cohere complement them via RAG (Retrieval-Augmented Generation). That is the equilibrium in 2026.

2. BloombergGPT — The Prototype Financial Domain LLM

BloombergGPT is a 50-billion-parameter model the Bloomberg AI team released in March 2023, holding the title of "first large-scale LLM trained on financial domain data." Training used roughly 700 billion tokens — half (363B) being FinPile, Bloomberg's 40 years of accumulated proprietary financial data (news, filings, transcripts, market data), the other half public corpora (C4, Pile, Wikipedia).

The paper (arXiv 2303.17564) reported benchmarks where BloombergGPT substantially beat same-50B-class BLOOM-176B and OPT-66B on financial tasks (FPB, FiQA SA, Headline, NER), while not regressing on general tasks. The key insight: "50% domain data + 50% general data" is the golden ratio that maximizes domain adaptation without catastrophic forgetting.

BloombergGPT itself was never released, but features like AI Assist, Document Search, and Earnings Call Summarization integrated into Bloomberg Terminal are used daily by Terminal's 300,000+ subscribers. Bloomberg BQuant + AI, launched in 2024, integrates BloombergGPT into quant workflows — describe a backtest in natural language and Python code is generated.

BloombergGPT's biggest limitation is its closed nature. Outside academic research, no developer can access it; weights are not public. This vacuum spurred open-source counterparts: FinGPT, FinRobot, the InvestLM family.

3. FinGPT / FinRobot — Open-source Financial LLMs

FinGPT (github.com/AI4Finance-Foundation/FinGPT) is an open-source project started in June 2023 by Columbia University's AI4Finance Foundation. The thesis: "LoRA fine-tuning a base model like Llama on financial data is far more cost-efficient than pretraining from scratch like BloombergGPT."

FinGPT standardized a data pipeline like this.

# Example FinGPT data pipeline
from fingpt.data import (
    YahooFinance,
    SECFilings,
    NewsAPI,
    TwitterCrypto,
    AnalystReports,
)

# 1) Data collection — integrate diverse sources
pipeline = (
    YahooFinance(symbols=['AAPL', 'MSFT'], period='5y')
    + SECFilings(forms=['10-K', '10-Q', '8-K'], year_from=2020)
    + NewsAPI(topics=['earnings', 'M&A'], days=30)
    + AnalystReports(sources=['JPM', 'GS', 'MS'])
)

# 2) Sentiment / event labeling
labeled = pipeline.label(
    sentiment_model='FinBERT',
    event_extractor='FinGPT-Event-v3',
)

# 3) LoRA fine-tune on Llama-3-8B as base
from fingpt.train import LoRAFinetuner
trainer = LoRAFinetuner(
    base_model='meta-llama/Llama-3-8B-Instruct',
    rank=16,
    alpha=32,
    target_modules=['q_proj', 'v_proj'],
)
trainer.fit(labeled, epochs=3, lr=2e-4)

Built on FinGPT, FinRobot (github.com/AI4Finance-Foundation/FinRobot) is an agent framework released in 2024. Rather than a single model, it orchestrates role-based agents — Market Forecaster, Document Analyzer, Trading Strategist, Risk Analyst, Customer Service Agent — in a LangGraph-style flow. The 2025 FinRobot Industry Chain Analysis workflow can auto-decompose composite queries like "find and analyze Korean listed companies in NVIDIA's supply chain."

Similar open-source projects include InvestLM (NYU Stern, 13B/65B sizes), PIXIU (FinMA base, 100+ instruction-tuned financial tasks), FinMA (part of PIXIU, Llama-30B based), and FinTral (Mistral-based multimodal, handles charts and tables). All can be downloaded from Hugging Face and self-hosted.

4. DocLLM (JPMorgan) — Financial Document-specialized LLM

DocLLM is a document-understanding LLM that the JPMorgan AI Research team published in January 2024 (arXiv 2401.00908). The core observation: financial documents like 10-K, 10-Q, prospectuses, and term sheets carry meaning in layout (tables, position, fonts), not just text.

Unlike multimodal models like GPT-4V and LayoutLM that take pixel images as input, DocLLM trains on OCR-extracted text + bounding-box coordinates alone. Without an image encoder it is much lighter, yet it surpasses GPT-4V on table-heavy and form-heavy financial documents.

JPMorgan integrated DocLLM as a core component of its in-house LLM Suite, reportedly cutting an analyst's manual 4-hour prospectus reading down to 20 minutes (disclosed in their Q3 2024 earnings call). Weights are not released, but the paper is detailed enough that open-source reimplementations (unilm/docllm) are active.

Similar document-specialized models include FinMA-Document, AlphaSense Smart Synonym, and Hebbia Matrix — all share "cross-page, table-aware reasoning" as the central capability.

5. Bloomberg Terminal + BQuant + AI — Integrated Environment

Bloomberg Terminal, launched in 1981, has been the standard workstation for financial professionals for 45 years. As of 2025 it has roughly 350,000 users at about 27,000 USD per annual license, with revenue estimated near 13 billion USD.

AI Assist, integrated into Terminal in 2024, builds on BloombergGPT to automate the following.

• News Summary: summarize last-24-hour news on a ticker or sector, estimate market impact
• Earnings Call Q&A Extraction: structure analyst questions and management answers from call transcripts
• Document Search: handle natural-language queries like "Apple's gross margin trajectory over last 8 quarters", auto-pulling relevant filing pages
• Code Generation: natural-language to Python/BQL conversion inside BQuant

BQuant is Bloomberg's Jupyter-based quant workbench (launched 2018). The 2025 BQuant 3.0 release made LLM code generation a first-class feature — type "bar chart of S&P 500 sector 1-year returns" and code, chart, and commentary are produced.

# BQuant natural-language to code example
# Input: "bar chart of S&P 500 sector 1-year returns"
# Auto-generated code (BQL = Bloomberg Query Language):
import bql, bqplot

bq = bql.Service()
universe = bq.univ.members('SPX Index')
returns = bq.data.px_last(dates=bq.func.range('-1Y', '0D'))
sectors = bq.data.gics_sector_name()

req = bql.Request(universe, {
    'sector': sectors,
    'return': (returns.last() / returns.first() - 1) * 100,
})
df = bq.execute(req).single().df()
df.groupby('sector')['return'].mean().plot.bar()

Bloomberg's competitor Refinitiv Eikon (rebranded to Workspace under LSEG) released its own LLM Workspace AI in 2024, but the consensus view is that it still lags Bloomberg on data depth and integration. FactSet partnered with OpenAI in 2025 to ship GPT-4-backed Search/Q&A; S&P Capital IQ has its own ChatIQ; PitchBook released PitchBook AI for natural-language querying of PE/VC data.

6. AlphaSense — Enterprise Research Search

AlphaSense (alpha-sense.com), founded in 2008, is a financial research search platform that closed a 400-million-USD Series F in April 2024 at roughly 4 billion USD valuation. Nearly every Wall Street megabank and over half the Fortune 500 use it.

The core value proposition is solving the analyst pain point of "knowing where the information lives, but spending all your time looking for it." Data sources are as follows.

• Filings (SEC, global): 10-K, 10-Q, 8-K, 20-F, S-1, fully indexed
• Broker reports: JPM, GS, MS, BofA, Citi, and 1,000+ institutions (license-based)
• News: Reuters, Dow Jones, Bloomberg, 5,000+ sources
• Conference calls / transcripts: real-time transcription plus search
• Expert calls (post-Tegus acquisition): 50,000+ expert interviews

The 2023 launch of AlphaSense Smart Synonyms lets an LLM auto-map abbreviations like "EV/EBITDA" to "enterprise value to EBITDA" as the same concept. Generative Search, launched in 2024, generates summaries with citations for natural-language queries; every citation links to the source page.

Query: "How is AMD's data center revenue growth different from NVIDIA's, as of Q4 2024?"

Response:
  AMD's data center revenue reached 3.8B USD in Q4 2024, up 69 percent year-over-year [1].
  NVIDIA's data center revenue for the same quarter was 18.4B USD, up 93 percent [2].
  The main differences are (a) AI accelerator share — AMD MI300X near 5 percent, NVIDIA
  H100/H200 at 88 percent; (b) software stack — CUDA lock-in; and (c) supply chain —
  TSMC CoWoS capacity allocated preferentially to NVIDIA [3][4].

[1] AMD Q4 2024 10-Q, page 14
[2] NVIDIA Q4 FY25 10-Q, page 11
[3] Morgan Stanley "AI Semis", 2025-02-15
[4] JPM "AI Datacenter Outlook", 2025-03-10

AlphaSense pricing is not public, but enterprise seats reportedly run 10,000 to 20,000 USD per year. It is hard for individual analysts to access; the main customers are hedge funds, banks, and asset managers.

7. Tegus — Expert Call Primary Research

Tegus, founded in Chicago in 2017, is an expert-call platform that AlphaSense acquired in June 2024 for 930 million USD. The core asset is roughly 75,000 expert-call transcripts — hour-long 1:1 interviews with former executives, customers, suppliers, and competitors' executives.

Tegus's differentiator is its pricing model: hosting a typical 30-minute call costs 1,000 to 2,000 USD, but Tegus's flat platform fee — around 40,000 USD per year for a small fund — buys unlimited access to every call. This dramatically lowered hedge funds' primary-research costs.

Tegus AI Notes, launched in 2024, auto-summarizes transcripts and compares opinions from multiple experts on the same company. After the AlphaSense acquisition, in 2025 Tegus data was integrated into AlphaSense Generative Search, so queries like "AMD's data center market share" now cite SEC filings, analyst reports, and expert calls together.

Similar expert networks include GLG (Gerson Lehrman Group), Third Bridge, and Guidepoint, all of which have released their own AI search tools, but few open their call transcripts as completely as Tegus.

8. Daloopa — Financial Data Extraction Automation

Daloopa, founded in 2019 and Series B in 2022, automates financial data extraction. The thesis is simple — "analysts spend 40 percent of their time copying numbers from 10-Ks and 10-Qs into Excel."

Daloopa's workflow looks like this.

1. Daloopa auto-collects SEC filings (10-K, 10-Q, 8-K, IR decks) for covered companies
2. A proprietary OCR + LLM pipeline extracts tables, footnotes, and MD&A into structured data
3. Daloopa's in-house labeling team (about 200 people) human-verifies key data
4. Pushed automatically into Excel, Google Sheets, or Bloomberg Terminal API
5. Auto-updated whenever a new filing drops

As of 2024 Daloopa tracks quarterly KPIs for 8,000+ global listed companies, processing 300 to 500 new filings per week. The main customers are hedge funds and sell-side analysts, with seats running roughly 10,000 to 20,000 USD per year.

Competitors include Calcbench (SEC XBRL specialist, cheaper), Sentieo (acquired by AlphaSense), and Visible Alpha (consensus estimates specialist).

9. Hebbia — Document Q&A Platform

Hebbia (hebbia.com), founded in 2020, raised a 130-million-USD Series B in July 2024 at a 700-million-USD valuation. Its flagship Matrix is a multi-document Q&A platform that "answers questions across hundreds of pages of documents simultaneously."

The Matrix workflow is to fill a matrix where rows are documents and columns are questions. For instance, when a PE fund receives 100-page CIMs (Confidential Information Memoranda) for 50 potential acquisition targets, it can issue one command — "extract EBITDA, revenue growth, customer concentration, leverage ratio, union status" — across all of them.

Example Matrix output:
        | EBITDA      | Rev Growth | Customer Conc | Net Debt | Union
TargetA | 24M USD     | +18% YoY   | top 5 = 42%   | 1.2x     | No (cite p.34)
TargetB | 8M USD      | +47% YoY   | top 5 = 78%   | 0.4x     | Yes (cite p.21)
TargetC | 51M USD     | flat       | top 5 = 12%   | 2.8x     | No (cite p.55)
...

Every cell traces back to a source citation that analysts can verify. Hebbia's main customers are PE funds, hedge funds, and consulting firms (McKinsey, BCG, Bain), with enterprise pricing around 30,000 to 50,000 USD per seat per year.

Similar categories include Brevit, Casetext (acquired by Thomson Reuters, legal-focused), Eve (legal), and Harvey (adopted by law firms like Allen & Overy).

10. Hedge Fund AI — Bridgewater, Renaissance, Two Sigma

Bridgewater Associates launched its Artificial Intelligence Equity Fund (AIA Fund) in December 2024, led by David McCormick (former CEO candidate). The fund started at roughly 2 billion USD with the operating philosophy that "AI integrates data, news, and macro signals to automate parts of decision-making." Not a fully black-box quant fund — portfolio managers review and incorporate the AI's recommendations, a human-in-the-loop model.

Renaissance Technologies' Medallion Fund is the legendary quant fund that has averaged roughly 39 percent net returns annually since 1988. Operating after Jim Simons's passing, it is closed to outside investors and runs only employee capital. ML has been in use since the 1990s, and even in 2026 it is widely regarded as the most sophisticated quant system, though details remain private.

Two Sigma, founded in 2001, manages roughly 60 billion USD with more than 1,500 data scientists — an ML-first hedge fund. The 2024-launched Venn platform is a SaaS that lets external asset managers analyze portfolio risk, while internally Two Sigma runs an LLM-based news and signal pipeline.

D.E. Shaw, AQR, and Citadel are all quant + ML funds. Citadel announced in 2024 an enterprise agreement with OpenAI to use GPT-4o internally for R&D.

Numerai, founded in 2015, is a crowdsourced quant fund: it publishes anonymized market data, data scientists worldwide submit models, and the platform ensembles the results to run its own fund. The structure aligns incentives via a token (NMR), and as of 2026 about 5,000 modelers participate.

# Example Numerai tournament submission
import numerapi
napi = numerapi.NumerAPI()

# 1) Download data — new round each week
napi.download_dataset('v5.0/train.parquet')
napi.download_dataset('v5.0/live.parquet')

# 2) Train model (LightGBM, XGBoost, or your own)
import lightgbm as lgb
model = lgb.LGBMRegressor(n_estimators=2000, learning_rate=0.01)
model.fit(train[features], train['target'])

# 3) Submit predictions
preds = model.predict(live[features])
napi.upload_predictions('preds.csv', model_id='your-model-id')

# 4) Stake — lock NMR tokens to your model, get rewarded or burned by performance
napi.stake(amount=10, model_id='your-model-id')

WorldQuant, Susquehanna, and Jane Street all run LLM/ML infrastructure too, but external disclosure is minimal.

11. BlackRock Aladdin — Asset Management OS

BlackRock Aladdin (Asset, Liability, Debt and Derivative Investment Network) began as BlackRock's internal risk system in 1988. As of 2026 it monitors roughly 21 trillion USD in assets — effectively the operating system of the asset-management industry. Beyond BlackRock itself, about 200 external institutions (BNP Paribas, MUFG, HSBC AM, etc.) license it.

Aladdin's core modules are as follows.

• Portfolio Management: real-time state of the whole portfolio, rebalancing simulations
• Risk Analytics: VaR, stress testing, scenario analysis, factor exposure
• Trading: order routing, best execution, TCA (Transaction Cost Analysis)
• Compliance: per-fund guidelines, automated regulatory-limit checks
• Operations: settlement, accounting, NAV calculation

Aladdin Copilot, launched in 2024, uses GPT-4o plus a proprietary model so a portfolio manager can ask in natural language "list the 5 bonds most affected by the credit-spread widening over the past 30 days." In 2025 the LLM was integrated into eFront Copilot (PE/alternatives tool) as well.

Competitors include MSCI BarraOne, Bloomberg PORT/AIM, and FactSet PA. Aladdin is most comprehensive, BarraOne leads in risk models, PORT leads in fixed income.

Vanguard, Fidelity, and State Street all run their own AI platforms — State Street's Alpha Platform, Fidelity's Fidelity AI Lab, and Vanguard's Personal Advisor AI are leading examples.

12. Bank AI — JPMorgan, Goldman, Morgan Stanley, Wells Fargo

JPMorgan Chase is the most aggressive AI investor among US banks.

• COiN (Contract Intelligence): launched 2017, auto-analyzes credit contracts, saving 360,000 lawyer-hours per year
• IndexGPT: launched 2023, auto-generates ETF themes
• LLM Suite: launched June 2024 — about 160,000 of 280,000 employees have access to the internal LLM, which routes between GPT-4o, Claude, and proprietary models
• DocLLM: the document-specialized model discussed above

JPMorgan's differentiator is internal LLM Suite governance. Every LLM call is logged, PII/MNPI (material non-public information) filters apply automatically, and per-model risk assessments are refreshed every quarter.

Goldman Sachs launched its GS-AI Platform in June 2024 — a proprietary LLM gateway, standardized function calling, and tailored workflows per trader, banker, and research analyst. The Banker Copilot automates pitch-deck creation and peer analysis for IB analysts.

Morgan Stanley partnered with OpenAI in early 2023 to launch AskRes (AI Research Assistant), exposing Morgan Stanley's 100,000+ research documents to natural-language access. Debrief — automatic advisor-client meeting summaries — was added in 2024.

Wells Fargo's Fargo is a voice AI assistant launched in 2023; in the mobile app it answers natural-language queries like "total grocery spend last week." It logged about 1.2 billion cumulative calls through 2025.

HSBC and Standard Chartered were the fastest among global banks to adopt LLMs. HSBC operates an in-house HSBC AI Markets platform; SC runs an OpenAI-backed SC GPT.

13. Korean Finance AI — KB, Shinhan, Hana, Woori, Toss

KB Kookmin Bank's KB Liiv has reached 14 million cumulative users since its 2017 launch. Liiv AI Banker, added in 2024, supports transfers, inquiries, and loan simulations via natural language. KB Financial Group released its proprietary financial LLM KB-GPT internally in 2025.

Shinhan Bank's Sol Bank AI launched in 2024, and Shinhan AI Copilot assists branch staff with product explanations and customer-complaint handling. Shinhan Financial reportedly became the first Korean financial group to sign a direct enterprise agreement with Anthropic Claude in 2025.

Hana Bank's Hana 1Q mobile banking app launched in 2018, and Hana AI Wealth was added in 2025 to provide wealth-management recommendations.

Woori Bank's WON-PAY is an integrated payment and transfer platform; in 2025 WON AI Service was integrated to support voice-based transfers and consultations.

Toss Bank / Toss Securities / Toss Payments is Korea's most mobile-native finance platform, running GPT-4-based customer-support chat, Toss Securities' AI stock recommendations, and Toss Payments' automated dispute handling since 2024. The Toss data-science team reportedly released its Korean financial LLM Toss-FinLM internally in 2025.

The Korean Financial Services Commission published its Financial AI Guidelines in 2024, requiring explainability, non-discrimination, and supervisory reporting whenever AI is used in credit scoring, wealth management, or insurance underwriting. Alongside the EU AI Act, this is the key compliance frame for Korean finance AI adoption.

14. Japanese Finance AI — MUFG, Nomura, NTT Data

MUFG (Mitsubishi UFJ Financial Group) released its proprietary financial LLM MUFG-GPT internally in 2024. It is used for credit analysis and corporate-banking support, and is notably a trilingual Japanese + English + Chinese model.

Nomura Holdings launched Nomura Research Assistant in 2025, exposing 70,000 of its own research notes plus global broker reports to natural-language search. Nomura's Nomura Securities Quant unit is expanding its ML-based algorithmic-trading business.

SMBC partners with Microsoft to run its Azure-OpenAI-backed in-house assistant SMBC-GPT. Mizuho launched the Anthropic-Claude-based Mizuho AI Lab in 2025.

NTT Data Banking AI is an NTT Data-operated banking-specialized AI platform. NTT Data wins many of Japan's banking-system SI contracts, and this platform is deployed at 50+ Japanese regional (prefectural) banks. These regional banks have small IT budgets and cannot run their own LLMs, so SI vendors like NTT Data provide shared infrastructure.

15. Insurance AI — Lemonade, Hippo, Tractable

Lemonade, founded in 2015 and IPO'd on NYSE in 2020, is a digital insurer whose key differentiators are two AI bots: Maya and Jim.

• Maya: quotes and binds new policies in under 90 seconds, collecting about 200 data points for auto-underwriting
• Jim: claims processing. Accident photos and recorded voice statements are analyzed by an LLM; simple claims are auto-paid in 3 seconds

In 2024 Lemonade reached roughly 2 million users across the US, EU, and UK, and reported auto-resolving about 35 percent of claims without human involvement. The loss ratio remains higher than traditional insurers, so the path to profitability is still slow.

Hippo Insurance (home) differentiates with IoT sensors, satellite imagery, and LLM claims processing. Root Insurance (auto) prices premiums via smartphone telematics. Both went public via SPACs but have struggled to hold value.

Tractable, founded in London in 2014, is a B2B AI company whose flagship model estimates damage and repair cost from photos of car accidents. It is deployed at 30+ global insurers — Allstate, GEICO, AXA, MS&AD — and a 2024 Series E priced it at roughly 1 billion USD.

Similar categories include Shift Technology (insurance fraud detection), Clearcover (auto), and Sure (embedded insurance).

16. PE / VC AI — Carta, Crunchbase, PitchBook

Carta, founded in 2012, is a cap-table platform managing about 40,000 private companies and PE/VC funds as of 2024. Carta AI launched in 2024 to answer natural-language queries like "show me companies in my portfolio whose last valuation is more than 6 months old." AI recommendations now sit inside 401(k) and ESPP employee-compensation modules too.

Crunchbase, founded in 2007, is a private-company data platform; Crunchbase AI launched in 2024 handles natural-language queries like "AI infrastructure companies that raised a Series A in the last 6 months" without SQL. In 2025 a predictive feature was added — which companies are most likely to raise their next round — and VCs use it for outbound sourcing.

PitchBook (a Morningstar subsidiary, founded 2007) holds more than 600,000 PE/VC/M&A deals. PitchBook AI launched in 2024 handles queries like "average ARR multiples for healthcare SaaS over the last 3 years" in natural language.

In PE/VC workflows the biggest AI footprint is in deal sourcing and portfolio monitoring. Sourcing combines Crunchbase, PitchBook, and LinkedIn through LLMs to surface companies matching a fund's investment theme daily. Monitoring automates KPI collection (Daloopa-style), quarterly report summarization, and crisis alerts.

Adjacent tools include Affinity (relationship intelligence, CRM), Visible (startup-to-investor reporting automation), and Mosaic (FP&A SaaS with AI integration).

17. Credit / Lending / Risk AI — Zest, Upstart, Featurespace

Zest AI, founded in 2009, is a credit-underwriting AI company whose core product is "machine-learning models that complement traditional FICO scores." Deployed at 200+ US credit unions and some mid-sized banks, it automates auto and personal-loan underwriting. Its differentiator is model explainability and adverse-impact verification (Adverse Action Reasons), required to comply with the US ECOA (Equal Credit Opportunity Act).

Upstart, founded in 2012 and Nasdaq-listed in 2020, is an AI lending platform that underwrites personal loans, auto loans, and small mortgages with its own model. The pitch — non-traditional data like education, employment history, and spending lets people who would be denied on FICO alone get approved — became harder when defaults spiked through 2022 to 2024 rising-rates cycles, and the stock fell sharply.

Featurespace, founded in Cambridge in 2008, is a fraud-detection company that Visa acquired for 950 million USD in 2024. Its flagship ARIC uses Adaptive Behavioral Analytics — learning the user's behavioral patterns over time, not transaction-by-transaction. HSBC, NatWest, and TSYS use it.

Similar categories include Sift (e-commerce fraud), Hawk:AI (AML), Feedzai (payments fraud), NICE Actimize (large-bank AML), and Resistant AI (document fraud).

18. Robo-advisors — Wealthfront, Betterment, Empower

Wealthfront, founded in 2008, is a robo-advisor managing roughly 50 billion USD as of 2024. Core products are auto-rebalancing, tax-loss harvesting, and smart-beta portfolios. Wealthfront AI launched in 2024 answers natural-language queries like "to reach 800,000 USD by age 60, how much should I invest per month."

Betterment, founded in 2008, manages around 45 billion USD with nearly the same product set as Wealthfront but with stronger 401(k) integration and a human-advisor option.

Empower (formerly Personal Capital, founded 2009, acquired by Empower Retirement in 2020) monitors roughly 1.5 trillion USD. More than a pure robo-advisor, it pairs an asset-aggregation dashboard with human advisors; Empower AI was integrated in 2025.

Korea has Aim, Fount, and Bulleo in the same category; Japan has WealthNavi and THEO. All run their own AI recommendation engines.

19. Algorithmic Trading — QuantConnect, NautilusTrader, Backtrader

Algorithmic trading is statistical and ML-heavy and largely separate from LLM territory, but it is one pillar of the finance AI ecosystem.

QuantConnect Lean (quantconnect.com) is an open-source algorithmic-trading engine where strategies are authored in C# or Python for backtesting and live trading. As of 2024 it has a community of about 300,000 quant developers and integrates with Interactive Brokers, Tradier, and others.

NautilusTrader (nautilustrader.io) is a Rust-core + Python-API high-performance backtest and live engine that has been rapidly adopted in the open-source quant community since 2023. It is strong on latency-sensitive strategies and multi-asset coverage (equities, futures, crypto).

Backtrader is a 2015-launched Python backtesting library; its low barrier to entry made it popular with beginners, but active updates have slowed.

Alternatives include Zipline (forked from Quantopian, currently maintained by Stefan Jansen), Vectorbt, and bt.

These tools do not directly combine with LLMs, but "describe a strategy in natural language and get backtest code" GPT integrations grew through 2024 and 2025. Composer.trade and Trade Ideas Holly AI are commercial products pursuing that direction.

20. Core Datasets — EDGAR, DART, EDINET, Compustat

Financial LLM quality depends directly on training and retrieval data quality. As of 2026 the most important public and paid data sources are as follows.

SEC EDGAR (US) (www.sec.gov/edgar) provides every filing for US-listed companies (10-K, 10-Q, 8-K, S-1, DEF 14A, 13F, etc.) for free. From 2024, XBRL-structured data is nearly universal, making machine processing far easier.

Korea DART (dart.fss.or.kr) is the disclosure system run by Korea's Financial Supervisory Service, free for Korean-listed-company filings. Programmatic access is via the Open DART API.

Japan EDINET (disclosure2.edinet-fsa.go.jp) is operated by Japan's Financial Services Agency with full XBRL adoption.

Yahoo Finance, IEX Cloud, and Polygon.io are the popular price-data sources. Yahoo is free but rate-limited, IEX provides IEX-exchange data without a full SIP license, Polygon offers full NMS data on a paid plan.

Compustat (S&P) and CRSP (University of Chicago Booth) are the standard datasets for academia and quant funds. Compustat covers fundamentals; CRSP provides backtest-grade price, dividend, and corporate-actions data. Academic licenses are expensive (tens of thousands of USD per year) but data quality is unmatched.

WRDS (Wharton Research Data Services) bundles these datasets into a single SQL interface — the de-facto standard for graduate quant research.

Alternative data includes YipitData (e-commerce traffic), Second Measure / Bloomberg Second Measure (card transactions), Thinknum (web-scraped KPIs), and Predata (geopolitics).

21. Evaluation Benchmarks — FinBen, FLUE, FinanceBench

Evaluating domain LLMs needs metrics different from general-purpose LLMs. The most-cited benchmarks as of 2026 are as follows.

• FinBen: 24 financial tasks (sentiment, NER, headline classification, QA, summarization, stock prediction) — github.com/The-FinAI/PIXIU
• FLUE (Financial Language Understanding Evaluation): 5 tasks across sentiment, headline, QA
• FinanceBench: 10,231 SEC-filing QA pairs — evaluation-only, not training
• FNS-2023, ConvFinQA: multi-turn conference-call QA
• MultiFin: multilingual financial classification

FinanceBench matters because it differs from generic QA — answers are human-verified against actual 10-Ks and 10-Qs, and many questions require cross-page reasoning over tables and footnotes. At publication in 2024, GPT-4 scored about 19 percent — a striking demonstration of "how weak general-purpose LLMs are on financial documents."

22. RAG Patterns — Finance-domain Retrieval

About 90 percent of financial LLM use cases are not pretraining but RAG (Retrieval-Augmented Generation). Financial RAG has these specific quirks.

# Example financial RAG pipeline (simplified)
from langchain.vectorstores import Pinecone
from langchain.embeddings import OpenAIEmbeddings
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.chat_models import ChatOpenAI

# 1) Chunking — tables go whole, text by paragraph
splitter = RecursiveCharacterTextSplitter(
    chunk_size=1500,
    chunk_overlap=200,
    separators=['\n## ', '\n### ', '\n\n', '\n', '. ', ' '],
)

# 2) Metadata — company, fiscal year, doc type, page, table id
def add_metadata(chunks, filing):
    return [{
        'text': c.page_content,
        'ticker': filing.ticker,
        'fiscal_year': filing.fy,
        'doc_type': filing.type,
        'page': c.metadata.get('page'),
        'section': c.metadata.get('section'),
    } for c in chunks]

# 3) Hybrid search — BM25 + dense embeddings
retriever = Pinecone(
    embeddings=OpenAIEmbeddings(model='text-embedding-3-large'),
    namespace='sec-filings',
    hybrid_alpha=0.7,  # 30 percent BM25, 70 percent dense
)

# 4) Metadata filtering — restrict to Apple 2024 10-K
results = retriever.similarity_search(
    'gross margin trajectory',
    filter={'ticker': 'AAPL', 'fiscal_year': 2024, 'doc_type': '10-K'},
    k=10,
)

# 5) Generate answer with citations
llm = ChatOpenAI(model='gpt-4o')
answer = llm.invoke([
    {'role': 'system', 'content': 'Cite source with [filing, page] format.'},
    {'role': 'user', 'content': f'Context: {results}\n\nQuery: ...'}
])

The key tricks: (1) tables as whole chunks — splitting a table breaks meaning. (2) Hybrid search — accounting vocabulary demands exact word matches, so BM25 stays in the mix. (3) Metadata filtering first — narrowing to company and year before semantic search greatly reduces hallucination. (4) Mandatory citation format — every answer must trace back to a source page.

AlphaSense, Hebbia, Daloopa, and Tegus all build on this RAG architecture; the differentiation is in data-curation quality and domain-specific post-processing.

23. Regulation — EU AI Act, US SR 11-7, Korean Financial AI Guidelines

Finance AI sits in the most heavily regulated domain.

EU AI Act (effective 2024, phased application 2026 to 2027) classifies credit scoring and insurance underwriting as "high-risk AI systems," requiring risk assessment, data governance, documentation, human oversight, accuracy, and robustness.

US SR 11-7 (Federal Reserve "Guidance on Model Risk Management", 2011) requires (1) development validation, (2) usage monitoring, and (3) a governance framework for every model used by a bank — LLMs included. Supplemental OCC/FDIC guidance in 2023 added interpretation for LLMs and generative AI.

Korean FSC AI Guidelines (effective 2024) codify five principles: explainability, non-discrimination, human oversight, model governance, and pre-disclosure. AI used in credit scoring, wealth management, and insurance underwriting must be pre-reported.

Japan's Financial Services Agency (JFSA) published "Principles for AI Utilization" in 2024 — a self-governance model that nonetheless addresses LLM hallucination risks explicitly.

This regulatory layer is the biggest constraint on LLM rollout pace, and is also the growth tailwind behind governance infrastructure (LangChain Tracing, Anthropic AGI Safety, Confident AI).

24. Limits and Open Problems

Even in May 2026 financial AI carries the following limits.

1) Hallucination and numerical precision: LLMs still misread numbers from tables and confuse accounting periods. As FinanceBench shows, general-purpose LLM accuracy is well below human.

2) Real-time data integration: LLMs have a training cutoff by default, so combining them with real-time market data requires RAG/Tool architecture. When that pipe breaks, you get wrong prices and stale timestamps.

3) Multimodal — charts and tables: GPT-4V and Claude 3.5/4 Vision can read charts to a degree, but precise numerical extraction is still more accurate via OCR plus rules.

4) Causal reasoning: LLMs catch correlations well but causation (why did the exchange rate move) poorly. Macro analysis remains human territory.

5) Regulatory governance cost: Model validation, documentation, and monitoring can cost 5 to 10 times the model itself for a single LLM rollout.

6) Data licensing: Bloomberg, FactSet, and S&P data may itself violate license terms when used for LLM training, so true domain LLM pretraining is feasible only for the few players who own data.

Because of these limits, the "AI replaces the portfolio manager" scenario remains distant in 2026. But analyst information-gathering, document processing, and simulation tools are being replaced by AI quickly, becoming productivity tools that let the same staff cover 10 times the tickers or deals.

25. Learning Path — Where to Start

If you are new to finance AI, recommended sequence.

1. Foundational finance + Python: Investopedia series, McKinney's "Python for Data Analysis", yfinance + Pandas for price-data analysis
2. Financial NLP intro: FinBERT paper, Hugging Face ProsusAI/finbert, FinGPT GitHub README
3. RAG practice: build an SEC 10-K Q&A bot with LangChain or LlamaIndex, use the EDGAR API
4. Benchmark evaluation: evaluate your own model or RAG against FinanceBench and FinBen
5. Try commercial tools: Bloomberg Terminal Anywhere academic license, AlphaSense trial
6. Quant intro: free QuantConnect courses, Hudson & Thames machine-learning books

Finance AI is not just NLP — it demands domain understanding of accounting, finance, and market microstructure. Having CFA Level 1 to 2 accounting and finance grounding will let you critically evaluate LLM output, which is invaluable.

26. References

• BloombergGPT paper — https://arxiv.org/abs/2303.17564
• Bloomberg AI blog — https://www.bloomberg.com/company/values/tech-at-bloomberg/
• FinGPT GitHub — https://github.com/AI4Finance-Foundation/FinGPT
• FinRobot GitHub — https://github.com/AI4Finance-Foundation/FinRobot
• InvestLM (NYU Stern) — https://github.com/AbaciNLP/InvestLM
• PIXIU / FinMA — https://github.com/The-FinAI/PIXIU
• DocLLM (JPMorgan) paper — https://arxiv.org/abs/2401.00908
• AlphaSense — https://www.alpha-sense.com/
• Tegus — https://www.tegus.com/
• Daloopa — https://daloopa.com/
• Hebbia — https://www.hebbia.com/
• BlackRock Aladdin — https://www.blackrock.com/aladdin
• JPMorgan AI Research — https://www.jpmorgan.com/technology/artificial-intelligence
• Goldman Sachs Engineering — https://www.goldmansachs.com/our-firm/engineering
• Morgan Stanley AI — https://www.morganstanley.com/articles/ai-research-assistant
• Bridgewater Associates — https://www.bridgewater.com/
• Two Sigma — https://www.twosigma.com/
• Numerai — https://numer.ai/
• Lemonade — https://www.lemonade.com/
• Tractable — https://tractable.ai/
• Carta — https://carta.com/
• Crunchbase — https://www.crunchbase.com/
• PitchBook — https://pitchbook.com/
• Zest AI — https://www.zest.ai/
• Upstart — https://www.upstart.com/
• QuantConnect Lean — https://github.com/QuantConnect/Lean
• NautilusTrader — https://github.com/nautechsystems/nautilus_trader
• SEC EDGAR — https://www.sec.gov/edgar
• Korea FSS DART — https://opendart.fss.or.kr/
• Japan EDINET — https://disclosure2.edinet-fsa.go.jp/
• FinanceBench — https://github.com/patronus-ai/financebench
• EU AI Act — https://artificialintelligenceact.eu/
• Korean FSC AI Guidelines — https://www.fsc.go.kr/