LLMs are changing how businesses work with information. But when it comes to forecasting, there is a big difference between generating an answer and generating a forecast you can trust.

The problem with asking an LLM to forecast

It is easy to see why people are trying it. Give an LLM some historical data, ask what happens next, and you get an answer in seconds.

The answer may even sound convincing.

But forecasting is not about producing a convincing answer. It is about producing a forecast that can be tested, measured, and trusted.

That is where LLMs become a tricky fit.

In one sense, LLMs and forecasting models do something similar: both predict what comes next in a sequence. The difference is in what they learned from and how they are judged.

An LLM learned the patterns of language. A forecasting model is fitted to your data and tested against what actually happened. In language, there are many acceptable ways to finish a sentence. In a forecast, only one outcome will actually happen.

A forecast needs to be repeatable

Ask an LLM the same question more than once and you can get different answers. For a conversation, that can be okay. For a forecast, it can be a problem.

A forecasting process should give you a stable result that can be compared with what actually happened. But LLMs have a built-in element of randomness, so the forecast can change even when you ask the exact same question twice. Change the wording of the prompt, and it can change again. That makes it much harder to know whether the model is responding to the data or simply to the way the question was asked.

This matters even more when forecasting several periods ahead. If each predicted value becomes part of the next step, small differences can build up over time.

In other words: a little randomness is harmless in a conversation. In a forecast, it gets in the way.

A forecast is more than a number

There is another important difference. A good forecast does not just tell you what is expected to happen. It also tells you how uncertain that expectation is.

Imagine a forecast of 12,400 units next month. That sounds precise. But what if the realistic range is between 10,800 and 14,300?

For an inventory decision, that range may matter more than the 12,400 point estimate.

Forecasting methods are designed to measure this uncertainty. An LLM can explain an existing prediction interval. But asking the LLM to invent its own statistically meaningful uncertainty range is a different story.

The lesson is simple: a forecast should tell us not only what we expect, but also what we don't know.

The future does not always look like the past

Historical data contains useful information. But business conditions change.

A new competitor can enter the market. Prices can change. A promotion can move demand. A supply problem can disrupt sales. A new regulation can change the relationship between two variables.

A recent trend can therefore be a useful signal without being a reliable prediction of what comes next.

An LLM can recognize patterns very well. But recognizing a pattern is not the same as understanding why the pattern exists or whether it will continue.

Real forecasts rarely depend on one variable

Most business forecasts are not simply about extending last year's sales into next year.

Demand can depend on price, promotions, inventory, competitors, economic conditions, weather and many other indicators.

The challenge is not just finding variables that moved together in the past. The challenge is finding variables that actually help predict the future.

This is where proper forecasting methods and backtesting matter. An indicator may look important in historical data and still add little value when the model is tested on periods it has not seen.

An LLM can help find potential indicators or summarize research. But deciding how much each indicator should influence the forecast is a different job.

How much does a 10% competitor price cut affect your demand? The answer depends on your customers, your market and your history. A forecasting model measures that relationship in your data and then tests it on periods it has not seen.

An LLM does neither. It gives you a plausible answer based on what it has read, not on what your data shows. And as we saw earlier, it may give you a different answer next time.

So, should we not use LLMs in forecasting?

We should, just not to produce the numbers. LLMs can be very useful in the forecasting process. The opportunity is to use them around the forecast, rather than asking them to become the forecasting engine.

Imagine your forecasting system has already produced the following:

·        Demand forecast: 12,400 units

·        Expected growth: +8%

·        Current inventory: 10,200 units

·        Supplier lead time: 21 days

·        Service-level target: 95%

·        Competitor price change: -10%

The LLM does not need to calculate the 12,400. The forecasting model has already done that.

Now the useful questions begin:

·        What is driving the forecast?

·        Does the inventory position look consistent with the forecast?

·        Which assumptions should we pay attention to?

·        Are any indicators telling us a different story?

·        What scenarios should we test?

This is where an LLM can become genuinely useful, with one condition. Its answers need to come from what the model tells it. If the forecasting system says what is driving the 12,400, the LLM can explain it. If it does not, the LLM will fill the gap with a plausible guess, and we are back to the original problem.

Give the LLM a business state, not a blank page

There is a big difference between asking an LLM, “What will demand be next month?” and giving it the current business state and asking, “What should we check before acting on this forecast?”

A structured business state might contain the latest forecast, inventory, prices, leading indicators, recent changes and the assumptions behind the model.

The same logic that applies to the forecast applies to the decision. A good decision should be transparent and repeatable, and someone needs to own it. So the LLM should not make the decision either. Its job is to work through the many pieces of information, give the decision maker a clear overview and help reveal issues and blind spots.

The forecasting system calculates the numbers. The LLM helps people make sense of them. People make the decision.

That simple separation makes the use case much more practical.

Where LLMs can help

Used this way, an LLM can add a useful layer on top of an existing forecasting process.

·        Explain what complex models are doing, including why a forecast changed.

·        Highlight conflicting signals or assumptions.

·        Summarize market intelligence and connect it to the forecast.

·        Help users think through different scenarios.

·        Turn technical forecasting output into language that business teams can use.

·        Help people ask better questions of the forecasting system.

None of these requires the LLM to calculate the forecast itself.

Where the LLM should stay out

The closer we get to the actual numerical forecast, the more important it is to use methods that can be tested and validated.

·        Do not let the LLM invent the official forecast when a validated forecasting model already exists.

·        Do not rely on it to create statistically meaningful prediction intervals from scratch.

·        Do not treat a relationship it finds in text or data as proof of causation.

·        Do not replace a validated model simply because the LLM gives a more convincing explanation.

·        Do not let it make high-impact operational decisions without appropriate controls.

The point is not to limit the LLM. It is to give it the right job.

A better way to think about it

The forecasting process can be simple:

·        Calculate: the forecasting model produces the forecast.

·        Structure: the forecast and relevant business information are brought together.

·        Reason: the LLM explains, questions and explores that information.

·        Decide: the business decides what to do.

If a new calculation is needed, the LLM can also ask the appropriate forecasting or simulation tool for the answer and then explain the result. The important part is that the number comes from the analytical system, not from the language model guessing.

The takeaway

LLMs are likely to change how people use forecasting systems. They can make forecasts easier to understand, make scenario analysis more accessible and help connect numbers with the wider business context.

But they do not need to become the forecast.

A language model can be a powerful forecasting assistant without being a trustworthy forecasting engine.

There is one more option we have not covered: asking an LLM to write the forecasting code for you. That is a step up from asking it for the numbers directly, but it brings risks of its own. Code that runs is not the same as a model that is right. We will look at that in a future post.

The best use of an LLM may not be to predict the future. It may be to help people understand the forecast, challenge its assumptions and decide what to do next.