Decision Trees

When a spreadsheet is not enough

A consumer hardware company had two credible paths for a new wearable: launch immediately after a $3 million tooling commitment, or spend $200,000 on a focused market study and decide afterward. Finance built a beautiful NPV (net present value, the discounted sum of future cash flows minus upfront investment) model in a spreadsheet. Marketing had conviction. Operations wanted certainty. The board meeting ended in a stalemate because nobody could explain, in one picture, how today's choice connected to uncertain demand six months out, or how the study changed the odds.

That stalemate is exactly where decision trees earn their place in managerial analytics. A decision tree is a diagram that maps sequential choices and uncertain events over time, with payoffs attached to each path. Unlike a static forecast, a tree forces you to show what you control, what nature controls, and how information arrives before the next fork. From Unit 4 on probability and Unit 5 on forecasting, you already know that uncertainty is not noise to ignore; it is structure to model. Decision trees are the bridge between probability arithmetic and a recommendation a general manager can defend.

Managers who skip trees often confuse average outcomes with decision logic. They might compare a single "expected revenue" line to a study cost without asking whether the study actually changes the decision. They might treat sunk spending as if it still matters at the next fork. They might pick the branch with the highest upside while ignoring a 40% chance of a covenant-breaking loss. Trees do not remove judgment. They make judgment visible so the board can debate assumptions instead of debating vibes.

Anatomy of a decision tree

A decision tree has three kinds of nodes and one clear direction of time. Time flows left to right (or top to bottom). You solve the tree by working backward from the end, a process called rollback.

Decision nodes (often drawn as squares) represent choices you control: launch now, wait, invest, divest, settle litigation, or run a pilot. At each decision node, you compare the rolled-back values of the branches and pick the best one according to your decision rule (usually maximize profit or EMV, defined below).

Chance nodes (circles) represent uncertain outcomes: demand is strong or weak, the study is favorable or not, the jury finds for you or against you. Each branch from a chance node carries a probability, and those probabilities on a single node must sum to 1.0. At chance nodes during rollback, you compute a weighted average of the downstream payoffs.

Terminal nodes (triangles or plain end points) are where the story stops. They hold the payoff for that path: profit, NPV, cost, or a utility score if you are modeling risk attitude. Only future cash flows and costs that differ between branches belong in the payoff. Past spending that cannot be recovered is sunk and should not enter the comparison at a later node (though it may explain why you are at this fork at all).

The power of the diagram is sequential clarity. A two-stage tree might read: (1) choose whether to test the market; (2) if you test, learn good or bad news; (3) then choose launch or cancel. A one-stage model that jumps straight from "test" to "average demand" hides the information structure. For the hardware company, the tree makes the study valuable only if it changes a later choice or shifts probabilities enough to alter EMV.

When building a tree, state information sets explicitly. What do you know at each decision node? If two branches merge incorrectly, you may allow a decision that uses information not yet revealed. That error is common in hiring and litigation models where people accidentally treat early outcomes as known before they happen.

Expected monetary value and the rollback procedure

EMV (expected monetary value) is the probability-weighted average of payoffs at a chance node. If outcome A pays $100 with probability 0.6 and outcome B pays $40 with probability 0.4, EMV = 0.6(100) + 0.4(40) = $76. At a decision node, the rolled-back value of the node is the maximum (or minimum, for cost problems) of the EMV or terminal values on its branches.

Rollback steps, in order:

1. Start at the rightmost terminal payoffs.
2. At each chance node, compute EMV across branches (check probabilities sum to 1).
3. At each decision node, select the branch with the best rolled-back value and mark inferior branches with a "prune" or cut symbol in teaching diagrams.
4. Move left until the root decision is resolved.

The risk-neutral EMV rule says: choose the branch with the highest EMV at the root. Most introductory MBA decision analysis uses this rule because it is transparent and auditable. Real firms may deviate when tail risk threatens survival, when covenants bind, or when executives are risk-averse (they weight losses more heavily than gains). Even then, compute EMV first. It is the baseline against which risk policy is articulated.

Consider a simple launch-or-delay problem without a study. Launch now yields $5 million if demand is strong (probability 0.55) and a loss of $2 million if demand is weak (probability 0.45). EMV(launch) = 0.55(5) + 0.45(−2) = 2.75 − 0.9 = $1.85 million. If delay preserves option value but adds carrying costs, you model delay as its own subtree. The arithmetic is not exotic. The discipline is in labeling what is uncertain and when.

Multi-stage trees add conditional probabilities. After a favorable study, strong demand might rise from 55% to 75% because the study is informative. After an unfavorable study, you might cancel and earn $0 rather than launch into bad odds. The EMV of the "study first" branch equals the EMV of the subtrees minus the study cost on that path. Compare root EMV across first-stage alternatives. The winner is optimal under EMV and stated assumptions.

Always show a check line for probability sums and for rollback arithmetic. A tree that looks persuasive with a hidden 0.35 + 0.50 branch set will mislead a board.

Risk attitude, tail outcomes, and when EMV is not enough

EMV treats a 50% chance of $1 million and a certain $500,000 as equivalent. Many managers do not. Risk-neutral analysis assumes you care only about long-run averages. Risk-averse analysis penalizes volatile outcomes; risk-seeking analysis favors lotteries with the same EMV. For corporate decisions, risk aversion often appears as policy, not as a fancy utility formula: maximum acceptable loss, minimum cash buffer, or "we do not bet the company on one launch."

Litigation is the classic teaching case. Fighting might have EMV −$5.2 million while settling at −$8 million looks worse on average, yet fighting carries a 40% chance of a −$22 million outcome that triggers distress. The general counsel is not irrational; the objective function includes staying solvent. Trees help because you can show the full distribution, not just EMV. A one-line expected value hides the fact that one branch is catastrophic.

Another nuance is correlation. If weak macro demand and weak product demand move together, treating branches as independent may understate downside. Advanced models use joint scenarios; introductory trees often assume independence but should flag the assumption when macro and product shocks align (as in recession lessons from scenario analysis in Lesson 2).

Real options (the value of waiting or staging investments, previewed in Lesson 3 on value of information) appear when delay or a pilot preserves the right to expand or abandon. A tree with a "wait" branch can capture option value that a static NPV misses. If ignoring delay looks like free money, your tree is probably missing a branch.

For communication, pair EMV with best case, base case, and worst case payoffs on the optimal path. Executives often ask: "What do we earn if we are wrong?" The tree answers that without abandoning EMV as the default decision rule.

Building trees that match the decision clock

Good trees mirror the decision calendar. Ask: What is decided today? What uncertainty resolves next quarter? What choices become available only after that resolution? Align payoffs to the same units (all in present-value dollars, or all in nominal year-one cash flow, but not mixed).

Relevance rules keep trees honest:

• Include only incremental cash flows that differ between branches.
• Exclude sunk costs from comparisons at later nodes (they are history, not futures).
• Include opportunity costs when a resource used on one branch cannot be used on another.
• Put information costs (study fees, pilot expenses) on the branch where they occur.

Document sources for probabilities: historical data, survey calibration, expert judgment, or Bayesian update from a prior (connecting to Unit 4). A tree with mystery probabilities is a fiction. At minimum, label them as judgment and plan sensitivity (Lesson 2).

Software helps at scale (tree add-ins, Python), but MBA fluency requires hand rollback on small trees. You should be able to explain every number on the slide without hiding behind a black box.

When the optimal decision is "close" (two branches within a few percent EMV), say so. Close decisions deserve sensitivity on key probabilities and payoffs. A study that costs $200,000 and raises EMV by $50,000 is a bad buy unless it reduces risk or satisfies a regulatory requirement. Closeness is strategic information, not a failure of analysis.

Drawing and validating trees in practice

A decision tree on a whiteboard and a decision tree in a board deck serve different purposes, but both must pass the same validation tests before anyone acts. Start by listing stages on a timeline: today, after the study, after launch quarter, after Year 1 demand resolves. For each stage, ask whether the firm makes a choice, nature reveals uncertainty, or the path ends. If you cannot assign one of those three labels, the stage is probably merged incorrectly or belongs in a different model.

Label every branch with incremental payoffs and conditional probabilities. A common validation exercise is the path audit: pick one complete path (for example, Study First → Favorable → Launch → Strong demand) and narrate it in plain English while tracing dollars. If the narrative requires information the decision-maker could not have at an earlier node, redraw the tree. Path audits catch the "clairvoyant manager" error faster than rollback algebra.

Software tools (PrecisionTree, TreePlan, Python libraries) help when trees grow beyond three stages. MBA fluency still requires that you can reconcile a small tree by hand. Before any tool run, solve a stripped version manually so you know which node should bind. When the software disagrees with your hand rollback, assume your diagram is wrong until proven otherwise. Typical software errors include double-counting study cost, forgetting to subtract investment on a branch, and entering probabilities that do not condition on the correct information set.

Peer review questions that strengthen trees before a capital committee sees them:

1. Do all chance-node probabilities on one node sum to 1.0?
2. Are payoffs in the same units (all NPV, all nominal Year-1 cash flow)?
3. Does each decision node compare only options available at that date?
4. Would any branch be chosen differently if sunk spending were removed from payoffs?
5. Is the EMV-optimal branch within a few percent of an alternative (flag for sensitivity)?

Finally, document rollback order in an appendix table: node name, node type, rolled-back value, optimal branch selected. Auditors and new analysts can reproduce the recommendation without reverse-engineering the diagram. That discipline connects directly to Lesson 5 communication: the board sees the optimal branch and the pruned alternatives, not a forest of unexplained numbers.

NovaWear, a fictional wearables division, must choose today between Launch Now (immediate $3 million tooling and marketing spend) and Study First ($200,000 market test, then launch or cancel depending on results). All payoffs below are incremental NPV in millions after investment on that path, consistent with the board's 12% hurdle rate already embedded in the figures.

Part A: Setup and tree structure

Stage 1 decision: Launch Now vs Study First.

If Launch Now: One chance node for Year-1 demand:

If Study First: Pay $0.2M now, then chance node for study result:

If favorable and NovaWear launches, updated demand probabilities apply:

If unfavorable, cancel with payoff $0 (no launch invest). Study cost $0.2M applies on the Study First branch regardless of outcome.

Probability checks: 0.55 + 0.45 = 1.00 ✓; 0.60 + 0.40 = 1.00 ✓; 0.75 + 0.25 = 1.00 ✓.

Part B: Rollback calculations

Launch Now (chance node):

EMV = 0.55(5.0) + 0.45(−2.0) = 2.75 − 0.90 = $1.85M

Study First, favorable subtree (launch is optimal vs cancel at $0):

EMV(launch | favorable) = 0.75(5.0) + 0.25(−2.0) = 3.75 − 0.50 = $3.25M

Study First, unfavorable subtree: Cancel → $0

Study First (chance node before study cost):

EMV = 0.60(3.25) + 0.40(0) = 1.95 − 0 = $1.95M

Subtract study cost on the branch: $1.95M − $0.2M = $1.75M

Root decision: Launch Now $1.85M vs Study First $1.75M → Launch Now wins on EMV by $0.10M.

Part C: Reconciliation and sensitivity hook

The study reduces EMV slightly in the base case because it costs $200,000 and still leads to launch after favorable news with residual weak-demand risk. It is not worthless (it avoids launch after bad news), but it is not worth buying on EMV alone here.

Sensitivity: If favorable study raised strong-demand probability to 0.85 (not 0.75):

EMV(launch | favorable) = 0.85(5.0) + 0.15(−2.0) = 4.25 − 0.30 = $4.05M

EMV(Study First) = 0.60(4.05) + 0.40(0) − 0.2 = 2.43 − 0.2 = $2.23M → Study First wins.

Break-even logic: managers should ask how informative the study must be to flip the decision. That threshold is often more actionable than the point estimate.

Part D: Managerial read

The CFO can tell the board: "We launch now on EMV, but the decision is fragile. A modest increase in study informativeness makes testing optimal. If marketing's real concern is avoiding a weak launch in public markets, the study buys reputational risk reduction not captured in EMV." The COO cares that Launch Now commits $3M immediately; Study First preserves flexibility worth discussing even when EMV is slightly lower.

Board questions: (1) What is our policy if EMV difference is only $100,000? (2) Do we have calibrated probabilities for post-study demand? (3) What is the worst-case cash need on Launch Now?

Harbor Labs faces a patent suit. Settle now for $8 million. Fight costs $2 million in legal fees and exposes Harbor to damages.

Part A: Fight subtree

Probability check: 0.25 + 0.35 + 0.40 = 1.00 ✓

Part B: EMV comparison

EMV(fight) = 0.25(13) + 0.35(1) + 0.40(−22) = 3.25 + 0.35 − 8.80 = −$5.2M

EMV(settle) = −$8.0M

EMV prefers fight (−5.2 > −8.0).

Part C: Risk policy read

Fight implies a 40% chance of −$22M, which may breach a loan covenant at −$15M cumulative cash impact. Settle is worse on average but bounded.

Part D: Managerial read

Document both recommendations: Analyst (EMV): fight. Treasurer (covenant): settle unless insurance or reserve funding is secured. The tree enables a risk policy conversation instead of a false "math says fight" headline.

Ridge Manufacturing chooses between Buy Machine ($500,000 today) and Lease ($150,000 per year for three years) to reduce unit costs. If Ridge buys, Year-1 demand is High (p = 0.50) with incremental profit $900,000 (after all costs except the $500,000 investment) or Low (p = 0.50) with incremental profit $200,000. Lease saves $280,000 per year in operating costs versus status quo for three years ($840,000 total benefit) minus lease payments ($450,000 total) = $390,000 net certain gain. Ridge is risk-neutral on EMV.

1. Draw the decision tree for Buy vs Lease (Buy has one chance node).
2. Compute EMV(Buy) and compare to Lease.
3. At what high-demand probability does Buy beat Lease? Show algebra.
4. If Ridge is risk-averse and refuses any outcome below −$100,000 net on Buy, does that change the decision?

Solution

1. Tree: Root decision → Buy (chance: High/Low) or Lease (terminal $390k certain).

2. EMV(Buy):

High net = 900 − 500 = $400k; Low net = 200 − 500 = −$300k

EMV = 0.5(400) + 0.5(−300) = 200 − 150 = $50k

Lease = $390k → Lease wins on EMV.

Check: 0.5 + 0.5 = 1.00 ✓

3. Break-even probability p for High:

p(400) + (1−p)(−300) = 390

400p − 300 + 300p = 390 → 700p = 690 → p ≈ 0.986

Buy beats Lease only if High demand is roughly 98.6% likely, far above 0.50.

4. Risk policy: Buy has a −$300k low branch, below −$100k floor. A risk-averse manager who applies a hard floor rejects Buy regardless of EMV. Lease remains chosen.

Explain in prose: EMV and risk policy can diverge when tail losses violate constraints. Ridge's board might publish a loss floor because a single bad demand year triggers layoffs and covenant review, even though long-run average profit favors buying when demand is high. The tree makes that tension visible: Lease eliminates the −$300k branch entirely at the cost of giving up $400k upside in the High state. A complete recommendation states both the EMV winner (Lease) and the policy rationale (floor), so operations does not interpret Lease as "demand will be low."

A biotech firm can Run Phase II ($4M) now or License Early for a certain $1.5M. If Phase II runs, success (p = 0.35) leads to NPV $20M; failure (p = 0.65) leads to NPV $0. Discounting is already in the NPV figures.

1. Compute EMV for Phase II vs License.
2. Should the firm run Phase II on EMV?
3. If success probability rises to 0.45, does the decision flip?

Solution

1. EMV(Phase II) = 0.35(20) + 0.65(0) − 4 = 7 − 4 = $3.0M (subtract $4M cost on path)

License = $1.5M

2. Phase II wins on EMV ($3.0M > $1.5M).

3. At p = 0.45: EMV = 0.45(20) − 4 = 9 − 4 = $5.0M, still beats license. Decision does not flip; margin widens.

Check: 0.35 + 0.65 = 1.00 ✓

Key takeaways

• Decision trees map sequential choices and uncertainties; solve by rollback from terminal payoffs.
• At chance nodes use EMV; at decision nodes choose the best rolled-back branch unless risk policy overrides.
• Only incremental, future payoffs belong on branches; sunk costs do not decide later forks.
• Compare root EMV across first-stage alternatives and stress-test close calls with probability sensitivity.
• When tail risk threatens survival, report EMV and policy-based choice side by side.

1. Sketch a two-stage tree for a real decision at your firm (launch, hire, lease) with dollar payoffs and at least one chance node.
2. What probability threshold would flip your optimal first-stage branch?
3. Continue to Lesson 2: Sensitivity and Scenario Analysis.