Two Different Questions
A predictive model answers a question about an outcome: for example, how likely is a claim to exceed a stated amount during a defined period? An action recommendation answers a different question: given the available choices and their consequences, what should be done? The forecast can inform the recommendation, but it does not determine it by itself.
This distinction matters whenever an insurer, service team, or other decision-maker can take an action that changes costs or outcomes. The decision depends not only on how likely an event is, but also on what the event would cost, what each action costs, and how outcomes differ under the available actions. These quantities should be stated in the context of the decision rather than hidden inside a risk score.
Here, \(X=x\) denotes the information available at the decision time for the case under consideration. A forecast such as \(P(Y=1\mid X=x)=0.30\) describes the chance of the event under the conditions represented by that forecast. It is not automatically the chance under every possible action. If taking an action can change the outcome, the comparison must account for the outcomes and costs associated with each action.
From Forecast to Choice
Suppose the feasible actions are \(a_1,\ldots,a_k\), and let \(C(a_j,Y)\) be the cost that results when action \(a_j\) is taken and outcome \(Y\) occurs. If the relevant probabilities under each action are available, the expected cost of action \(a_j\), for a case with information \(X=x\), is
For a cost-minimizing decision, choose an action with the smallest expected cost among the feasible choices. A different objective, such as maximizing expected net benefit, can be handled similarly by comparing expected benefits minus costs. In either case, the action set and the objective must be specified. A prediction alone does not state them.
For a simple binary event, suppose not acting leads to a loss \(L\) if the event occurs and zero loss otherwise. Suppose an intervention costs \(c\), prevents the event completely, and has no other costs or consequences. If the event probability without intervention is \(p\), the expected cost of not acting is \(pL\), while the expected cost of intervening is \(c\). Therefore, intervention is preferred when \(c<pL\).
The threshold follows by comparing the two expected costs: intervention is preferred when \(c<pL\); dividing both sides by the positive amount \(L\) gives \(p>c/L\). This result applies only under the stated assumptions: the intervention prevents the event, costs \(c\), and has no additional effects. If it only reduces the event probability, carries a cost when the event occurs, or creates other benefits or harms, the expected costs must include those terms.
Worked Example: A High Event Risk Does Not Guarantee Outreach
A hypothetical insurer predicts whether a policyholder will have a preventable service lapse during the next year. For one policyholder, the predicted probability without outreach is 0.20. A one-time outreach costs $120 and, based on appropriate action-specific evidence, is expected to reduce the probability of a lapse to 0.10. Each lapse costs the insurer $1,000. Assume the outreach cost is incurred whether or not a lapse occurs, and there are no other costs or benefits. Should outreach be recommended?
Method. Compare the expected total cost under each action, using the lapse probability appropriate to that action.
No outreach. The expected cost is \(0.20\times \$1{,}000=\$200\).
Outreach. The expected cost is the outreach cost plus the expected lapse cost: \(\$120+(0.10\times \$1{,}000)=\$120+\$100=\$220\).
Because \(\$200<\$220\), the cost-minimizing choice under these assumptions is no outreach. The predicted risk of 0.20 is useful, but it does not by itself justify outreach; the action’s cost and its expected effect on the outcome also matter. The calculation checks: the two action-specific expected costs are $200 and $220, so outreach is $20 more expensive in expectation.
Common mistake avoided: comparing the outreach cost of $120 with the expected lapse cost of $200 under no outreach, then recommending outreach because $120 is smaller. That comparison ignores the remaining expected lapse cost after outreach. The correct comparison is $200 against $220.
When Several Actions Are Available
Many decisions are not a choice between acting and doing nothing. There may be several service levels, inspection options, or other feasible actions. In that case, compare the expected cost or benefit for every relevant option. Selecting an action because it is more intensive, or because it targets the highest-risk cases, is not enough: it must perform better on the stated objective.
The forecast and the recommendation also have different information needs. A prediction can be useful even when no action is available, because it describes what is expected to happen. A recommendation requires, in addition, a defined set of feasible choices and a way to compare their consequences. If actions may change the outcome, those consequences should be action-specific. A risk probability observed under one operating policy should not automatically be treated as the outcome probability under a different policy.
Worked Example: Choosing Among Three Service Levels
A hypothetical insurer is considering service for accounts at risk of a billing interruption during the next month. For one account, a model estimates the following probabilities under three options. The probabilities are assumed to come from evidence suitable for comparing these options. A billing interruption costs $1,000. Standard service costs $80 per account, and intensive service costs $180; no service has no service cost. Each service cost is paid regardless of whether an interruption occurs. Which option minimizes expected total cost?
Method. For each option, add its service cost to the event probability under that option multiplied by the event cost.
No service. The expected cost is \(0+(0.30\times \$1{,}000)=\$300\).
Standard service. The expected cost is \(\$80+(0.18\times \$1{,}000)=\$80+\$180=\$260\).
Intensive service. The expected cost is \(\$180+(0.10\times \$1{,}000)=\$180+\$100=\$280\).
The minimum is $260, so standard service is preferred among these three options. Although intensive service has the lowest event probability, its greater cost makes its expected total cost higher than standard service’s. The check is \( \$260<\$280<\$300\), so the selected option is indeed the least costly under the given figures.
Common mistake avoided: choosing intensive service because it produces the lowest event probability. The objective is to minimize total expected cost, not the event probability alone.
Prediction Quality and Decision Quality
A forecast can be accurate for its stated target and still be insufficient to choose an action. A probability of an event under usual conditions does not tell the decision-maker whether an intervention will change that probability. Nor does it specify the value of preventing the event, the intervention’s cost, or the existence of other options. Those are inputs to the decision, not properties of the risk score.
The reverse distinction matters too. An action that is optimal under a particular cost structure does not make its underlying prediction universally useful. Changing the cost of an event, the cost of an intervention, or the set of feasible actions can change the preferred action even when the predicted risks stay exactly the same. Thus, a recommendation should be tied to the decision context in which its costs and consequences were compared.
When an action can affect the outcome, a recommendation requires information about what would be expected under the available alternatives. For example, a probability estimated among cases that received outreach may not, on its own, show what the same cases’ outcomes would have been without outreach. A defensible comparison needs suitable evidence or explicit assumptions for the action-specific probabilities. If those probabilities are not known, the analysis should say so rather than present a risk ranking as a proven action recommendation.
A practical framing sequence is to state the predicted outcome and its time window, identify what information is available at the decision time, list the feasible actions, and define the consequence used to compare them. The prediction and the action-specific comparison can then be reported separately. This makes clear which statements are forecasts and which depend on costs, preferences, or evidence about action effects.
Worked Example: Inspection With a Partial Reduction in Loss Probability
A hypothetical property insurer must decide whether to order a remote inspection for a building. Without inspection, the probability of a specified damage event during the coming year is 0.12. An inspection costs $70 and is expected to reduce the event probability to 0.04 by enabling a repair. The event costs the insurer $800, and the inspection cost is incurred whether or not damage occurs. Assume there are no other consequences. Should the insurer order the inspection?
Method. Compare expected total costs under no inspection and inspection. The inspection does not eliminate the event, so its residual expected damage cost must be included.
No inspection. Expected cost is \(0.12\times \$800=\$96\).
Inspection. Expected cost is \(\$70+(0.04\times \$800)=\$70+\$32=\$102\).
Since \(\$96<\$102\), no inspection is preferred under the stated assumptions. The inspection reduces expected damage cost by \(\$96-\$32=\$64\), but it costs $70; its net expected cost increase is \(\$70-\$64=\$6\). The check agrees with the direct comparison: $102 is $6 greater than $96.
Common mistake avoided: comparing the $70 inspection cost only with the $96 expected damage cost before inspection. That omits the $32 expected damage cost that remains after inspection.
Exam Traps and a Reliable Distinction
A useful written answer keeps the forecast separate from the choice. For instance, “The estimated probability of an interruption is 0.30 without service” is a prediction statement. “Standard service minimizes expected cost among the three options given the stated probabilities and costs” is a decision statement. The second statement depends on the specified action set, objective, and action-specific probabilities.
- Equating high risk with high action value. A high event probability may still not justify an expensive action, especially if the action has little effect.
- Using one probability for every action. If an action changes the outcome, compare probabilities or outcome distributions that correspond to each action.
- Leaving costs out of the choice. Minimizing event probability is not the same as minimizing expected cost.
- Claiming more than the evidence supports. A forecast under observed conditions does not, by itself, establish the outcome under a different action.
- Forgetting the residual loss. If an intervention reduces risk rather than eliminating it, include the expected loss that remains after the intervention.
Check Your Understanding
For each problem, distinguish the forecast from the action comparison and use the stated assumptions.
- A hypothetical insurer estimates a 0.08 probability of a $2,000 loss without a device. The device costs $100 and reduces the probability to 0.02. It has no other consequences. Compare expected costs and state the preferred option. (Answer: No device costs $160; device costs $100 + $40 = $140, so choose the device.)
- A service team estimates a 0.25 probability of an event costing $600 without service. Basic service costs $50 and reduces the probability to 0.18. What are the expected costs, and which option is less costly? (Answer: No service costs $150; basic service costs $50 + $108 = $158, so choose no service.)
- For a hypothetical account, the event costs $900. No intervention leaves event probability at 0.20; option A costs $40 and reduces it to 0.12; option B costs $100 and reduces it to 0.05. Which option minimizes expected cost? (Answer: No intervention costs $180; A costs $40 + $108 = $148; B costs $100 + $45 = $145, so choose B.)
- An intervention fully prevents an event that would cost $500. It costs $75, with no other consequences. What is the event-probability threshold above which intervention has lower expected cost? If the predicted probability is 0.12, what is the preferred action? (Answer: The threshold is \(75/500=0.15\); at 0.12 the expected loss without action is $60, so choose no intervention.)
- A hypothetical inspection costs $90. Without inspection, the event probability is 0.16; after inspection it is 0.07. The event costs $1,200. Calculate both expected costs and explain why the lower event probability does not settle the decision by itself. (Answer: No inspection costs \(0.16\times \$1{,}200=\$192\); inspection costs \(\$90+0.07\times \$1{,}200=\$174\), so inspection is preferred because its expected total cost is lower.)