Prediction = Data × Assumptions
Knowledge Bank · Shelf: Decision Science
Every forecast is a product of two things: the data it was built from and the assumptions it carries. We write the frame as Prediction = Data × Assumptions. It is the working frame we bring to any forecast we read, and it is the reason this house publishes none of its own.
The frame is the owner's own, carried in his study notes for years. It appears in his study index, the "Frameworks & Concepts Index," listed among his recurring frames. His gloss, from a meeting note in the same notebook: "How well the assumptions are determines how good the decision is." The same meeting note pairs the frame with the domain expert's role in data-driven decisions — the person who knows, from the work itself, which assumptions the decision can bear and which it cannot.
The multiplication sign is the content
The frame could have been written with a plus sign. It was not, and the choice carries the meaning.
Addition would let a strong factor compensate for a missing one. Multiplication does not. Worthless data times elegant assumptions is a worthless prediction. Rich data times unexamined assumptions is the same. If either factor is near zero, the product is near zero, however impressive the other factor looks.
That symmetry is the point. A forecast cannot be rescued by the half that looks good. The two halves are joint, and the weaker one sets the ceiling.
The data factor
The data half is the one people discuss. In his Wharton Advanced Business Analytics notebook, a Week 45 essay titled "Predictions = Data + Assumptions" sets out what belongs to it.
Training-data quality decides what a model can learn. Bias–variance is the standing trade: a model can fit its training data's noise and call it signal, which is overfitting. Test data exists to catch exactly that, by asking the model to perform on data it has never seen.
None of this is exotic. It is the ordinary discipline of holding data to account before trusting what comes out of it.
The assumptions factor
The assumptions half is the one that hides.
Every forecast carries assumptions about what will stay constant: growth, margins, behaviour, the world. Most forecasts present their conclusion in large type and their assumptions in small type or not at all. The eye lands on the number; the conditions underneath it go unread.
The frame's working instruction is plain. To read any forecast, find its assumptions first. Then change one and watch the prediction move. The amount it moves is the amount the forecast was the assumption.
A worked illustration from applied work
The frame is not only a reading habit. It shows up in the owner's applied work, as recorded in his notebook.
The frame also shows up in a peer-review exchange from his Wharton coursework. Reviewing a classmate's roadmap, his note on the accuracy lift it claimed was dry: it "seems a significant business case." The number was doing persuasion work.
On his own roadmap — a repair-cost prediction use case drawn from his operations work — classmates noted it "will add value in saving WACC while budgeting." His reply made the assumption explicit instead: he normalized repair-cost variances quarterly or semi-annually. That is, he put the assumption on the table — that costs are comparable across periods once normalized — and chose its period deliberately, rather than letting a default choose it.
The prediction mattered because a decision hung on it: a budget. That is when assumptions deserve the most light, not the least.
The machine-learning version
His analytics coursework adds the same discipline in machine-learning terms.
Choosing a loss function and a metric is an assumption about what errors cost. Precision and recall trade against each other. Evaluating a model is itself a decision about which mistakes matter. None of these choices is given by the data; each is made by a person, and each shapes what the model appears to say.
The house rule that follows
From this frame comes a house rule: we publish no forecasts.
We publish numbers, decompositions, and the assumptions panels behind them, so that a reader can run the confession test on any forecast. Ours would face it too, which is one reason we make none.
The test is simple to state. Find the assumptions. Move one. Watch the prediction move. What moves is what the forecast was made of.
The limits, stated plainly
The frame is a working heuristic from one practitioner's study notes. It is not a theorem and not a measurement. The multiplication here is a picture of joint dependence, not a formula anyone can evaluate.
It cannot grade a forecast by itself. It tells you where to dig, not what you will find.
It says nothing about which assumption is true. It says only that the prediction cannot be better than the weaker of its two factors.
And some forecasts fail for reasons outside both factors: the world changes in ways no dataset or assumption set contained. The frame does not rescue them.
Sources: the owner's study index, "Frameworks & Concepts Index"; his Wharton Advanced Business Analytics notebook, Week 45 essay "Predictions = Data + Assumptions," and the meeting note recorded in the same notebook.