Decision Trees

Why it matters

A hard decision under uncertainty is a tangle — until you draw it: your choices and chance’s coin-flips as one branching map. Then the best first move falls out, by working backwards from the payoffs at the tips.

For example: you can drill a wildcat oil well for $1M, or walk away. Drilling isn’t one outcome — it’s a gamble: maybe a 30% shot at striking oil worth +$4M net, a 70% chance of a dry hole that costs you the $1M. Held in your head, the two paths blur into a queasy “it might pay off, it might ruin me.” Drawn as a tree, the fog clears. Your choice is a fork — drill or walk. The drill branch runs into chance, which splits into strike and dry. Put the dollars at the tips, weight them by their odds, and the drill path is worth, on average, 0.30 × (+$4M) + 0.70 × (−$1M) = +$0.5M — better than the $0 of walking away. So you drill, clear-eyed about the 70%. The picture didn’t change the odds; it made them decidable.

What it reveals. Which first move is best when the outcomes are uncertain — by laying every choice, every chance event, and every payoff on one map, then computing the average worth of each path so they can actually be compared.
How it changes the read. You stop arguing about a single scary outcome (“what if the well’s dry?”) and start weighing the whole distribution — every branch, weighted by its odds — so a good gamble with an ugly downside stops masquerading as a bad bet.
When to foreground it. Any decision that branches: options with different risk profiles, or stages where a later choice depends on how an earlier gamble turns out, that are too tangled to compare by gut.
What you’d miss without it. The value of information — the same picture shows whether paying for a test (a seismic survey, a pilot, a second opinion) that resolves the uncertainty before you commit is worth more than its price.
Where it misleads. A tree is exactly as honest as the numbers fed into it — invented probabilities and a forgotten branch produce a confident wrong answer. And the highest average payoff isn’t always the right choice: someone who can’t survive the bad branch should weigh the ruin, not the mean.

How it works

A man is standing on a patch of Texas scrubland with a decision that could make or break him. The geology says there might be oil under his feet. Drilling a well will cost him a million dollars he can’t easily lose. If he hits, the lease is worth four million net; if the hole comes up dry, the million is simply gone. He could also just walk away and keep his money. He turns it over and over and gets nowhere, because the decision keeps collapsing into a single dreadful question — what if I spend the million and it’s dry? — and that question has no answer, only a knot in his stomach.

So he does something simple on the back of an envelope. He draws a little square for the choice that’s his to make, and from it he draws two lines: drill and walk away. The walk-away line is easy — it ends in a circle of zero; he’s out nothing, in nothing. The drill line is the interesting one, because after he drills, the outcome is no longer up to him — it’s up to the earth. So he marks that handoff with a different shape, a circle, to mean now chance decides, and from the circle he draws the two things chance can do: strike, ending at +$4 million, and dry, ending at −$1 million. Above each of those two lines he writes his best honest guess at the odds: three in ten it strikes, seven in ten it’s dry. In a few pen-strokes the knot in his stomach has become a map — his moves as squares, the world’s moves as circles, every road ending in a number he can read.

Now comes the move that turns the map into a decision, and it runs backwards. He starts at the tips, where the dollars are, and folds the tree inward toward the choice. At the chance circle he asks: what is this gamble worth on average? Three times in ten he gets four million; seven times in ten he loses one. So the average is 0.30 × (+$4M) + 0.70 × (−$1M) — that’s $1.2 million of expected upside minus $0.7 million of expected loss, leaving +$0.5 million. He writes that number on the circle: the drill gamble, taken many times over, is worth half a million. Then he steps back to his own square and simply compares the two branches he can choose between — drill, now stamped +$0.5M, against walk-away at $0. The drill branch wins. The tangle that had no answer has produced a clean one: drill — not because the well is safe (it’ll be dry seven times in ten) but because the rare big strike more than pays for the frequent small losses. He goes in eyes open.

That is a decision tree, and the backwards step is called folding back (or rolling back) the tree: you push the uncertainty toward the present by replacing each chance fork with its expected value — its probability-weighted average payoff — and replacing each choice with its best branch, until the whole map reduces to a single recommended first move. The whole apparatus is just three honest commitments made visible: here are my real options, here is how likely each outcome is, and here is what each outcome is worth. Put them on paper and the best choice is forced, not felt.

And then the picture quietly does something cleverer. Suppose a geologist offers our wildcatter a seismic survey for $100,000 that would tell him, before he commits the million, whether this particular hole is wet or dry. Is the test worth buying? Drawn as a tree, the answer is right there. With perfect foreknowledge he’d drill only the wet holes and skip the dry ones — so you draw that world, fold it back the same way, and compare it to drilling blind. The gap between “decide knowing” and “decide guessing” is the most the information could possibly be worth; if that gap clears the $100,000 price, you buy the survey. This is the value of information — the single most underrated thing a decision tree tells you. It’s why the same little map that picks your move can also tell you when the smartest move is to not decide yet, and pay instead to make the uncertainty go away.

The discipline, and the danger, both live in the inputs. A tree built on probabilities someone invented to make the math tractable, or with an option quietly left off because it felt unfamiliar, will hand you a wrong answer wearing the costume of a precise one. And the average is not the whole story: a gamble worth +$0.5 million on average is still a coin that comes up “lose a million” seven times in ten, and a person who would be wiped out by that loss should not be seduced by the mean. Used well, the tree is the most honest instrument there is for a branching decision — it makes you say your options out loud, price your uncertainty, and look the downside in the eye. Used carelessly, it launders a guess into a number. The craft is knowing the difference.

Framework & implementation

Origin and evidence

The decision tree is the core apparatus of decision analysis, the discipline that gave structured form to choices under uncertainty. Its canonical statement is Howard Raiffa’s Decision Analysis: Introductory Lectures on Choices Under Uncertainty (1968), which set out the tree, the rollback (folding-back) algorithm, and the value of information as a teachable, rigorous method — building on the Bayesian-decision program Raiffa developed at Harvard with Robert Schlaifer, and on Ronald Howard’s parallel coining of “decision analysis” (1966). The technique reached practising managers a few years earlier through John F. Magee’s “Decision Trees for Decision Making” in the Harvard Business Review (1964), which is where the now-standard squares-for-choices, circles-for-chance diagram and the very wildcatter-style investment examples entered business practice. (Magee’s article predates stable digital identifiers and carries no DOI; it is cited here by reference rather than link.) The deeper lineage runs back to the expected-utility foundations of von Neumann and Morgenstern and to Bayesian probability; the modern operational descendant is Ronald Howard’s influence diagram, a more compact cousin for large problems. Decision trees are now standard in operations research, management science, medical decision-making, and the teaching of decision theory.

Applications and common uses

The decision tree is a working instrument wherever a consequential choice branches into uncertain outcomes — used both to pick the move and to price the uncertainty.

Capital projects and R&D. Drill-or-walk, build-or-buy, expand-now-or-wait, advance-the-drug-to-the-next-trial — staged investments where each commitment buys a gamble on the next, and folding back tells you whether to commit, defer, or kill.
Medical decision-making. Treat, test, or watch-and-wait, with branches for diagnostic accuracy and treatment response — decision trees (often in quality-adjusted life-years rather than dollars) are a backbone of clinical decision analysis and cost-effectiveness work.
Valuing information. The standout use: computing what a market study, a pilot, a diagnostic test, or a second opinion is worth, by comparing “decide knowing” against “decide guessing” — and so deciding when to pay to resolve uncertainty before committing.
Risk and contingency analysis. Litigate-or-settle, insure-or-self-insure, and project risk registers, where the tree makes the probability-weighted cost of each path explicit and exposes which contingencies dominate.
Sequential strategy. Multi-stage business and policy decisions where today’s choice is best understood only by reasoning back from how tomorrow’s choices will play out — exactly the recursive structure folding back was built for.

In every case the move is the same: name the real options, draw the chance forks, price the tips, fold back to the best first move — and check, with a sensitivity sweep, which numbers the answer truly rests on.

Failure modes and when not to use it

The protocol’s characteristic ways of going wrong are catalogued in its Common Failure Modes, with two more the lens names in passing:

False precision. Probabilities and values quoted to several decimals when their basis is rough estimate, dressing a guess as a measurement. The tell is decimal-place confidence over an estimated input. Round inputs to honest precision and report sensitivity bands, so the reader sees the answer’s true firmness.
Pruning by anchor. A material option left off the tree because it’s unfamiliar or unattractive, not because it’s infeasible — so the “best” branch is only best among the ones that got drawn. The tell is a branch excluded on taste rather than feasibility. Include every material option and let the expected-value calculation, not the analyst’s comfort, do the pruning.
Terminal-value mismatch. Tips quoted in mixed units — dollars here, utility points there, a qualitative score elsewhere — so the expected-value sum is comparing unlike things. The tell is an outcome scale that shifts between branches. Convert everything to one unit (typically monetary or utility equivalent) and verify every material consequence is actually on the tip.
Forgetting to fold back. Building the full tree and then reading the answer off the shape — picking the branch that looks attractive — without doing the backwards expected-value pass that is the whole point. A tree never folded back is a diagram, not a decision.
Ignoring risk attitude. Taking the highest expected value as the answer when the decision-maker is risk-averse and the losing branch is ruinous — the expected value is what you’d get on average over many plays, but a one-shot decision-maker who’d be wiped out by the bad branch should weigh that ruin, not the mean. Here the protocol hands off to prospect theory and loss aversion: the right choice may be the lower-mean, lower-variance branch.

When not to reach for it. When the outcomes genuinely can’t be assigned even rough probabilities — deep uncertainty where the odds themselves are unknowable — the expected-value machinery has nothing to compute, and a scenario or robustness approach fits better. When the decision turns on hard constraints or clean multi-criteria trade-offs rather than chance, a constraint-mapping or multi-criteria read is the right instrument. And when the choice is simple and one-shot with an obvious dominant option, a full tree is ceremony — the protocol earns its keep on decisions that branch, stage, or hide their best move behind an ugly downside, not on ones the gut already gets right.

Decision Architecture — the analysis this protocol serves; integrates outcomes, constraints, stakeholders, and failure modes into one structured recommendation, with the decision tree supplying the alternatives-and-odds layer.
Bayesian Reasoning — the discipline of updating probabilities as evidence arrives; it supplies and revises the very numbers a decision tree’s chance forks depend on, and underwrites the value-of-information calculation.
Prospect Theory — why the highest expected value isn’t always the right pick: it describes how a real decision-maker actually weighs gains, losses, and odds, the correction a tree’s risk-neutral average needs.
Margin of Safety — the complementary stance for the ruinous branch: rather than averaging over a catastrophe, build in enough slack that the bad outcome can’t sink you.