# Freedom Fund: AI-Assisted Investing Research

Research brief — September 11, 2026

This document is a research and engineering brief, not investment advice. It does not establish a recommendation to buy, sell, short, borrow, use derivatives, or give an AI system access to capital.

## Executive conclusion

There is not yet strong public evidence that a general-purpose large language model (LLM), operating autonomously, can produce durable live trading alpha after realistic costs, taxes, liquidity constraints, and model risk. Recent live-style benchmarks find that many agents struggle to beat simple baselines, and that risk control is more important to robustness than general model intelligence. [AI-Trader](https://arxiv.org/abs/2512.10971) and [StockBench](https://arxiv.org/abs/2510.02209) are useful starting points because they test sequential decisions rather than static finance questions.

The best-supported near-term use of AI is as a constrained information and workflow layer around a conventional quantitative portfolio engine:

1. retrieve and verify evidence;
2. convert evidence into structured, timestamped features;
3. propose hypotheses or candidate signals;
4. challenge the thesis and identify missing data;
5. let deterministic code enforce portfolio, liquidity, cost, and loss constraints;
6. record every input, prompt, output, decision, fill, and post-mortem.

This is an engineering inference from the sources, not a proven profit recipe. FINRA describes industry use of AI for research, portfolio management, smart order routing, price optimization, best execution, and risk/compliance, while warning that autonomous systems can fail in conditions not represented in training data. [FINRA: AI Applications in the Securities Industry](https://www.finra.org/rules-guidance/key-topics/fintech/report/artificial-intelligence-in-the-securities-industry/ai-apps-in-the-industry)

## What people are building

### Research and synthesis agents

These systems use LLMs to search filings, news, transcripts, macro releases, and internal research; summarize the evidence; compare a company or asset with a predefined checklist; and produce a cited investment memo. [FinRobot](https://github.com/AI4Finance-Foundation/FinRobot) is an open-source example of a multi-agent financial-analysis platform. [FinMem](https://arxiv.org/abs/2311.13743) explores layered memory and a decision module for LLM trading agents.

The useful output is a structured record, not a paragraph that says “buy.” Each record should include the source URL, publication timestamp, data vintage, extracted fact, transformation, confidence, counter-evidence, and expiry date.

### Quantitative signal agents

In this pattern, the agent helps discover or explain candidate features, but the signal is calculated by reproducible code. The initial candidates should be deliberately ordinary: momentum/trend, mean reversion, relative value, event windows, volatility, liquidity, and simple factor exposures. The agent is tested on whether it improves research speed or signal quality against a locked baseline—not on whether its prose sounds convincing.

### Execution and risk agents

AI can help produce an order plan, detect unusual fills, reconcile positions, summarize exposure, and flag stale or contradictory data. This may have a better risk/reward profile than asking an LLM to forecast the next price. The execution engine should remain deterministic: allowed instruments, order types, maximum size, price collars, exposure limits, loss limits, and kill-switch behavior should be enforced outside the model.

### Reinforcement-learning and multi-agent systems

Open-source projects such as [FinRL](https://github.com/AI4Finance-Foundation/FinRL) and [FinRL-X](https://github.com/AI4Finance-Foundation/FinRL-Trading) show how researchers are experimenting with reinforcement learning, portfolio weights, and modular trading infrastructure. These are useful research references, not evidence of guaranteed live returns. RL should come after reliable data, baselines, cost models, and walk-forward validation because reward design, seed sensitivity, regime shifts, and transaction costs can dominate the result.

## Strategy lanes worth testing

The ordering below is a research priority, not a return ranking.

### 1. Diversified benchmark and allocation

Start with buy-and-hold, equal-weight, cash, and simple trend baselines. This creates a control group and prevents an agent from receiving credit for market beta. The agent’s first job is research and rebalance documentation.

### 2. Trend and momentum

Test whether persistent price or cross-sectional strength survives costs and turnover. Let code calculate rankings, positions, and sizing. Let the agent classify regime evidence, surface contradictory macro conditions, and explain why a signal is active. The central risk is crowding and rapid reversal.

### 3. Mean reversion and statistical arbitrage

Test relative relationships, spread stability, and mean-reversion half-life. Require an explicit regime-break test and cap gross exposure. The central risks are non-stationarity, correlation breakdown, leverage, and execution slippage.

### 4. Event-driven and financial NLP

Extract structured information from filings, earnings, guidance, and verified news. Enforce publication-time cutoffs so the model cannot see later revisions. The central risks are latency, ambiguous language, duplicate reporting, and trading after the information is already priced.

### 5. Cost-aware execution

Test whether the system improves fill quality, reduces turnover, and avoids trading when expected edge is smaller than fees, spread, slippage, funding, and borrow. FINRA identifies smart order routing, price optimization, and best execution as active AI use cases. [FINRA: AI Applications in the Securities Industry](https://www.finra.org/rules-guidance/key-topics/fintech/report/artificial-intelligence-in-the-securities-industry/ai-apps-in-the-industry)

### 6. RL allocation

Keep this as a later experiment. Compare multiple seeds and policies against simple baselines, lock the out-of-sample period, and report the complete trial count. Do not use a single impressive run as evidence of robustness.

## The data-first model

The Freedom Fund system should treat data provenance as part of the investment model:

- **Market data:** versioned OHLCV/order-book data with corporate actions, exchange calendars, delistings, and a declared timestamp convention.
- **Company facts and filings:** use canonical EDGAR submissions and XBRL facts where applicable. The SEC says its APIs are updated as filings are disseminated and exposes submissions and extracted XBRL data through `data.sec.gov`. [SEC EDGAR APIs](https://www.sec.gov/search-filings/edgar-application-programming-interfaces)
- **Macro data:** use real-time/vintage-aware series. FRED documents that economic observations may be revised and that ALFRED real-time periods can retrieve what was known at a historical date. [FRED real-time periods](https://fred.stlouisfed.org/docs/api/fred/realtime_period.html)
- **News and events:** store the source, publication time, retrieval time, canonical text or hash, and the exact subset exposed to the model.
- **Prompts and models:** version the model name, system prompt, tools, temperature/settings, retrieval set, and output schema.

The agent should never receive a magically clean “historical” dataset that contains information unavailable at the decision time. Every feature needs an as-of timestamp. Data that can be revised needs a vintage or revision policy.

## Evaluation protocol

Every strategy should produce an immutable experiment record before results are viewed:

1. **Specify:** hypothesis, universe, horizon, rebalance rule, data cutoff, costs, constraints, and invalidation rule.
2. **Replay:** run chronological train/validation/test splits with point-in-time data and no test-set tuning.
3. **Stress:** use walk-forward windows, multiple random seeds where learning is involved, cost sensitivity, liquidity shocks, regime slices, and missing/stale-data tests.
4. **Correct for selection:** count every trial. The [Deflated Sharpe Ratio](https://doi.org/10.2139/ssrn.2460551) addresses performance inflation from multiple testing and non-normal returns. [White’s Reality Check](https://doi.org/10.1111/1468-0262.00152) addresses data-snooping risk when many models are searched.
5. **Paper trade:** run the exact production data and execution path without real orders, recording expected and realized fills.
6. **Review:** compare against simple baselines and decide whether the evidence justifies any small, capped live allocation.

Minimum reporting should include net return, volatility, maximum drawdown, downside risk, Sharpe/Sortino, turnover, hit rate, profit factor, exposure, concentration, liquidity, slippage, fees, funding/borrow, capacity, and the distribution of daily and trade-level outcomes. Return alone is not enough.

## Freedom Fund architecture

The proposed control flow is:

`Canonical data → feature code → research agent → quant candidate → risk engine → paper executor → audit ledger`

The LLM can propose and critique. It cannot override the risk engine, invent missing data, silently change a prompt, or place an order outside the allowed schema. The ledger should make it possible to answer: “What did the system know, what did it believe, what did it do, what did it cost, and what happened afterward?”

NIST’s AI Risk Management Framework is a useful operating model: govern, map, measure, and manage continuously, with defined human oversight and a way to deactivate systems whose outcomes drift from intended use. [NIST AI RMF Core](https://airc.nist.gov/airmf-resources/airmf/5-sec-core/)

## Goal math

Assuming $200 on September 11, 2026, a target of $1,000,000 on August 22, 2027, no deposits or withdrawals, and smooth calendar-day compounding, the account needs a 5,000× increase over 345 days:

- daily compounded pace: approximately 2.50%;
- weekly equivalent: approximately 18.86%;
- 30-day equivalent: approximately 109.72%.

This is a mathematical stress test, not a reasonable expected-return assumption. A serious plan should model separate scenarios for additional contributions, business income, time extensions, lower targets, and the probability of total loss. CFTC warns that AI cannot predict sudden market changes and that guaranteed or extraordinary bot returns are a fraud warning. [CFTC: AI Won’t Turn Trading Bots into Money Machines](https://www.cftc.gov/LearnAndProtect/AdvisoriesAndArticles/AITradingBots.html)

## Source inventory

1. FINRA. “[AI Applications in the Securities Industry](https://www.finra.org/rules-guidance/key-topics/fintech/report/artificial-intelligence-in-the-securities-industry/ai-apps-in-the-industry).” Industry use cases and autonomous-system risks.
2. FINRA. “[Algorithmic Trading](https://www.finra.org/rules-guidance/key-topics/algorithmic-trading).” Testing, validation, supervision, and post-deployment review.
3. CFTC. “[AI Won’t Turn Trading Bots into Money Machines](https://www.cftc.gov/LearnAndProtect/AdvisoriesAndArticles/AITradingBots.html).” Investor protection and AI-return-scam warnings.
4. NIST. “[AI Risk Management Framework Core](https://airc.nist.gov/airmf-resources/airmf/5-sec-core/).” Governance, mapping, measurement, management, oversight, and deactivation.
5. SEC. “[EDGAR Application Programming Interfaces](https://www.sec.gov/search-filings/edgar-application-programming-interfaces).” Official filings and XBRL data access.
6. Federal Reserve Bank of St. Louis. “[FRED API: Real-Time Periods](https://fred.stlouisfed.org/docs/api/fred/realtime_period.html).” Historical data vintages and revisions.
7. Fan et al. “[AI-Trader: Benchmarking Autonomous Agents in Real-Time Financial Markets](https://arxiv.org/abs/2512.10971).” 2025 preprint; live-style multi-market agent benchmark.
8. Chen et al. “[StockBench: Can LLM Agents Trade Stocks Profitably In Real-world Markets?](https://arxiv.org/abs/2510.02209).” 2025 preprint; contamination-free sequential stock-trading benchmark.
9. Yu et al. “[FinMem: A Performance-Enhanced LLM Trading Agent with Layered Memory and Character Design](https://arxiv.org/abs/2311.13743).” 2023 preprint; memory and decision architecture.
10. Zhou et al. “[FinRobot: AI Agent for Equity Research and Valuation with Large Language Models](https://arxiv.org/abs/2411.08804).” 2024 preprint; multi-agent equity research architecture.
11. Bailey and López de Prado. “[The Deflated Sharpe Ratio](https://doi.org/10.2139/ssrn.2460551).” Journal of Portfolio Management, 2014; multiple-testing and non-normality correction.
12. White. “[A Reality Check for Data Snooping](https://doi.org/10.1111/1468-0262.00152).” Econometrica, 2000; data-snooping correction.
13. AI4Finance Foundation. “[FinRL](https://github.com/AI4Finance-Foundation/FinRL)” and “[FinRL-X](https://github.com/AI4Finance-Foundation/FinRL-Trading).” Open-source research and infrastructure references; not performance guarantees.