Where Does the LLM Fit?

Most people start with the wrong request:

“Build me a tool that flags risky trades.”

It sounds reasonable.

But it asks the LLM to do the wrong job.

An LLM should not decide whether a trade is risky. It should explain a risk decision after the real risk system has already measured the risk and chosen the action.

Think of it like a hospital

A good risk system has three separate jobs.

1. Measure

In a hospital, this is the blood test.

The test might say:

Your inflammation marker is 9.2

In trading, this is the risk model.

The model might say:

Margin-call probability = 0.88

This job is about producing a number.

Not a feeling.

Not a sentence.

A number.

2. Decide

In a hospital, the doctor follows a protocol.

For example:

If the marker is above 8.0, admit the patient to ICU.

In trading, the policy rule does the same thing.

For example:

If margin-call probability is above 0.80, freeze leverage and alert the risk desk.

This job is about choosing the action.

3. Communicate

In a hospital, the nurse explains the situation to the family in plain language.

In trading, this is where the LLM belongs.

The LLM can write:

This member has an 88% estimated margin-call risk, mainly due to high leverage and correlated positions. Leverage has been frozen and the case has been sent to the risk desk for review.

That is the right job for the LLM.

It explains the decision.

It does not make the decision.

The common mistake

Nobody asks the nurse to run the blood test, diagnose the patient, and decide the treatment.

But that is exactly what people do when they ask an LLM:

“Is this trade risky?”

That question sounds simple, but it gives the LLM too much responsibility.

Why the LLM should not be the risk model

Ask an LLM if something is risky, and it may say:

This account shows concerning risk characteristics and significant leverage.

That sounds serious.

But what does it actually mean?

Does it mean there is an 80% chance of a margin call?

A 15% chance?

A 3% chance?

You cannot tell.

That is the problem.

Language can sound confident without being precise.

Risk systems need numbers that can be tested against history.

An LLM gives you persuasive wording.

A risk model gives you a probability.

For trading risk, you need the probability first.

The fix: answer four questions before building

Before writing code, answer these four questions.

1. What is the real number?

Do not start with:

This trade is risky.

Start with something measurable:

Margin-call probability = 0.88

That number can be checked later.

If the model says 0.88, you can ask:

Did accounts like this actually margin call around 88% of the time?

That is how you improve the system.

You cannot improve vague language in the same way.

2. What can we actually do about it?

An alert is not an action.

An alert just says, “Someone should look at this.”

That is often too vague.

Instead, define real actions:

• Do nothing
• Send a soft warning
• Freeze leverage and send to human review

Now the system knows what to do.

The LLM can explain the action after it happens.

3. What does it cost to be wrong?

Not all mistakes have the same cost.

If you warn a healthy account by mistake, maybe the cost is small.

Maybe the member is annoyed.

Maybe support gets one extra message.

Maybe it costs $5 in friction.

But if you miss a real margin call, the cost can be much higher.

It might mean financial loss.

It might mean legal risk.

It might mean a serious review from the risk team.

That could easily cost $2,000 or more.

These two errors are not equal.

So the system should not treat them equally.

4. Where does the LLM actually belong?

The LLM belongs last.

First, the risk model measures the risk.

Then, the policy decides the action.

Then, the LLM explains the result in plain English.

That order matters.

The LLM should not say:

I think this account is risky.

It should say:

This account has an 88% estimated margin-call risk. The main drivers are 4.2x leverage and two highly correlated positions. Based on policy, leverage has been frozen and the case has been sent to the risk desk.

That message is useful because it includes:

• The number
• The reason
• The action already taken

What the finished system looks like

Here is the simple structure.

Risk model

Job:

Generate the risk score.

Example:

Margin-call probability = 0.88

Decision logic

Job:

Apply the policy rule.

Example:

If probability is above 0.80, freeze leverage and notify the risk desk.

LLM

Job:

Write the message in clear language.

Example:

Member is at 88% margin-call risk, driven by 4.2x leverage across two correlated positions.

Feedback loop

Job:

Learn from what actually happened.

Example:

Did the margin call happen? Use the outcome to retrain the model.

The important shift

The goal is not to make the alert sound more dramatic.

The goal is to make it more useful.

A bad alert says:

This account appears risky.

A better alert says:

This account has an 88% margin-call probability. The risk is driven by 4.2x leverage across two correlated positions. Leverage has been frozen and the risk desk has been notified.

That is the difference.

The first one sounds worried.

The second one helps someone act.

Final rule

Do not ask the LLM to decide risk.

Use the model to measure risk.

Use policy to decide the action.

Use the LLM to explain the decision clearly.

That is where the LLM fits.