Predicting event outcomes when data is incomplete

The decisions where you most want a prediction are exactly the ones where you have the least data. By the time the data is complete, the event has usually resolved and the prediction is moot. So the real skill isn't predicting with perfect information — it's predicting well with incomplete information, which means being disciplined about what you know, what you're assuming, and how much that gap should widen your range.

This post is about that discipline: how to predict an event's outcome when the data is partial, and why an honest prediction gets more useful, not less, when it admits what it can't see.

Incomplete data is the normal case, not the exception

People treat missing data as a reason to delay the prediction. Usually it's a reason to make one now, because waiting for completeness means waiting until the decision is irrelevant. The question is never "do I have enough data to be certain?" — you won't be — but "given what I have, what's the realistic spread of outcomes, and how wide should the gaps make it?"

A prediction handles incompleteness explicitly. It states what it's inferring, flags which inferences are load-bearing, and widens its outcome range to reflect the uncertainty rather than faking precision it hasn't earned.

Make the unknowns explicit

The single most important move when predicting with partial data is to separate what you know from what you're assuming, out loud. A vague prediction blends the two and hides the risk; a good one says "given that X is true (known) and assuming Y (uncertain), here are the outcomes — and note that if Y is wrong, the prediction flips."

This is where MiroFish's assumption ledger does its heaviest lifting. With complete data the ledger is a nicety; with incomplete data it's the core of the prediction, because the assumptions are the prediction. Reading them tells you exactly which missing fact would most change the answer — and therefore which one is worth the effort to go find.

How outcomes branch under uncertainty

A partial-data prediction tends to produce a wider spread than a data-rich one, and that width is honest signal, not a flaw:

• Resolves as the known factors suggest (~40%): The information you do have is decisive enough that the unknowns don't swing it. The prediction is relatively confident here precisely because the knowns dominate.
• Swung by an unknown (~40%): One of the gaps turns out to be load-bearing, and the outcome hinges on it. This branch is large exactly because the data is incomplete — and it tells you where to look.
• Surprised by something off-board (~20%): A factor nobody was tracking decides it. Bigger when the event is novel and the analogues are thin.

A prediction that collapsed all this into one confident answer would be lying about the second and third branches. The width is the truth.

The variable to chase

When data is incomplete, the prediction does something uniquely valuable: it tells you which missing piece of information would most reduce your uncertainty. Not all unknowns matter equally. Some gaps barely move the outcome; one or two are decisive. The prediction's sensitivity analysis points straight at the high-value unknown — and that's where you should spend your limited time and effort gathering more, rather than chasing data that wouldn't change the answer anyway.

That reframes incomplete-data situations entirely. The prediction isn't just an estimate; it's a research plan. It tells you the one fact worth finding before you act.

When to admit you can't predict it

There's a floor. If the load-bearing unknowns are genuinely unknowable and the event is precedent-free, the honest output is "this is low-confidence — treat the range as very wide." A prediction tool that respects you will say that rather than fabricate a number, the same discipline argued in crypto scenario predictions and why some predictions are more reliable than others.

Predicting with incomplete data is the general case behind every other post in this category — policy changes and market reactions included. Master the discipline of naming your unknowns and you'll predict everything else better.