Business Problem Formulation and Analytics Design

Introduction

Machine learning projects rarely fail because an algorithm cannot be trained. They more often fail because the organization has not defined the business problem precisely, selected an appropriate unit of analysis, measured the right outcome, or connected predictions to decisions. Business analytics therefore begins before data preparation and model fitting. It begins with disciplined problem formulation.

Problem formulation translates a managerial concern into an analytical task that can be answered with data and acted upon by people or systems. For example, “reduce customer churn” is a strategic objective, not yet a machine-learning problem. It must become a question such as: “Which active customers are most likely to cancel their subscription during the next 30 days, and which of them should receive a retention offer?” This formulation identifies a target, a prediction horizon, a population, an intervention, and a decision.

This chapter presents a practical design framework for translating business challenges into classification, regression, forecasting, clustering, recommendation, or anomaly-detection tasks. It also explains units of analysis, targets, features, constraints, baselines, key performance indicators (KPIs), decision workflows, and the CRISP-DM process.

Learning Objectives

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

1. Distinguish a business objective from an analytical question and a machine-learning task.
2. Select an appropriate analytical task for a managerial problem.
3. Define the unit of analysis, observation window, target, prediction horizon, and features.
4. Recognize leakage, sampling problems, confounding, and misaligned evaluation designs.
5. Establish useful business and analytical baselines.
6. Define model, business, operational, and responsible-use KPIs.
7. Design a decision workflow linking predictions to actions.
8. Apply CRISP-DM critically to a business analytics project.
9. Evaluate limitations, governance requirements, and implementation risks.

1. From Managerial Concern to Analytical Specification

A managerial concern describes an undesirable condition or an opportunity. An analytical specification makes the concern measurable and operational. A useful translation has six elements:

A well-formed problem statement can be written as follows:

For [population], use information available at [decision time] to estimate or describe [target] over [horizon], so that [decision-maker or system] can take [action], evaluated by [business and responsible-use criteria].

This statement prevents a common error: building a technically impressive model that answers a different question from the one the organization needs answered.

Descriptive, predictive, and prescriptive questions

Descriptive analytics asks what has happened or what is happening. Diagnostic analytics asks why a pattern may have occurred. Predictive analytics estimates an unknown or future outcome. Prescriptive analytics recommends or triggers an action. Machine learning is most often used for predictive or recommendation tasks, but the business value usually arises at the prescriptive stage.

A prediction is not automatically a recommendation. A model may estimate that a customer is likely to churn, but this does not prove that a discount will prevent churn. Prediction describes risk; intervention requires evidence about treatment effects, costs, and alternatives.

2. Choosing the Analytical Task

The target, decision, and data structure determine the task. The following categories are related but not interchangeable.

Classification

Classification predicts a categorical outcome. Binary classification has two classes, such as fraud/not fraud or renew/not renew. Multiclass classification has more than two mutually exclusive classes, such as low, medium, and high risk. Multilabel classification allows several labels simultaneously, such as a support ticket involving both billing and technical issues.

For a binary target, let y ∈ {0,1} denote the observed class and p(x) = P(y = 1 | x) the estimated probability conditional on features x. A threshold can convert the probability into an action:

Here τ is the decision threshold. It need not be 0.5. If missing a positive case is much more costly than contacting an unnecessary case, a lower threshold may be appropriate.

Regression

Regression predicts a numeric quantity, such as claim cost, customer lifetime value, delivery time, or monthly sales. The target is continuous or approximately continuous. Regression is suitable when the magnitude of the outcome matters, not merely its category.

A basic error measure is mean absolute error:

Here n is the number of observations, yᵢ is the actual value, and ŷᵢ is the prediction. MAE has the same units as the target and is therefore often easy to explain. It treats all absolute errors proportionally. Its interpretation assumes that the observations and target measurement are defined consistently.

Forecasting

Forecasting predicts values indexed by time, such as daily demand, weekly cash flow, or hourly call volume. Forecasting is a special form of prediction because temporal order matters. Future information must not be used to predict the past, and evaluation should reproduce the intended deployment sequence.

A simple baseline for a one-period-ahead series is the naive forecast:

This says that the next value equals the latest observed value. It is often surprisingly competitive and provides an essential benchmark. Seasonal settings may require a seasonal-naive baseline, such as using the value from the same weekday in the previous week.

Forecasting questions should specify the forecast origin, horizon, frequency, and aggregation level. “Predict sales” is incomplete; “predict daily units sold for each store seven days ahead using information available at the end of each day” is operationally meaningful.

Clustering

Clustering groups observations without a predefined target. It may support customer segmentation, store typology, or product portfolio analysis. A cluster is a statistical grouping, not automatically a meaningful market segment. Managers must interpret whether clusters differ in ways that support distinct actions.

Clustering requires choices about representation, scaling, distance, number of groups, and stability. For example, income measured in currency can dominate a distance calculation if variables are not appropriately scaled. Clustering should be judged by both technical coherence and managerial usefulness.

Recommendation

Recommendation systems rank items, content, offers, or actions for a user or context. The central output is usually an ordered list rather than a single class. Examples include products for a shopper, articles for a reader, or maintenance actions for an engineer.

A recommendation problem must define the candidate set, eligibility rules, ranking objective, exposure constraints, and feedback mechanism. Observed clicks are not necessarily preferences: an item cannot be clicked if it was never shown. This creates selection bias and makes offline evaluation difficult.

Anomaly detection

Anomaly detection identifies observations that differ substantially from an expected pattern. It is useful for fraud screening, cybersecurity, equipment monitoring, and quality control. An anomaly score is not the same as proof of misconduct or failure. Alert thresholds must account for investigation capacity and the cost of false alarms.

Anomalies may be point anomalies, contextual anomalies, or collective anomalies. A transaction may be normal in general but unusual for a particular account; a sequence of individually ordinary events may be suspicious collectively.

3. Defining the Data-Generating Structure

Unit of analysis

The unit of analysis is the entity represented by one analytical row. It may be a customer, account-month, transaction, store-day, machine reading, support ticket, or product-user pair. The unit must match the decision being supported.

A customer-level retention model and a transaction-level fraud model should not use the same row definition. If the organization makes one decision per customer per month, a customer-month may be appropriate. If the model is trained on multiple rows for a customer but the intervention is made once, aggregation and dependence must be considered.

Observation window, outcome window, and prediction point

The observation window contains information allowed to construct features. The outcome window contains the period in which the target is measured. The prediction point separates the two.

For a customer-month model, data observed from 1 January through 31 March might predict churn during April. Information recorded after 31 March must not enter the feature set, even if it is available in a data warehouse today. This temporal rule is central to preventing leakage.

Target definition

A target should be observable, consistently labeled, and connected to a decision. Important design choices include:

• Event definition: What exactly counts as churn, default, conversion, or failure?
• Horizon: Over what period is the event measured?
• Censoring: What happens when the observation period ends before the event can be observed?
• Exclusions: Which cases are not eligible for prediction?
• Label delay: When does the outcome become reliably known?

Ambiguous labels produce ambiguous models. For instance, “inactive customer” could mean no login for 30 days, no purchase for 90 days, or account closure. These are different targets.

Features and feature availability

Features are variables used to make a prediction or construct a grouping. They may describe behavior, transactions, context, product attributes, or operational conditions. A feature is acceptable only if it is available at the prediction point, measured reliably, and permissible under policy.

Feature selection should consider causal and operational meaning. A variable can be highly predictive because it records a decision already made by the organization. Such a feature may produce excellent historical accuracy while adding no value at deployment.

Leakage

Leakage occurs when information unavailable at decision time enters training or evaluation. Examples include a post-default collections code in a default model, a cancellation reason in a churn model, or a future sales total in a demand forecast. Leakage inflates performance estimates and can cause severe deployment failure.

Leakage may be direct, indirect, temporal, or caused by preprocessing. Randomly splitting repeated customer records can also allow near-duplicate information from the same customer to appear in both training and test sets. Splitting must reflect the deployment setting.

4. Constraints, Baselines, and KPIs

Business and operational constraints

A model operates within constraints. These may include latency, budget, staffing, inventory, legal requirements, customer experience, explainability, and integration with existing systems. A technically optimal model may be unsuitable if it cannot return a score within the required time or if no team can act on its alerts.

Constraints should be specified before model selection. Examples include:

• A fraud screening service must respond in less than one second.
• A marketing team can contact no more than 10,000 customers per week.
• A credit decision requires an understandable reason code.
• A recommendation must exclude unavailable or restricted products.

Baselines

A baseline is a simple reference against which a proposed approach is compared. Baselines may be:

1. Current practice: the existing manual or rule-based process.
2. Statistical: majority class, historical mean, naive forecast, or seasonal-naive forecast.
3. Operational: the action selected under current staffing or budget constraints.

Without a baseline, a model metric has little meaning. A classifier with 95% accuracy may be poor if 95% of cases belong to the majority class. A forecast with low error may still be inferior to a simple seasonal rule.

Evaluation metrics

Metrics must reflect the decision. For binary classification, a confusion matrix distinguishes true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN). Common measures include:

Precision answers: among predicted positives, how many were positive? Recall answers: among actual positives, how many were identified? Both denominators must be nonzero, and their interpretation depends on a clearly defined positive class.

Accuracy is:

It can be misleading under class imbalance. A business may instead use a cost-weighted measure. If C_FP is the cost of a false positive and C_FN the cost of a false negative, expected error cost can be written as:

This simplified expression assumes that true-positive and true-negative benefits are handled separately or are not included. In a complete business case, all relevant revenues, costs, and intervention effects should be included.

For regression, MAE or root mean squared error may be used. Root mean squared error is:

RMSE uses the target's units but penalizes large errors more strongly than MAE. It is appropriate when large errors are disproportionately undesirable and the target scale is meaningful.

Business KPIs and model KPIs

Model metrics are not business outcomes. A retention model may improve recall while generating unprofitable offers. A recommendation model may increase clicks but reduce margin or customer trust. KPIs should therefore be layered:

Calibration is important when probabilities guide resource allocation. A calibrated group of cases assigned probability 0.7 should contain approximately 70% positive outcomes, subject to sampling variation and a stable definition of the target.

5. Decision Workflows and the Value of Prediction

A decision workflow specifies what happens before, during, and after scoring:

1. Eligibility: identify cases that may receive a decision.
2. Data capture: assemble features available at the decision point.
3. Scoring: generate a probability, value, rank, segment, or anomaly score.
4. Policy: apply thresholds, budgets, exclusions, and escalation rules.
5. Action: contact, approve, inspect, replenish, recommend, or defer.
6. Outcome capture: record the action and eventual result.
7. Monitoring: assess performance, drift, fairness, and operational effects.

The economic value of a model depends on action, not prediction alone. Suppose a retention campaign costs $8 per contacted customer, and a successful save contributes $40. If a model identifies a customer with an estimated probability of successful saving p, the simplified expected contribution before other costs is:

Contacting is attractive under this simplified assumption when p × 40 − 8 > 0, or p > 0.20. In practice, this threshold must also consider customers who would have stayed without intervention, treatment effects, capacity, and fairness. The calculation illustrates why a probability threshold should be connected to economics rather than chosen automatically.

Uplift or treatment-response modeling is often more appropriate when the objective is to identify who will change behavior because of an intervention. A high-risk customer may be impossible to retain, while a low-risk customer may remain regardless of treatment. The best target for an offer is the customer with favorable incremental response, not necessarily the highest predicted risk.

6. CRISP-DM as a Design Framework

CRISP-DM—Cross-Industry Standard Process for Data Mining—organizes analytics work into six iterative phases:

1. Business understanding: define objectives, decisions, stakeholders, constraints, and success criteria.
2. Data understanding: collect initial data, describe variables, explore quality, and identify potential bias or leakage.
3. Data preparation: select cases, clean data, construct features, label targets, and create analysis-ready datasets.
4. Modeling: select techniques, train models, tune parameters, and document assumptions.
5. Evaluation: assess technical performance and confirm that results answer the business question and satisfy constraints.
6. Deployment: deliver scores, reports, decisions, or services; monitor performance and maintain the solution.

The phases are iterative, not a one-way staircase. Evaluation may reveal that the target is poorly defined; deployment may reveal that a feature is unavailable in production; business understanding may change after stakeholder interviews.

CRISP-DM is valuable because it places business understanding first and deployment inside the project rather than treating it as an afterthought. It is not, however, a complete governance framework. Modern projects should add explicit documentation of privacy, fairness, security, model risk, data lineage, human oversight, and post-deployment monitoring.

7. Worked Example: Designing a Demand-Forecasting Project

A grocery retailer reports frequent stockouts and excess spoilage. The initial request is “use AI to improve inventory.” A disciplined formulation proceeds as follows.

Business objective. Reduce stockouts and waste while maintaining service levels.

Decision. Determine replenishment quantities for each product-store combination.

Unit of analysis. Product-store-day, because replenishment decisions are made at that level.

Prediction target. Units sold on the next day, or possibly demand adjusted for stockouts. The distinction matters: observed sales may understate true demand when shelves were empty.

Prediction horizon. One day ahead initially, with a later extension to multiple days.

Features. Recent sales, day of week, price, promotion, holiday indicators, weather where relevant, inventory position, delivery schedule, and product attributes. Every feature must be available before the replenishment decision.

Baseline. A seasonal-naive forecast using sales from the same weekday in the previous week.

Metrics. MAE in units for operational interpretation, weighted error for high-value products, stockout rate, waste, and service level. The retailer should not optimize forecast error while ignoring inventory consequences.

Constraints. Supplier minimum order quantities, shelf capacity, delivery calendars, budget, perishability, and system response time.

Evaluation design. Use time-ordered validation. A model trained on earlier dates should be tested on later dates. A random split would allow future patterns to influence training and produce an unrealistic estimate.

Decision workflow. Forecast demand, combine it with current inventory and lead time, apply order constraints, route exceptions to planners, and record realized sales, stockouts, waste, and overrides.

Risks. Stockouts censor demand; promotions may change both demand and availability; a forecast can alter ordering and therefore alter future data. The system must monitor drift caused by changing prices, competitors, seasons, and supply conditions.

The correct outcome is not simply “the model has lower MAE.” The outcome is improved replenishment performance under the retailer’s constraints, demonstrated through a controlled pilot or carefully designed comparison.

8. Illustrative Case: Retention Analytics at Northstar Mobile

This is a fictional illustrative case.

Background and problem

Northstar Mobile, a subscription provider, experiences rising monthly cancellations. Managers ask the analytics team to “predict churn and save customers.” The team first interviews marketing, customer service, finance, and compliance staff. They discover that retention agents can contact 5,000 customers weekly, and that discounts are expensive. The true decision is therefore not merely who is likely to churn, but whom to contact and what action to offer.

Analysis and formulation

The team defines the unit as customer-week for active customers eligible for service. The prediction point is the end of each week. Features include usage changes, payment events, complaint history, contract status, and service interruptions observed up to that point. The target is account cancellation during the following 30 days. Customers who are already scheduled to end service for a non-preventable reason are excluded from the intervention population.

The team identifies a serious leakage variable: a cancellation workflow code entered after a customer calls to terminate service. It is removed. They also discover that customer records are repeated across weeks, so evaluation is time-based and customer-aware rather than a careless random row split.

Approach

A majority-class classifier and the existing rule-based campaign form the baselines. The team evaluates recall and precision at the 5,000-customer capacity, calibration, contact cost, retention rate, and customer complaints. Because an observational churn model cannot establish that a discount will save a customer, Northstar runs a randomized test among eligible high-score customers: some receive the offer, while a control group receives normal service.

The test shows that the highest-risk customers are often already determined to leave and respond weakly to discounts. Moderately high-risk customers with recent service problems respond more strongly to service recovery. Northstar therefore creates two policies: service recovery for one segment and a limited offer for another. Customers in a protected category are monitored for disparate contact rates and outcomes, and the business documents legitimate reasons for any differentiated treatment.

Outcome

The final evaluation considers incremental retention, contribution after offer cost, agent workload, complaint rates, and fairness indicators. The model is deployed with a weekly batch score and a dashboard for campaign managers. Performance is monitored because customer behavior, tariffs, and competitor offers change over time.

Learning points

1. Churn prediction and churn prevention are different problems.
2. The unit, target horizon, and eligibility rules determine what the model means.
3. Leakage can be hidden in operational workflow variables.
4. A capacity limit makes ranking and top-k evaluation more relevant than overall accuracy.
5. Intervention experiments are needed to estimate incremental business value.
6. Governance and monitoring continue after deployment.

9. Common Failure Modes and Best Practices

Failure mode: starting with an algorithm

Choosing a neural network before defining the decision encourages metric chasing. Begin with the decision, target, timing, and baseline. Model complexity should be justified by incremental value.

Failure mode: confusing correlation with actionability

A predictive feature may be associated with an outcome without being a useful lever. A model can identify risk without showing how to reduce it. Pair predictive work with experiments, domain knowledge, or causal analysis when intervention is central.

Failure mode: optimizing a proxy

Clicks, time on site, or model accuracy may be convenient proxies but can conflict with profit, satisfaction, safety, or equity. Document the relationship between the proxy and the ultimate objective and test for unintended effects.

Failure mode: ignoring missingness and selection

Missing values may reflect process failure, customer behavior, or eligibility rather than random absence. Similarly, historical decisions determine which cases appear in the data. Document data-collection processes and consider whether the training population represents deployment.

Failure mode: treating deployment as completion

Data distributions, customer behavior, policies, and software dependencies change. Monitor data quality, feature drift, target performance when labels arrive, decision outcomes, latency, overrides, and fairness. Establish ownership for retraining and rollback.

Ethical, legal, and governance considerations

Analytics design should apply data minimization, purpose limitation, access control, retention rules, and appropriate security. Sensitive attributes may require special handling, but omitting them does not guarantee fairness because correlated variables may remain. Fairness assessment should consider the relevant decision, population, metric, and legal context; no single fairness metric is universally correct.

Organizations should provide meaningful explanations where decisions affect people, maintain audit trails, validate data lineage, and preserve human review for high-impact cases. Automated recommendations should not be presented as objective facts. Governance should specify who can approve deployment, who investigates complaints, and how affected individuals can seek correction or review.

10. Contemporary Developments

Modern analytics design increasingly includes real-time scoring, automated feature pipelines, foundation models, synthetic data, and machine-learning operations (MLOps). These technologies can improve speed and scale, but they do not remove the need for target definition, causal reasoning, validation, or governance.

Generative AI may assist with problem discovery, documentation, or text classification, but generated outputs require validation and controlled data handling. Automated machine learning can compare algorithms efficiently, yet it cannot determine whether a target represents the organization’s true objective. Responsible design remains a human and organizational responsibility.

Best-Practice Design Checklist

Before modeling, confirm that:

• the decision and decision owner are identified;
• the unit of analysis matches the decision frequency;
• the prediction point, observation window, and outcome horizon are explicit;
• the target is observable and operationally meaningful;
• features are available at deployment time;
• leakage risks and exclusions are documented;
• baselines and constraints are defined;
• metrics include business and responsible-use outcomes;
• evaluation reflects time, grouping, and deployment conditions;
• an action, capacity rule, and escalation process exist;
• data protection, fairness, security, and audit requirements are addressed;
• monitoring, ownership, and rollback plans are specified.

Chapter Summary

Business analytics begins with formulation rather than algorithm selection. A managerial objective must be translated into a decision, population, unit of analysis, target, timing structure, features, constraints, and success measures. Classification predicts categories, regression predicts numeric values, forecasting predicts time-indexed outcomes, clustering discovers groups, recommendation ranks alternatives, and anomaly detection identifies unusual cases. Each task has distinct data and evaluation requirements.

Reliable design prevents leakage by respecting information available at the prediction point. It uses baselines to establish whether complexity adds value and selects KPIs that connect model behavior to business, operational, and responsible-use outcomes. A decision workflow explains how scores become actions and how outcomes feed monitoring and improvement. CRISP-DM provides an iterative structure from business understanding through deployment, but contemporary projects also require privacy, fairness, security, documentation, and ongoing governance. The central principle is simple: a successful machine-learning project is not merely accurate; it supports a clearly defined decision and creates sustainable value under real-world constraints.

Key Terms

Unit of analysis: The entity represented by one analytical observation or row.

Target: The outcome a model is designed to predict or describe.

Feature: An input variable used to generate a prediction, grouping, ranking, or score.

Prediction point: The time at which information is collected and a decision may be made.

Observation window: The period from which permitted feature information is collected.

Outcome window: The period during which the target outcome is measured.

Classification: Prediction of a categorical outcome.

Regression: Prediction of a numeric outcome.

Forecasting: Prediction of a time-indexed future value.

Clustering: Grouping observations without a predefined target.

Recommendation: Ranking items, actions, or content for a user or context.

Anomaly detection: Identification of observations that differ from expected patterns.

Data leakage: Use of information unavailable at the intended prediction time.

Baseline: A simple reference method used for comparison.

Precision: The proportion of predicted positives that are actual positives.

Recall: The proportion of actual positives that are identified.

Calibration: Agreement between predicted probabilities and observed frequencies.

CRISP-DM: An iterative process framework covering business understanding, data understanding, preparation, modeling, evaluation, and deployment.

Concept drift: Change over time in the relationship between inputs and outcomes.

Decision workflow: The sequence connecting eligibility, data, scoring, policy, action, outcomes, and monitoring.

Review Questions

Short Answer Questions

1. What is the difference between a business objective and a machine-learning task?
2. Why must the unit of analysis match the decision being supported?
3. Define data leakage and give one business example.
4. Why can accuracy be misleading for an imbalanced classification problem?
5. What are the six phases of CRISP-DM?

Long Answer Questions

1. Explain how a managerial concern can be translated into a complete analytical specification. Use a business example.
2. Compare classification, regression, forecasting, clustering, recommendation, and anomaly detection in terms of targets and applications.
3. Discuss how observation windows, outcome windows, and prediction points prevent leakage.
4. Explain why model KPIs and business KPIs should be evaluated separately but connected.
5. Describe the role of constraints, human oversight, and monitoring in deploying a machine-learning system.

Analytical / Application Questions

1. A bank wants to “predict loan defaults.” Design the unit of analysis, target horizon, features, baseline, evaluation approach, and decision workflow. Identify at least two governance risks.
2. A retailer has a model that predicts which customers will buy after receiving a coupon. Explain why a churn or purchase-propensity model may not identify the customers most worth treating. Propose an evaluation design.
3. A call center forecasts hourly call volume. Explain why a random train-test split could be misleading, select an appropriate baseline, and recommend operational and business KPIs.

Multiple Choice Questions with Answers

Question 1

Which element specifies the entity represented by one analytical row?

A. Target

B. Unit of analysis

C. Prediction threshold

D. Baseline

Correct Answer: B. Unit of analysis

Question 2

A model uses a cancellation code entered after a customer requests termination. What problem does this create?

A. Underfitting

B. Class balancing

C. Data leakage

D. Clustering instability

Correct Answer: C. Data leakage

Question 3

Which task predicts a numeric quantity such as claim cost?

A. Regression

B. Classification

C. Clustering

D. Anomaly detection

Correct Answer: A. Regression

Question 4

What does recall measure in binary classification?

A. The proportion of predicted positives that are correct

B. The proportion of actual positives that are identified

C. The proportion of all predictions that are correct

D. The average size of prediction errors

Correct Answer: B. The proportion of actual positives that are identified

Question 5

Which is an appropriate baseline for a one-period-ahead time series when no strong seasonal pattern is assumed?

A. Predict the latest observed value

B. Predict the largest historical value

C. Randomly shuffle the time series

D. Use the test target as a feature

Correct Answer: A. Predict the latest observed value

Question 6

Why is a random split often inappropriate for forecasting evaluation?

A. It always increases sample size

B. It can allow future information to influence evaluation of past-like predictions

C. It prevents calculation of MAE

D. It eliminates seasonality

Correct Answer: B. It can allow future information to influence evaluation of past-like predictions

Question 7

Which CRISP-DM phase focuses on defining objectives, stakeholders, decisions, and success criteria?

A. Business understanding

B. Data preparation

C. Modeling

D. Deployment

Correct Answer: A. Business understanding

Question 8

If a retention contact costs $8 and a successful save contributes $40, which condition makes the simplified expected contribution positive?

A. p < 0.08

B. p > 0.20

C. p > 0.40

D. p = 8

Correct Answer: B. p > 0.20

Question 9

What is the main purpose of a baseline?

A. To guarantee fairness

B. To replace business understanding

C. To provide a reference for judging added value

D. To remove the need for monitoring

Correct Answer: C. To provide a reference for judging added value

Question 10

Why does a high predicted churn probability not necessarily identify the best customer to contact?

A. Predicted probabilities cannot be stored

B. Risk prediction does not establish that an intervention will change the outcome

C. Churn is always a clustering problem

D. High-risk customers are never eligible for action

Correct Answer: B. Risk prediction does not establish that an intervention will change the outcome