0. Decision and Data Foundations

Decision Framing

Convert a broad business request into actions, outcomes, constraints and evidence

Decision Framing

Replace the topic with a choice

“Analyse late deliveries” names a topic. HarborMart must choose among actions:

  • accept the requested window;
  • offer a later window;
  • reroute fulfilment;
  • add temporary capacity;
  • leave the current process unchanged.

An analysis is decision-ready only when at least two feasible alternatives exist.

A six-part decision statement

At checkout, the fulfilment manager chooses an available delivery promise for one proposed order, to maximise expected contribution after service-failure cost, subject to capacity, safety and customer-commitment constraints.

PartTest questionFailure mode
ownerwho is authorised to act?“the business” owns nothing
actionwhat can change now?predicting an immutable outcome
unitone customer, order, item or day?mixed denominators
horizonwhen is the outcome realised?target observed too late
objectivewhich consequences count?optimising clicks while harming margin
constraintswhat cannot be violated?infeasible recommendation

Worked decision table

Suppose an order yields £18 contribution if on time. A late delivery loses £7 through refunds and future attrition. Offering a later slot preserves £15 contribution and reduces late risk.

ActionLate probabilityExpected value
accept requested slot0.35180.35(7)=£15.5518-0.35(7)=£15.55
offer later slot0.08150.08(7)=£14.4415-0.08(7)=£14.44

Under only these consequences, accepting has higher expected value. If a late event instead costs £15, the values become £12.75 and £13.80, so the later slot wins. The prediction did not change; the decision changed because the loss did.

Threshold from consequences

Let pp be late probability, VAV_A on-time contribution for accepting, CLC_L late cost, and VBV_B the value of a safer alternative. Accept when

VApCLVB.V_A-pC_L \ge V_B.

Therefore the maximum acceptable risk is

p=VAVBCL.p^*=\frac{V_A-V_B}{C_L}.

With VA=18V_A=18, VB=15V_B=15 and CL=15C_L=15, p=0.20p^*=0.20. A universal 0.5 classification threshold would be economically arbitrary.

Prediction, intervention and policy

QuestionObject needed
Which orders will be late?P(Y=1X)P(Y=1\mid X)
Would adding a picker prevent lateness?E[Y(1)Y(0)X]E[Y(1)-Y(0)\mid X]
Where should three picker-hours go?effects plus costs and capacity constraints

A risk model can prioritise review. It cannot by itself estimate the effect of the review or allocate scarce staff.

Value of information

Information is valuable only when it can change an action enough to improve expected outcomes.

If both a 10% and a 30% late-risk estimate lead to accepting the order, refining the estimate between those values has no immediate decision value. Near the 20% threshold, a modest improvement in calibration may matter greatly.

This is why model error should later be translated into decision regret, not reported only as RMSE or AUC.

Quick check

A campaign team asks, “Which customers are likely to churn?” What three facts are still missing before modelling?

Answer
At minimum: the action available for a flagged customer, the decision time/horizon, and the benefit and cost of intervention. Also define churn, eligibility, capacity and the no-action baseline.

Foundation

Next: Data Contracts

Copyright © 2026