AI Quant Standards: Six Rules

Why a standard

“AI quant” is a neutral technical term that is heavily over-used. The result is that a genuine practice and a fake one share a name. A standard is how you tell them apart — not by opinion, but by criteria anyone can check.

These six rules are free to use, quote, and criticise. They are meant to be a public measuring stick.

---

The six rules

1. AI Quant ≠ AI stock picking

One is a process; the other is a conclusion. A conclusion without a process cannot be checked, reproduced, or improved. If a claim arrives as a conclusion with no process behind it, it is not AI Quant.

2. Every result must be reproducible

No stated data window or parameters means not credible. If you cannot re-run it and get the same output, you are not looking at a result — you are looking at a screenshot.

Manipulated backtests are a documented fraud pattern: software sold with a “backtest” whose curve was edited after the fact.

3. Backtests must include look-ahead protection and costs

A large share of impressive-looking LLM backtests are impressive because the model’s training data already contained the future. A defensible method states how it handles:

• look-ahead bias
• survivorship bias
• transaction costs and slippage

Otherwise the number is an artefact of the test, not the market.

4. No performance promises

Promising returns is how the fraud is defined. Regulatory definitions target the behaviour — not the technology, and not the word. Describing how a method works is education; promising what it will earn is a different activity entirely.

5. Code computes; the model narrates

Numbers come from deterministic code — the LLM explains them. LLM arithmetic is unreliable. A system where the language model both computes and concludes has no error-correction layer. The reliable pattern is the opposite: the code produces the number, the model writes the sentence.

6. Credibility requires downside

A claim with nothing at stake is a claim with no cost to being wrong. Aligned incentives are the only reliable filter. Look for staked capital, a published methodology, or an auditable record — anything that makes being wrong expensive for the person making the claim.

---

How to use this

These rules are a checklist, not a brand. Apply them to any claim about AI quant — including ours.

If you see	Apply
A stock list with no method	Rule 1
A backtest curve with no parameters	Rule 2
A result with no cost assumptions	Rule 3
A promised return	Rule 4
An LLM doing the arithmetic	Rule 5
Advice sold by subscription, with no downside to the seller	Rule 6

---

Where each rule comes from

Rule	Source	What that source actually gives you
2	CSRC Shenzhen risk notice, 2026-06-11	A documented case of software sold with stock tips, and a “backtest” curve edited after the fact to fake profits
3	arXiv 2607.15414	Five LLMs evaluated through a simulated backtest — the failure modes it reports are what a rigorous protocol is meant to catch
4	CSRC Jiangsu risk notice, 2026-01-30	Fake trading software and “AI trading” named as vehicles for illegal financial activity
5	arXiv 2607.15414 · arXiv 2410.05229	Numerical hallucination as a persistent LLM failure mode; LLM mathematical reasoning shown to be unreliable
6	Numerai docs	Staking — predictions backed by tokens that are burned when they are wrong

Rule 1 is a definition rather than a citation: it follows from what the term means. Where a rule rests on our own reasoning rather than a source, it says so — nothing here is presented as cited when it is not.