Why Deterministic Forecasting Beats LLM Guesses

A forecast used to move capital has to be reproducible and auditable, and a language model’s free-form guess is neither. That is the whole argument in one line: the number should come from an explicit method you can run twice and take apart, not from a prompt that gives you a plausible figure and hides how it got there.

This is not a complaint about model quality. Modern language models reason well and often produce a sensible-looking estimate. The problem is a different one, and it does not go away as the models improve. A forecast is not just a number. It is a number plus the trail of assumptions behind it, and that trail is what an investor actually argues with, stress-tests, and later grades. Strip the trail away and you are left with something that reads like analysis but cannot be used like analysis.

What “deterministic” actually means here

A deterministic forecast is one produced by a fixed method: you name the drivers, you state the assumptions, you apply the same calculation every time. Feed it the same inputs and it returns the same answer, today and next month, on your machine and on mine. More important than the repeatability is the visibility. Every step is on the page. You can point at any number in the output and trace it back to an assumption you can name and change.

A free-form model answer works the opposite way. You ask for a revenue estimate, it produces one, and the reasoning that led there lives inside a process you cannot inspect and may not get twice. Ask again and the number can shift. Ask slightly differently and it can shift more. None of that makes the model bad at reasoning. It makes the output unfit for a job that requires you to defend, reproduce, and revisit the exact figure.

A forecast you cannot reproduce is a story about the future, not a tool for allocating capital.

Reasoning quality and reproducibility are different things

It is worth separating two ideas that get blurred together. Reasoning quality is how good the thinking is. Reproducibility is whether you can get the same result again and see how it was formed. People assume better reasoning removes the need for reproducibility. It does not. They are orthogonal.

You can have excellent reasoning that is useless because it is not reproducible: a sharp analyst who gives you a great number but cannot tell you which assumption to change when the world moves. And you can have mediocre reasoning that is still useful because it is fully transparent: a plain model where every input is visible, so you can see the weak assumption and fix it yourself.

For capital, transparency usually wins the tie. Not because the transparent method is smarter, but because you can improve it. A forecast whose assumptions are on the table invites disagreement, and disagreement is how forecasts get better. A confident number with no visible working shuts that down. You either accept it or you do not, and neither is analysis.

A forecast is only useful if you can grade it later

Here is the part that matters most and gets discussed least. The point of writing down a forecast is not to be right on the day. It is to be checkable afterwards. Three months later the company reports, reality lands somewhere, and you want to know exactly why you were off so you can fix the method. That post-mortem is only possible if the original forecast exposed its drivers.

Take a real, public example. When a company guides to something like “high single-digit volume growth in the band of about 8 to 10 percent” alongside a margin band it says it intends to hold, that guidance becomes an input you can anchor a forecast to and then grade quarter by quarter. If you built your revenue estimate from a stated volume assumption and a stated price assumption, and volume comes in soft, you can see immediately that the miss was volume, not price, and adjust. That is a driver-based forecast doing its job. The mechanics of building one from the ground up are their own subject, covered in how analysts forecast revenue before earnings.

Now try to do the same post-mortem on a free-form guess. The number was 12 percent growth, reality was 7, and you have nothing to open up. Was the error in volume, in mix, in a segment you never isolated? You cannot tell, because the guess never separated them. You learn nothing, so you improve nothing, and next quarter you are guessing again. A forecasting process that cannot teach you from its own misses is not a process. It is a slot machine with good vocabulary.

Where the language model actually belongs

None of this means keep AI out of forecasting. It means put it where it is genuinely strong and keep it out of the one seat it should not hold.

Language models are very good at the work that feeds a forecast. Reading every page of a filing, pulling the segment lines, tracking what management guided to and how the wording shifted, reconciling a number that appears in three documents, drafting the first-pass narrative. This is reading and structuring, and it is exactly the tireless grunt work that used to eat an analyst’s week. Letting a model do it well is a real gain, and it is a different claim from letting a model be the forecast.

The line to hold is this: the model gathers and structures the inputs; the method turns inputs into the number. The assumptions stay explicit and human-owned. The calculation stays fixed and reproducible. The model’s judgement can inform an assumption, but it does not get to be the assumption in a form nobody can see. This is the same discipline that separates a grounded estimate from a confident one in can AI predict company earnings: the useful part is fast, sourced processing, not an oracle’s number.

There is a related failure worth naming. When the inputs are wrong, no amount of reasoning saves the forecast, and a fluent model will happily reason its way to a precise number on top of bad data. Getting the inputs clean and correct matters more than the cleverness stacked above them, which is the point of data quality beats model quality. A deterministic method at least makes the bad input visible, because it sits there as a named assumption you can check, rather than dissolving into the middle of an answer.

The honest limits of determinism

Determinism is not a promise of accuracy, and it is important not to oversell it. A reproducible forecast can be reproducibly wrong. If the assumption is bad, the fixed method will return the same bad number every time with perfect consistency. Repeatability is a property of the process, not a guarantee about the future.

But that is exactly why it is the right foundation. A wrong deterministic forecast tells you where it is wrong. You open it, find the assumption that broke, argue about it, change it, and the whole thing recomputes cleanly. A wrong free-form guess just tells you it was wrong. The first is a mistake you can learn from. The second is a mistake you are doomed to keep making, because you were never allowed to see how it was made.

Real understanding of a company’s numbers has to respect the structure underneath them rather than treat them as prose, and forecasting inherits that same demand, a theme picked up in building an AI that understands financial statements. The forecast is only as trustworthy as your ability to take it apart.

So the preference is not aesthetic. It is practical. Between a brilliant number you cannot reproduce and a plain one you can inspect, the inspectable one is worth more to anyone actually deploying capital, because it can be checked, defended, and improved. The goal was never a machine that sounds certain. It was a method that stays honest about how it reached the number, so that when the number is wrong, and sometimes it will be, you can find out why.