Map sequential investment commitments in a decision tree
17 min read
Lesson objectives
- Build a tree separating decisions, chance outcomes and learning costs at their commitment points.
- Compare information value and actual flexibility with liquidity constraints and error probabilities.
Please sign in to use all the tools
Once you sign in, the reading tools are yours: larger or smaller text, typeface, line spacing, focus mode, highlighting and bold, and notes saved to your account and gathered in My study.
A portfolio team proposes either immediate expansion of a new service or a small test followed by an expansion decision. One presentation says testing reduces risk but adds expansion costs to every test branch, even when the test would lead to stopping. Another values expansion as though the decision-maker already knew its final outcome. Both obscure the order of information and commitment. In this lesson, a decision tree represents what the team knows at each commitment, what remains uncertain and what it can do after evidence arrives. The aim is a simple, traceable calculation rather than an elaborate diagram suggesting unsupported precision. The output is an executable conditional policy: what is paid now, what happens after each possible result, and who is authorised to decide?
Define the decision, timing and monetary unit first. This training case uses thousands of riyals and net values over one short horizon. It assumes no material timing difference requiring discounting within the example; a real decision needs a common valuation date when timing differs. Separate choosing expansion from the future state of service success. The board can choose expansion but cannot choose success. Place the test result between the decisions only if it arrives before the expansion commitment and the board can respond by changing its choice. A test received after an irreversible contract may assist operations, but does not provide the investment flexibility claimed for it.
Read the tree from left to right to explain the sequence of commitments, then calculate from right to left to identify the policy. At a chance node, multiply each outcome by its probability conditional on the preceding information and add the products. At a decision node, compare choices actually available then, subject to stated constraints, and select according to the declared rule. Expected value is the explicit calculation criterion in this example, with possible loss and liquidity limits shown separately. An organisation unwilling to accept a particular loss may reject a positive-expected-value option. An average does not remove a risk limit or replace the board's authority to accept exposure.
Add a test costing 40 that produces a positive or negative signal. Each signal occurs with probability one half in the supplied model. After a positive signal, success probability becomes 80%; after a negative signal, it becomes 20%. These conditional probabilities differ from the probability of observing a signal. Do not multiply 80% by 800 while omitting failure after a positive signal. A positive signal is not guaranteed success: a 20% failure probability remains in that branch. Nor does a negative signal guarantee failure. It changes the estimate, and limited expansion could remain appropriate if a different option existed outside this comparison.
Ask where probabilities come from rather than treating the sponsor's confidence as evidence. They may derive from a relevant earlier trial, a model using documented inputs, or expert assessment with explicit ranges. Different populations and operating conditions weaken transferability. The Green Book's general approach covers alternatives, risk and uncertainty; it does not provide success probabilities for this investment. The MIT reference on perfect information helps distinguish the value of knowledge obtained before a decision, rather than proving local test accuracy. Where evidence is weak, show how the result changes when signal discrimination deteriorates instead of hiding estimation behind many decimal places.
Count the test fee once at the point where it is paid. If 40 is paid before observing the signal, every tested path bears it, whether expansion follows or not. Deduct it before averaging the two branches, or deduct it within each branch and then average; both methods agree when applied consistently. Deducting before the average and again inside each branch double-counts the fee. Conversely, omitting it from the stopping path makes testing appear free whenever expansion is rejected. Separating learning expenditure from expansion expenditure makes errors easier to inspect and reveals how much money is exposed to information that may rationally lead to abstention.
Distinguish testing from hypothetical perfect information. Perfect information reveals whether the service would succeed or fail before expansion is chosen, without error in this theoretical comparison. The policy then expands in the successful half of cases to receive 800, and stops in the failing half for zero. Its value before an information price is 400, compared with the best uninformed decision of 200. The gross expected value of perfect information is therefore 200. This is neither a realistic test price nor a promise that complete knowledge is obtainable. It is a reference ceiling within the same model. Imperfect testing cannot exceed that ceiling while the underlying choices, information and values remain unchanged.
Inspect contractual commitments between stages. Testing may require an initial payment of 40, followed by another 300 if expansion is selected. Net outcomes of 800 and minus 400 already include that 300 in the return calculation; deducting it again would count investment cost twice. Liquidity analysis still needs the separate 300 payment on an expansion path. Retain two linked views: a net economic result for each path and a cash schedule showing maximum funding commitments and timing. Net outcome is not the same as the amount of cash needed at the decision point. A positive-expected-value policy can be infeasible under a funding ceiling before benefits arrive.
Test sensitivity to weaker discrimination. If conditional success becomes 60% after positive and 40% after negative, with equally likely signals, the original 50% success probability remains consistent. Expansion is then worth 320 after positive and 80 after negative, both above zero. The policy expands in both branches, yielding 200 before the fee and 160 after it. Here the signal does not change the decision among current options, so it adds cost without improving expected outcome. It might provide a separate operational benefit, but that benefit must be explicitly measured and included rather than hidden under an undefined risk-reduction label.
Waiting may also change value. A six-month test could miss a demand season or supplier opportunity, so outcomes of 800 and minus 400 may no longer apply. Calculate the results of delayed expansion instead of automatically borrowing immediate-expansion values. If the test period permits redesign, distinguish learning followed by redesign from learning followed by expansion of the same design. Each new version requires its own cost, timing and supported probabilities. This prevents the test from being credited with options that do not exist, or with eliminating every loss after a positive signal.
For a tree with several stages, check probabilities at each chance node rather than adding every probability written on the diagram. Positive and negative sum to one; success and failure within the positive branch also sum to one, as do those within the negative branch. A complete path probability is the product of successive conditional probabilities. In the basic model, a positive signal followed by failure has probability one half times one fifth, or 10% of all cases. That differs from the 20% failure probability within positive cases alone. Name each population to avoid presenting a conditional percentage as the overall loss rate to the board.
In the board note, present the chosen policy and its adverse path rather than only its average. Testing avoids expansion after negative, but a false positive can still lead to expansion and a loss of 400 plus the fee of 40. That path occurs with probability 10% in this model; testing therefore does not eliminate loss. Also show maximum cash needs, the stopping condition and responsibility for reviewing probability validity. A contractual change preventing stopping, or a higher fee, requires reassessment of the affected choice. Retain the original decision version so reviewers can distinguish what was known at commitment from knowledge available only afterward.
Transfer the skill to an investment containing two real commitments separated by expected information. Draw choices genuinely available at each point, then remove branches the organisation cannot select contractually. Review cost locations and probabilities with finance and the evidence owner, and identify assumptions that change the recommendation. If learning does not alter the decision, explain that and quantify any separate purpose. A tree's value lies in exposing commitment order, learning and choice limits, rather than node count or a large claimed information value. Retain a recalculable model and a recommendation able to answer a simple question: what will we actually do if the result is the one we dislike?
Derive the decision threshold instead of trying arbitrary probabilities. Let success probability be p. Expansion's expected outcome is 800 times p plus minus 400 times one minus p. Rearranging gives 1,200 times p minus 400, which equals zero when p is one third. Above one third, expansion beats the zero option under expected value alone; below it, abstention is preferred. This threshold depends on outcomes of 800 and minus 400 and an abstention value of zero. Different returns, a limited-expansion option or a loss constraint change it. One third must not be memorised as a universal investment rule.
Also investigate whether a supplier has transferred the probability of success after a positive result from another population without adjusting the original success rate. Discriminatory performance alone does not determine what a positive result means. We need prior success prevalence, the positive rate among successes and the positive rate among failures. An expected-count table makes the relationship visible without incorrectly assuming independence. The same test can have weaker predictive value when success is rare in the new service. Request the supplier's measurement population and outcome-verification method, then recalculate for local conditions rather than importing an attractive headline accuracy figure.
Lesson summary
The lesson evaluates an investment policy by working backward through a tree. It distinguishes signal probabilities from conditional success, and gross information value from value after the fee. Stopping rights, cash requirements and population consistency are checked before expansion is recommended.