ModelAtlas AI Thesis
AI will become a core layer in equity decision workflows
Our view is not that AI is a magic oracle today. Our view is that (1) measurable signal already exists in specific setups, (2) model capability is improving quickly, (3) investor adoption is rising, and (4) the most credible near-term use case is supervised decision support rather than blind auto-execution.
There is already signal
Recent studies suggest LLM setups can extract useful equity signal in specific contexts.
Model quality is improving fast
GPT-5.4 and Claude 4.6 push frontier capability deeper into finance and knowledge work.
Adoption is real
Retail and institutional investors already show growing willingness to use AI in the workflow.
The right shape is copilot
The strongest evidence still supports supervised research support, not blind auto-execution.
Evidence-drivenComparative model evaluationNot investment adviceOpen cockpit
1) Evidence that AI can help in stock analysis
Research is not unanimous, but several studies already report useful predictive signal and workflow efficiency.
Finance Research Letters (2024)

LLM sentiment strategies can extract signal in equities

In a large U.S. news dataset (2010–2023), LLM-driven sentiment showed stronger predictive power than traditional lexicon approaches.

74.4% directional accuracy (paper result)
Open source
Expert Systems with Applications (2025)

Chain-of-thought setups improved financial forecasting in one benchmark

A study comparing LLM forecasts with analyst baselines reported higher earnings-direction accuracy and positive backtest alpha in its sample.

64.35% vs 58.37% (study sample)
Open source
Finance Research Letters (2025)

Prompt robustness still matters in stock recommendations

A 2025 study on LLM stock recommendations found statistically significant variation across repeated prompts, rephrasing, and system-prompt changes.

Reason to compare models and standardize inputs
Open source
2) Why we think this trend should accelerate
Capability and enterprise adoption curves suggest compounding progress rather than stagnation.
OpenAI GPT-5.4 release (2026)

GPT-5.4 pushes further into professional knowledge work

OpenAI positions GPT-5.4 as designed for professional work, with stronger performance in spreadsheet modeling, knowledge work, web research, and lower factual error rates versus GPT-5.2.

87.3% in investment-banking modeling tasks; 33% fewer false claims
Open source
Anthropic Opus 4.6 release (2026)

Claude Opus 4.6 targets economically valuable knowledge work

Anthropic frames Opus 4.6 as a stronger model for finance, legal, research, and long-context agent workflows, with a 1M token context window in beta.

Industry-leading GDPval-AA performance in finance/legal knowledge work
Open source
Anthropic Sonnet 4.6 release (2026)

Claude Sonnet 4.6 compresses high-end capability into a cheaper tier

Anthropic says Sonnet 4.6 matches Opus 4.6 on OfficeQA and materially upgrades long-context document comprehension and knowledge-work performance.

Reason premium document work is getting cheaper and more available
Open source
3) Investor behavior is already shifting
As more people use AI to support allocation decisions, model narratives can influence flows more directly.
World Economic Forum (2025)

Retail investors already show meaningful openness to AI-guided investing

The World Economic Forum’s 2025 outlook reports that 28% already use AI chatbots for investing guidance, 41% would trust AI chatbots with personal financial information, and 42% would invest more with AI chatbot support.

28% already use AI chatbots; 42% would invest more with one
Open source
PwC Global Investor Survey (2025)

Institutional investors want more AI exposure, but with clearer disclosure

PwC’s 2025 Global Investor Survey shows investors want companies to allocate more capital to technological transformation, while also demanding much better disclosure around AI strategy, returns, and innovation.

92% want more tech transformation; only 37% say AI disclosure is enough
Open source
World Economic Forum analysis (2025)

The advice layer is likely to become more hybrid, not less

A World Economic Forum analysis on AI and wealth management argues that AI-driven tools are moving into the advice workflow quickly, but the most credible path remains hybrid models where trust, supervision, and human judgment still matter.

Supports the copilot thesis more than a pure robo-advisor thesis
Open source
4) Why we build this as a copilot, not an autopilot
The strongest recent evidence suggests the winning pattern is supervised research workflow, with transparency and human override still central.
ESMA EU evidence (2026)

Most financial-market AI use is still internal and supervised

Recent EU securities-market evidence points to AI being used mainly for internal workflows, with human oversight remaining the default.

92% internal use cases; 90% human in the loop
Open source
ESMA guidance (2024)

Regulators already frame AI investing as a supervised activity

ESMA’s guidance highlights data quality, opacity, overreliance, and privacy as core issues when firms use AI in retail investment services.

Why our product is copilot-first, not auto-trading
Open source
Anthropic Project Glasswing (2026)

Frontier labs themselves are using restricted rollouts for higher-risk capability jumps

Anthropic’s Project Glasswing describes Claude Mythos Preview as meaningfully stronger than Opus 4.6 on cyber benchmarks and frames it as a model being handled with government discussions and controlled access, not broad consumer rollout.

Reason to keep ModelAtlas as supervised research support with explicit gating and human override
Open source