From Prediction to Decision

Introduction

Predictive analytics estimates what is likely to happen; decision-making determines what should be done. This distinction is central to business analytics. A model may predict that a customer is likely to leave, a payment is likely to default, or demand is likely to exceed inventory. None of these predictions automatically specifies the best action. The organisation must compare available actions, consequences, costs, benefits, constraints, and risks.

The movement from prediction to decision therefore requires a decision framework. Such a framework connects model outputs to operational choices through decision rules, expected value, cost–benefit analysis, scenario analysis, sensitivity analysis, optimisation, and informed human judgement. The quality of a decision depends not only on predictive accuracy but also on whether the selected action creates value in the real business environment.

Learning Objectives

By the end of this chapter, you should be able to:

• distinguish prediction, prescription, and decision-making;
• translate probabilities and forecasts into explicit decision rules;
• calculate expected monetary value and compare alternatives;
• incorporate costs, benefits, constraints, and uncertainty into decisions;
• use scenario and sensitivity analysis to test decision robustness;
• explain how optimisation supports resource allocation;
• identify appropriate and inappropriate uses of human judgement;
• evaluate ethical, governance, and implementation issues in data-driven decisions.

1. Prediction Is Not a Decision

A prediction describes a future state or outcome. For example, a model may estimate a 0.70 probability that a loan applicant will default within twelve months. A decision is an action, such as approving, rejecting, or referring the application. A decision rule specifies how predictions are converted into actions.

A useful distinction is:

A predictive model is only one component of a decision system. Other components include the objective, feasible actions, costs, benefits, constraints, timing, accountability, and monitoring process.

For classification, the output may be a probability p = P(Y = 1|X), where Y is the outcome and X represents observed information. The probability is not itself a command. A threshold or economic rule is needed. For example, a bank might investigate applications with p ≥ 0.35, while another institution might use p ≥ 0.60 because its investigation capacity and risk tolerance differ.

2. Decision Rules and Thresholds

A decision rule maps information to an action. In its simplest form:

Here, p is the predicted probability of an event and t is the decision threshold. The threshold should not be selected solely because it produces a convenient accuracy score. It should reflect the relative consequences of errors and the value of intervention.

Consider a fraud-screening system. A transaction flagged as fraudulent but actually legitimate is a false positive. A transaction not flagged but actually fraudulent is a false negative. If the cost of a missed fraud is much greater than the cost of reviewing a legitimate transaction, a lower threshold may be justified. If manual review capacity is limited, the organisation may need a higher threshold or a ranking system that reviews only the highest-risk cases.

For a binary decision with an intervention and no intervention, an economically based rule compares expected values. Let:

• p = probability that the target event occurs;
• B = benefit of intervention when the event occurs;
• C_I = cost of intervention;
• C_N = cost of not intervening when the event occurs.

The expected value of intervention is:

The expected value of no intervention, expressed as a cost, is:

Intervention is preferred when EV(I) > EV(N). Thus:

Rearranging gives:

This threshold assumes that the intervention benefit, intervention cost, and avoided cost are appropriately measured, additive, and relevant to the same decision period. It also assumes that the model probability is calibrated: a group predicted at 0.40 should experience the event approximately 40% of the time.

Worked Example: Customer Retention

A retailer predicts whether a customer will make no further purchase during the next six months. A targeted retention offer costs £8 per customer. If the customer would otherwise leave, the offer produces £30 of contribution margin through retained purchases. If the offer is not made and the customer leaves, the expected lost contribution is £50. The threshold for intervention is:

A customer with a predicted departure probability of 0.18 should receive the offer under this simplified rule. The expected value of the offer is:

The expected value of no intervention is:

Although intervention has a negative expected value relative to an ideal zero-cost baseline, it is better than no intervention because −£2.60 is greater than −£9.00. The comparison must therefore be between feasible alternatives, not between an action and an imaginary outcome with no cost or risk.

3. Expected Value and Cost–Benefit Analysis

Expected value combines possible outcomes with their probabilities. For an action a with outcomes 1, 2, ..., n, the expected monetary value is:

where p_i is the probability of outcome i, V_i is its monetary value, and ∑p_i = 1. Values should be measured consistently, such as net contribution, profit, or total cost avoided. Probabilities should represent the relevant population and time horizon.

Worked Example: Inventory Decision

A retailer is deciding whether to order 1,000 or 1,500 units of a seasonal product. Demand may be high with probability 0.60 or low with probability 0.40. The estimated profits are:

For 1,000 units:

For 1,500 units:

The 1,500-unit order has the higher expected monetary value by £400. However, this conclusion depends on the probabilities and profit estimates. It may be inappropriate if the organisation cannot tolerate the £20,000 difference between the high- and low-demand outcomes, or if the low-demand probability is poorly estimated.

Expected value is a rational comparison rule when probabilities and monetary values are credible and the decision-maker is willing to evaluate uncertain outcomes by their averages. It does not mean that the expected outcome will occur. Nor does it automatically account for risk aversion, cash constraints, reputational damage, or non-monetary consequences.

Cost–Benefit Analysis

Cost–benefit analysis identifies all material incremental costs and benefits associated with an option. Typical categories include:

• direct operating costs;
• technology and implementation costs;
• staff time and training;
• expected revenue or contribution;
• avoided losses;
• customer and employee effects;
• regulatory, legal, or reputational consequences;
• opportunity costs.

Only incremental effects should normally be included. A cost that occurs regardless of the decision should not be treated as a benefit of one option. Fixed costs may matter for affordability, but they do not necessarily affect the comparison between alternatives.

If cash flows occur across multiple periods, future values can be discounted. For a single future cash flow:

where CF_t is the cash flow at time t, r is the discount rate per period, and PV is present value. This formula assumes a constant per-period rate and cash flow measured at the end of each period. For several cash flows, present values are summed. A positive net present value indicates that discounted benefits exceed discounted costs relative to the chosen baseline.

4. Decision Trees and Value of Information

A decision tree represents decisions, uncertain events, probabilities, and outcomes in chronological order. Decision nodes show choices; chance nodes show uncertain events. The standard procedure is to calculate values from right to left, multiplying each outcome by its probability and then selecting the best action at each decision node.

Figure: A simple decision tree for an inventory-ordering decision.

The expected value of perfect information (EVPI) estimates the maximum amount worth paying to know the future state with certainty. It is calculated as:

EVPI is an upper bound, not a guarantee that information will produce that amount of benefit. Real information is imperfect and costs money. Its value is also reduced by delays, implementation limitations, and the organisation’s ability to act on it.

5. Scenario Analysis and Sensitivity Analysis

A single expected value can conceal important assumptions. Scenario analysis examines coherent combinations of assumptions, such as:

• optimistic, base, and pessimistic demand;
• high, central, and low conversion rates;
• alternative inflation or interest-rate conditions;
• different levels of model accuracy or intervention effectiveness.

Scenarios should be plausible and internally consistent. They are not forecasts with false precision. Their purpose is to explore how a decision behaves under different conditions.

Sensitivity analysis changes one input at a time, or systematically varies several inputs, to identify which assumptions most affect the result. In the inventory example, the difference between ordering 1,500 and 1,000 units is:

If the high-demand probability is q, the larger order is preferred when:

Simplifying:

Thus, the larger order is preferred when the probability of high demand exceeds approximately 57.14%. Since the estimated probability is 60%, the decision is favourable but not highly robust. A modest change in the demand estimate could reverse it.

A break-even analysis identifies the value of an input at which two alternatives have equal value. Break-even points are especially useful for communicating model dependence to managers. Tornado charts, one-way data tables, and stress tests can support this analysis, although the underlying assumptions must remain visible.

6. Optimisation and Resource Allocation

Optimisation selects the best feasible combination of decisions according to an objective function and constraints. Prediction supplies estimates such as demand, probability of response, or expected processing time; optimisation uses these estimates to allocate scarce resources.

A linear programming model can be expressed as:

subject to:

Here, x_j is the quantity of activity or decision j, c_j is its contribution, a_ij is the amount of resource i used by activity j, b_i is the available amount of resource i, and Z is total contribution. Linear programming assumes that contributions and resource use are linear, divisible where required, and sufficiently known for the planning problem.

For example, a marketing team may choose campaigns subject to a £50,000 budget and a 400-hour staffing limit. A predictive model estimates expected conversions for each campaign. Optimisation then determines the campaign mix that maximises expected contribution, subject to budget, staffing, audience, and compliance constraints.

A critical danger is optimisation against inaccurate or biased predictions. If a model systematically overestimates one campaign type, the optimiser may allocate too much budget to it. The solution may be mathematically optimal for the model but poor in reality. Robust optimisation, constraints on exposure, minimum exploration, and periodic recalibration can reduce this risk.

7. Human Judgement Under Risk and Uncertainty

Risk refers to situations in which relevant outcomes and probabilities can be reasonably specified. Uncertainty refers to situations in which probabilities, outcomes, or causal relationships are poorly known. The distinction is not absolute, but it helps determine how much confidence to place in expected values.

Human judgement remains important when:

• the data do not represent a major structural change;
• the decision involves unmeasured qualitative factors;
• consequences are irreversible or ethically sensitive;
• the model is operating outside its training population;
• legal or professional accountability requires review;
• stakeholders possess contextual knowledge unavailable in the data.

Judgement can also introduce inconsistency, anchoring, overconfidence, confirmation bias, and selective override. The appropriate response is not to remove people automatically, but to design controlled human oversight. Useful practices include documenting override reasons, separating evidence from preference, using predefined escalation rules, conducting calibration training, and auditing outcomes for different groups.

A human-in-the-loop process should specify whether the person is reviewing, approving, overriding, or merely accountable for the system. “Human oversight” is not meaningful if the reviewer lacks time, information, authority, or understanding of the model.

8. Applied Case: Predictive Maintenance in a Distribution Centre

The following is an illustrative case.

Background

A distribution centre operates 120 automated conveyor motors. Unexpected motor failure causes an average disruption cost of £18,000, including lost throughput and emergency repair. A predictive model estimates the probability of failure within the next 30 days for each motor.

Problem

The engineering team can perform preventive maintenance on only 15 motors per month. Each maintenance intervention costs £1,200. Maintenance prevents 80% of predicted failures, and the model’s probabilities are reasonably calibrated. Management must decide which motors to service.

Analysis

For a motor with failure probability p, the expected cost of no maintenance is:

The expected cost of maintenance is:

Maintenance is preferred when:

The economic break-even probability is therefore approximately 8.33%. Motors above this probability have positive expected savings from maintenance, assuming the stated values are accurate.

Approach and Solution

The centre ranks motors by expected net savings, not merely by a generic risk score. For each motor:

For a motor with p = 0.20:

Expected saving equals £1,680. The team selects the 15 motors with the greatest expected net savings, while checking safety-critical equipment and operational constraints.

Outcome

After three months, the centre compares predicted and actual failure rates, maintenance effects, false alarms, and missed failures. It also checks whether the 15-motor capacity constraint caused high-risk equipment to remain untreated. If the model performs poorly for a particular motor type, the model and the decision rule are reviewed rather than blindly expanded.

Learning Points

The case illustrates that a probability becomes useful only when linked to an economic consequence and a feasible action. Ranking by probability alone may select a motor with high risk but low avoidable loss, while expected net saving incorporates both risk and impact. Capacity, safety, calibration, and post-decision monitoring are essential parts of the solution.

9. Governance, Ethics, and Implementation

Data-driven decisions can distribute opportunities and burdens. Credit, recruitment, insurance, healthcare, and fraud decisions may affect legally protected or vulnerable groups. A technically accurate model can still produce unacceptable outcomes if data reflect historical discrimination, if proxy variables reproduce protected characteristics, or if affected individuals cannot challenge errors.

Good governance includes:

• a clearly documented purpose and decision owner;
• data quality, provenance, and access controls;
• validation on relevant and sufficiently recent populations;
• calibration and subgroup performance analysis;
• monitoring for drift and changing base rates;
• explainable decision logic appropriate to the audience;
• appeal, correction, and escalation mechanisms;
• records of model versions, thresholds, overrides, and outcomes;
• proportionate privacy and security safeguards.

The threshold itself should be governed. Changing a threshold changes the number and type of people or transactions affected, as well as false-positive and false-negative rates. Threshold changes should therefore be tested, approved, documented, and monitored.

Contemporary systems increasingly combine predictive models with generative AI, optimisation engines, and automated workflows. These technologies can accelerate analysis, but they do not eliminate the need for reliable data, economic validation, human accountability, and controls against fabricated outputs or inappropriate automation. An AI-generated recommendation should be treated as an input to a governed decision process, not as an unquestionable decision-maker.

10. Best-Practice Decision Workflow

A practical workflow is:

1. Define the decision, objective, time horizon, and baseline.
2. List feasible actions and constraints.
3. Identify the relevant predicted outcomes and validate their meaning.
4. Convert predictions into costs, benefits, or utility values.
5. Establish a decision rule or optimisation model.
6. Calculate expected values using consistent units and assumptions.
7. Test scenarios, break-even points, and sensitive assumptions.
8. Assess risk, uncertainty, fairness, privacy, and operational feasibility.
9. Assign human accountability and escalation procedures.
10. Implement on a controlled basis and monitor outcomes.
11. Compare realised outcomes with predictions and update the process.

The baseline must be explicit. A new decision should be compared with what would happen without the proposed intervention, not with an unrealistic zero-cost alternative. Similarly, benefits should be incremental and should not be counted twice.

Challenges and Limitations

First, model probabilities may be poorly calibrated even when rankings are useful. A ranking model can identify higher-risk cases without accurately estimating the magnitude of risk. Second, causal effects are not the same as predicted outcomes. A customer predicted to leave may not be saved by an offer; prediction identifies risk, whereas causal analysis estimates treatment impact.

Third, data and conditions change. Demand shocks, policy changes, new competitors, and altered customer behaviour can invalidate historical relationships. Fourth, expected value can conceal tail risks and unequal consequences. Fifth, optimisation may amplify model errors and create concentrated exposure. Finally, implementation costs, staff resistance, privacy restrictions, and limited intervention capacity may make a theoretically attractive decision impractical.

Chapter Summary

Prediction estimates likely outcomes, while decision-making selects actions under objectives, constraints, costs, benefits, and uncertainty. Decision rules convert probabilities into actions, and economically justified thresholds depend on the relative consequences of intervention and non-intervention. Expected monetary value compares uncertain alternatives by probability-weighted outcomes, but it requires credible probabilities, consistent monetary measures, and an appropriate baseline.

Decision trees clarify sequential choices and uncertainty, while the value of information indicates how much improved information could be worth. Scenario analysis explores coherent alternative futures; sensitivity and break-even analysis reveal whether a conclusion is robust. Optimisation allocates scarce resources but can magnify prediction errors if its inputs are biased or inaccurate. Human judgement is valuable for context, ethics, and unusual cases, but should be structured and audited. Effective decision systems combine predictive evidence, economic reasoning, operational constraints, governance, and continuous monitoring.

Key Terms

Decision rule: A specified mapping from information or predictions to an action.

Decision threshold: A cutoff used to determine which action is selected.

Expected monetary value: The probability-weighted average monetary value of an action’s possible outcomes.

Cost–benefit analysis: Systematic comparison of incremental costs and benefits of alternatives.

Baseline: The outcome expected under the current or reference course of action.

Calibration: Agreement between predicted probabilities and observed event frequencies.

False positive: An intervention or alert generated when the target condition is absent.

False negative: Failure to identify a case in which the target condition is present.

Decision tree: A representation of decisions, uncertain events, probabilities, and outcomes.

Expected value of perfect information: The maximum value of knowing uncertain future states before acting.

Scenario analysis: Examination of decisions under coherent alternative assumptions or futures.

Sensitivity analysis: Investigation of how results change when input assumptions vary.

Break-even point: The input value at which competing alternatives have equal value.

Optimisation: Selection of the best feasible solution according to an objective and constraints.

Objective function: The mathematical expression representing what an optimisation model seeks to maximise or minimise.

Constraint: A limit or requirement restricting feasible decisions.

Risk: Uncertainty for which outcomes and probabilities can be reasonably specified.

Model drift: Decline in model relevance or performance as data and relationships change.

Human-in-the-loop: A decision process in which people review, approve, or can override automated outputs.

Expected treatment effect: The estimated change caused by an intervention, distinct from predicted risk.

Review Questions

Short Answer Questions

1. Why is a predictive probability not, by itself, a decision?
2. Define calibration and explain why it matters for threshold selection.
3. What is the difference between scenario analysis and sensitivity analysis?
4. State two limitations of expected monetary value.
5. What is the role of a constraint in an optimisation model?

Long Answer Questions

1. Explain how false-positive and false-negative costs influence a classification threshold. Use a business example.
2. Describe the process of converting a predictive model into a governed operational decision system.
3. Explain how decision trees and expected value can support decisions under uncertainty.
4. Discuss the benefits and risks of using optimisation after predictive modelling.
5. Evaluate the role of human judgement in high-impact automated decisions.

Analytical / Application Questions

1. An intervention costs £15, avoids a £90 loss if an event occurs, and otherwise has no benefit. Derive the minimum event probability at which intervention is preferred, assuming no other costs.
2. A firm must choose between two projects with uncertain outcomes. Design a table for expected monetary value, then identify three assumptions that should be tested through sensitivity analysis.
3. A predictive model ranks customers by default risk, but investigation capacity is limited to 5% of applications. Recommend a decision process that combines ranking, economic value, fairness monitoring, and human review.

Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes a prediction from a decision?

A. A prediction always has a higher monetary value than a decision.

B. A prediction estimates an outcome, whereas a decision selects an action.

C. A prediction is always causal, whereas a decision is always descriptive.

D. A prediction can be made only after an action has occurred.

Correct Answer: B. A prediction estimates an outcome, whereas a decision selects an action.

Question 2

An intervention costs £10, provides a £30 benefit when an event occurs, and avoids a £20 loss that would otherwise occur when the event happens. What is the minimum event probability for intervention to be preferred?

A. 0.10

B. 0.20

C. 0.25

D. 0.50

Correct Answer: B. 0.20

Question 3

What does calibration mean in a probabilistic prediction model?

A. The model has the highest possible accuracy.

B. Predicted probabilities correspond reasonably to observed frequencies.

C. The model uses only numerical variables.

D. The model produces identical predictions for all groups.

Correct Answer: B. Predicted probabilities correspond reasonably to observed frequencies.

Question 4

Which analysis changes assumptions to examine whether a conclusion is robust?

A. Sensitivity analysis

B. Data encryption

C. Descriptive reporting

D. Random sampling only

Correct Answer: A. Sensitivity analysis

Question 5

In a decision tree, a chance node represents:

A. A decision-maker’s selected action

B. A resource constraint

C. An uncertain event with possible outcomes

D. The final accounting entry

Correct Answer: C. An uncertain event with possible outcomes

Question 6

Which expression represents expected monetary value?

A. The largest possible outcome only

B. The sum of each outcome multiplied by its probability

C. The average of probabilities without considering outcomes

D. The cost of the most likely action

Correct Answer: B. The sum of each outcome multiplied by its probability

Question 7

In an optimisation model, what is a constraint?

A. The quantity the model seeks to maximise

B. A limit restricting feasible decisions

C. A predicted probability only

D. A model performance metric that must equal zero

Correct Answer: B. A limit restricting feasible decisions

Question 8

Why can ranking customers by predicted risk alone be insufficient for intervention decisions?

A. Risk rankings never contain useful information.

B. Ranking ignores the costs, benefits, and feasibility of intervention.

C. Ranking automatically guarantees fairness.

D. Risk predictions cannot be compared across customers.

Correct Answer: B. Ranking ignores the costs, benefits, and feasibility of intervention.

Question 9

Which situation most strongly supports structured human review?

A. A low-impact decision with stable, well-monitored outcomes

B. A decision involving irreversible consequences and missing contextual information

C. A decision where no action is possible

D. A decision with no affected stakeholders

Correct Answer: B. A decision involving irreversible consequences and missing contextual information

Question 10

What is the main purpose of a baseline in cost–benefit analysis?

A. To provide a reference against which incremental effects are compared

B. To ensure every cost is counted twice

C. To replace all probability estimates

D. To guarantee that the preferred option has no risk

Correct Answer: A. To provide a reference against which incremental effects are compared