# AI hallucination in finance

AI hallucination in finance is an AI-generated financial claim, figure, citation or explanation presented as factual despite lacking support from reliable evidence.

## What it means

An AI system can produce a convincing answer that includes an unsupported earnings figure, an invented news item or a citation that does not exist. In finance, such errors matter because a seemingly specific claim can change how someone assesses a stock or an order.

A hallucination is about the **support for a claim**, not whether its conclusion sounds plausible. An actual figure quoted with the wrong reporting period is also unreliable, even if the figure itself exists. Missing or outdated data requires a separate warning: an AI system should not fill a gap with a confident guess.

## How it is calculated

AI hallucination has no standard financial formula or universal score. A practical manual check can instead measure the share of factual claims in an answer that remain unsupported after verification:

`Unsupported-claim rate (%) = (U / N) × 100`

- `U` is the number of checked factual claims for which the stated evidence cannot be found or does not support the claim, measured in claims.
- `N` is the total number of factual claims checked, also measured in claims. The rate is undefined when `N = 0`.

This is a checking metric, not the probability that an AI answer is false. Define what counts as a factual claim before counting. A broken link alone does not prove a claim is false; it means the cited evidence has not verified it.

## Worked example

Suppose an AI summary of a hypothetical US stock contains five checkable claims. A reader compares each with the relevant company filing or market-data record, using the same reporting period and share units.

1. The claimed quarterly revenue of US$100 million matches the filing: supported.
2. The claimed quarterly earnings of US$20 million matches the filing: supported.
3. The claim that revenue rose 10% uses the correct comparison periods: supported.
4. A claimed US$5 million dividend cannot be found in the cited filing: unsupported.
5. A claimed trading halt has no support in the exchange notice the answer cites: unsupported.

Here, `U = 2` claims and `N = 5` claims, so the unsupported-claim rate is `(2 / 5) × 100 = 40%`. That percentage describes only this small set of checked claims. It does not estimate how often the system hallucinates in general, and it does not establish that either unsupported event did or did not occur elsewhere.

## Why it matters before you trade

Pre-trade analysis depends on the inputs as well as the reasoning. An invented trading halt could distort an assessment of whether execution is possible. A misquoted volume figure could distort a liquidity assessment; a false earnings claim could change how event risk is understood. A fluent explanation does not repair an incorrect input.

Check consequential claims against a source appropriate to the claim: company filings for reported results, exchange notices for trading status and clearly dated market data for prices or volume. Check the ticker, currency, units, date and reporting period. If a claim cannot be verified, treat it as unconfirmed rather than converting it into a trading fact.

## How PreTrAIde uses it

PreTrAIde performs pre-trade analysis on a user-built stock basket. Its quantitative layer produces deterministic, reproducible estimates and scores; AI agents then reason about each stock and the basket using those scores, recent news and analyst ratings. Its guardrails correct AI output that contradicts specified data: a suspended stock is always NO-GO, a GO with a success probability below 50% is downgraded, and probabilities are clamped to 5%–95%. Guardrails move verdicts only in the conservative direction.

These checks address specified contradictions; they do **not** establish that every AI statement, news item or source is accurate. The GO, NO-GO or MODIFY verdict is decision support, not a recommendation. Users place any trades themselves through their own broker.

## Common mistakes

- Treating a precise number or polished citation as evidence without opening the underlying source.
- Confusing an unsupported claim with a proven false claim; both call for verification, but they are not identical.
- Checking a figure without checking its date, currency, unit or reporting period.
- Assuming a guardrail on verdicts verifies every sentence in an AI explanation.

---
Source: https://pretraide.com/glossary/ai-hallucination