Method note 04

The assumptions

All models are wrong. This one says where.

The statistician George Box put it in six words: all models are wrong, but some are useful. A model of an economic decision is not a law like the ones physics writes down. It is a deliberate simplification that trades accuracy for tractability: a small set of assumptions that explains a large range of behavior, never perfectly, usually well.

A portfolio model is the same kind of object. It cannot be true; it can only be useful. And a model that hides its assumptions cannot even be judged on that standard. This note sets out the assumptions under the model, what each one buys, where it fails, and what the model does about it.

Where the model sits

There are three ways to support a portfolio decision. Scoring and ranking are easy to explain, but they are not a model at all: no assumption is stated, no limit is represented, so there is no trade-off and no price, and nothing can be tested. Full simulation sits at the other end. It is closer to reality, but it needs data no organization has, and it cannot be explained in the room where the decision is made.

The model takes the middle: constrained optimization over whole projects, with every limit priced. It solves in seconds, its assumptions can be listed, and each result arrives with the price of the limit that caused it. That price is an opportunity cost, what the portfolio gives up because the limit is there, and it is what makes the result explain itself.

Ten assumptions

1. Strategy can be expressed as weighted priorities. This turns intent into a scale on which value can be added. It fails when executives disagree or their weights are unstable, so every weight carries a range, and decisions that would change within it are flagged.

2. Impact can be rated on one scale per priority. This makes very different projects comparable. It fails when the scale saturates or means different things for different priorities, so measured and judged criteria are placed on the same ratio scale.

3. Value adds across projects. This keeps the problem exactly solvable. It fails when projects reinforce or cancel each other. Where the dependence is explicit, it can enter as a constraint; where it is not, it is the model's known blind spot.

4. Projects are funded whole. This reflects real choices rather than fractions of projects. It fails when a project can be phased or scaled, and phases can then enter as separate options. The consequence is that prices move in steps, not smoothly.

5. Costs and values are known. Every model must start somewhere, and every estimate is uncertain. The model's answer is not to pretend otherwise but to measure how wrong an estimate can be before the decision changes.

6. Capacity adds: a person is a person. This assumes constant returns, which keeps planning simple. It fails because output rises with specialization in an area and falls with the cost of coordination, and because a small internal core directs everyone else. Measuring those effects is part of the research programme.

7. Money, internal people and agile capacity are separate limits. Each then has its own price. They do trade against each other, which is why internal capacity is used first and external capacity is bought at the margin, at a stated price.

8. Costs and capacity are known by period. This lets multi-year work fit the plan. Timing slips, so cost and capacity are carried as amounts per period rather than single totals.

9. The organization has one set of priorities. This gives one objective to optimize. Units want different things, so every number that would change a decision has an owner who must agree to it.

10. Priorities substitute for one another at fixed rates. A weighted sum treats them as perfect substitutes, which keeps the objective linear. Real strategy prefers balance, and a weighted sum pushes toward corners: everything into the priority with the best value per dollar. Diminishing value once a target is met is the economic remedy, and testing it is part of the research programme.

The preferences behind the weights

Economics asks three things of preferences: that any two options can be compared, that the comparisons hang together, and that more is better. Each holds less well inside an organization. Executives often cannot say which of two priorities matters more. A single person is usually consistent, but a committee can produce rankings that go in circles, as Condorcet showed and Arrow proved inescapable. And past its target, more of a priority adds little.

The model also asks more than economics does. Textbook preferences only rank. This model needs a ratio scale, on which twice the value means twice, because it divides value by cost and prices every limit. That is a heavier demand on people, and it is why the weights are drawn out with care rather than voted.

The answer to Box

Most builders of models answer Box by claiming their model is close enough. This one answers differently. It does not need to be true; it needs the decision to survive its errors.

For each decision the model measures how far an estimate, a weight or a limit must move before the decision changes. A decision that survives a large error does not depend on the model being right: the model was useful even though it was wrong. A decision that changes on a small error depends on a number nobody can know that precisely, and the model says so and hands it back to people.

The model does not claim to be right. It claims to know which of its answers can be trusted.

What the model leaves to people

The assumptions mark the edge of what the model can know, and beyond that edge people decide. Where the margin is wide, the optimizer proposes and people confirm. Where it is narrow, people review, with the numbers that would change the answer in front of them. Where an assumption does not hold, people decide, and the model prices the choice: overriding it is always allowed, and its cost is shown and recorded.

Stating the assumptions is what makes that division honest. People can see what the model took for granted, and where their own judgment has to take over.

Cite as: Gruia, M. (2026). The Assumptions: All Models Are Wrong. This One Says Where. Frontnovo Method Notes.