Predictive Analytics in Human Resource Management A Hands-on Approach

1. 1Start from a business symptom, not from a survey. Define the problem precisely: where and what, over which period (e.g., sales in the southern region have dropped while other regions run smoothly). Why it happened comes later.
2. 2Write the conceptual model before collecting anything. State the hypothesis (e.g., a sales decline is caused by adverse employee attitudes or by lower product quality) and name the dependent variable (sales, customer satisfaction, attrition) and the independent variables (the attitudes).
3. 3Choose which attitudes to collect using experience, conversations with experts or seniors, and theory. Do not collect every attitude. Select the ones your model says affect the outcome.
4. 4Build the data foundation first. Keep attitude data in an employee database alongside behaviour and on-job performance, as Google does, so attitudes can be linked to outcomes at the individual level.
5. 5Categorise the analytical problem as classification, clustering, or association before choosing a tool. Use correlation on historical data to test whether frontline attitudes or retention move with a business outcome. Use clustering to find employee segments. Use classification to predict who leaves or stays.
6. 6Segment before you act. Identify what differentiates groups, then work out what each group values most and treat each differently. Vary HR actions by employee life cycle.
7. 7Interpret and validate the results, then convert them into recommendations. If an attitude finding points to a further cause (e.g., dissatisfaction with pay in one region), flag it as a follow-up investigation, much as a doctor prescribes further tests.
8. 8Map the process as inputs, transformation, and outputs (as in Tables 1.1 and 1.2) so that the attitude-to-action logic is explicit and repeatable.

A global restaurant chain wanted higher customer satisfaction. Correlation analysis of historical data on frontline employees and customer satisfaction showed a significant relationship between the two, so the chain chose to improve frontline retention to raise revenue. It built a predictive model on individual employee data and found distinct clusters of employees, labelled 'socialisers', 'potential leaders', 'entrepreneurial taskmasters' and 'conservative taskmasters'. The conservative taskmasters were the most consistent performers. Performance depended more on career development opportunities and cultural norms than on the variable compensation then offered. Acting on these findings improved customer satisfaction, speed of service, sales and attrition rates. The lesson for job attitudes: the pay-centred assumption was wrong, and the data redirected investment toward development and culture.

• If a performance metric such as sales or turnover misses its standard in one unit, then treat it as a symptom and test whether employee attitudes explain it before assuming a quality or market cause.
• If a correlation links a workforce condition (e.g., frontline retention) to a customer or revenue outcome, then build a predictive model on individual employee data before committing investment.
• If analytics shows adverse attitudes tied to a specific grievance such as pay dissatisfaction in one region, then log it as a defined follow-up investigation rather than a conclusion.
• If segments differ in what they value, then design a differentiated action for each segment, and adjust timing by life-cycle stage (Convergys found that paying increments every six months rather than yearly made employees more likely to stay).
• If you cannot choose which attitudes to measure, then fall back on theory, expert and senior input, and experience to narrow the list.

• Collecting attitude data with no model. The tell: a large survey database with no stated dependent variable or hypothesis.
• Assuming compensation drives performance and attitudes. The tell: variable pay is increased while the data points to career development and cultural norms.
• Treating one attitude finding as the final answer. The tell: pay dissatisfaction in a region is reported as the cause with no follow-up test.
• Keeping attitude data separate from performance and behaviour records. The tell: attitudes cannot be joined to outcomes per employee, so no prediction is possible (the problem Google's single database avoids).
• Applying one uniform action to all employees. The tell: the same retention offer goes to every segment, ignoring differing values and life-cycle stage.
• Claiming direct ROI for attitude programmes. The tell: the case rests on a precise return figure, though the source says direct ROI on human capital cannot be shown and only the chain from HR investment to attitudes to business outcomes can.

• ▸Business symptom: state the metric that missed its standard, where, and over what period.
• ▸Hypothesis: write the attitude you believe drives the outcome and the competing explanation (e.g., product quality).
• ▸Dependent variable: name the outcome measured (sales, customer satisfaction, attrition).
• ▸Independent variables (attitudes): list the attitudes to collect and why (experience, expert input, or theory).
• ▸Data sources: record where attitude, behaviour and performance data sit, and how they are linked per employee.
• ▸Problem type and tool: mark classification, clustering, or association, and name the chosen technique.
• ▸Segments found: describe each employee group and what it values most.
• ▸Differentiated actions: list the action per segment and its life-cycle timing.
• ▸Follow-up investigations: list any new issue the result exposed (e.g., regional pay dissatisfaction) and who owns it.

1. 1Fix the definition before you collect anything: turnover intent is the reflection of the probability that an individual will change his/her job within a certain time period (Sousa-Poza and Henneberger, 2002). Write the time window into the question (the passages use 'the upcoming year'), so every respondent answers about the same horizon.
2. 2Separate it from actual turnover. Year-end figures show actual turnover, but they cannot show who intends to leave. Treat intent as a distinct measurement, taken periodically, not as a retrospective count.
3. 3Name intent as the target variable of your predictive model. The source states that for the prediction of turnover, turnover intent can be used as the target variable. Every predictor you gather must be something you can tie to this target.
4. 4Scope the problem narrowly in step 1 of the holistic approach. The source example is turnover prediction for software engineers in the IT industry in India, a sector with high annual turnover. Pick one population and one sector context, not 'all employees'.
5. 5Build the predictor set from records that already exist, going beyond exit interviews: tenure, performance ratings, skill levels, and previous experience. Associate each one with turnover intent, as the IT-firm example does. Add other attributes your data supports (Experian used about 200, including supervisor performance and commuting distance).
6. 6Run the prediction with the supplied tools: KNN (and ANN) in R, with the confusion matrix and ROC as performance metrics. Check the confusion matrix before you trust any accuracy figure.
7. 7Close the loop with prescription. Classify the likely voluntary leavers, because voluntary turnover is avoidable with timely intervention. Map each risk driver to an action, as Nielsen does by tenure: for first-year hires, managers are reminded to make contact.
8. 8Re-measure on a schedule. Firms gauge intent periodically; compare each wave with the actual turnover that follows to see whether the model is predicting leavers.

An IT firm in India faces high annual turnover among software engineers. Exit interviews arrive too late, so the analyst goes to employment records instead. She pulls tenure, performance ratings, skill levels, and previous experience, and associates each with turnover intent, which she defines as the probability of changing jobs in the coming year. Intent is the target variable. She fits a KNN model in R and checks the confusion matrix and ROC. The model flags high-intent groups early, and HR designs retention measures for them. This follows the source's description of how Indian IT firms predict turnover better and lower annual turnover. The Experian parallel shows the scale of the payoff: attrition prediction on about 200 attributes cut turnover by 2–3% over one year. For retention actions, Nielsen's rule by tenure applies: for first-year hires, managers are prompted to contact them.

• If you only have year-end turnover figures, then you cannot see intent; add a periodic intent measure and use it as the target variable.
• If an exit is voluntary (a resignation), then treat it as avoidable and design a timely intervention. If it is involuntary (dismissal, death), keep it out of the intent analysis.
• If model inputs are limited to exit-interview content, then widen to employment records (tenure, performance ratings, skill level, previous experience) and test each against intent.
• If the model flags risk by tenure band, then prescribe by band: for first-year hires, schedule manager contact points, as Nielsen does.
• If the departing group is low performers, then the turnover may be functional; if it is key performers, it is dysfunctional and merits intervention.

• Leaving the time window undefined. The tell: respondents' answers are not comparable and the model cannot be checked against a defined year of actual turnover.
• Equating intent with actual turnover. The tell: you report year-end turnover as if it measured who planned to leave, and cannot say who was at risk beforehand.
• Relying on exit interviews alone. The tell: insight arrives only after people have left and no one can be targeted in advance.
• Predicting without prescribing. The tell: a risk list exists but no retention action is tied to tenure or other drivers, so turnover does not fall.
• Treating all turnover as bad. The tell: you spend retention effort on departures that are functional (movement of unproductive employees), which is dysfunctional use of resources.

• ▸Population and sector: state the single group modelled (e.g., software engineers in IT).
• ▸Intent definition and window: write the probability-of-leaving statement and the time period (e.g., upcoming year).
• ▸Target variable: name the intent measure and how often it is gathered.
• ▸Predictor list: tenure, performance rating, skill level, previous experience, plus any other attributes you hold (e.g., supervisor performance, commuting distance).
• ▸Data sources: list the record systems each predictor comes from.
• ▸Model and metrics: tool chosen (KNN or ANN in R) and the confusion matrix / ROC results.
• ▸Voluntary vs involuntary split: how each exit type is coded and handled.
• ▸Retention action by risk group: the intervention tied to each driver or tenure band (e.g., first-year manager contact).
• ▸Review date and actual-turnover comparison: when intent predictions are checked against real exits.

1. 1Define the event before touching data. Turnover is movement across the membership boundary, and the costly kind is voluntary. Write the outcome as a binary target (left voluntarily or stayed) and decide how involuntary exits are handled. The book's models treat the target as a factor variable, so the definition has to be exact before coding.
2. 2Draw the problem before modelling it. Use an influence diagram and sort variables into dependent (turnover), independent (e.g. supervisor performance, commuting distance), moderator and mediator. This stops you from feeding in variables whose role you cannot explain.
3. 3Build the employee database first. The source names the employee database as the foundation of any analytical application. Pull one row per employee with the attributes you can defend, then log what was recorded and when. Experian-style prediction relied on about 200 attributes, which only works if the data was already being kept.
4. 4Prepare the data for the chosen tool. For an artificial neural network, normalise the dataset (the book's R walkthrough does this). For K-nearest neighbour, set the target as a factor. Pick the tool using the book's 'basis to choose analytical tools for business problems', not preference.
5. 5Split the dataset into training and test portions, and fit the model on the training portion only. The book's demonstrations do this in R for both ANN and KNN. Never score the model on rows it learned from.
6. 6Evaluate on the held-out portion with a confusion matrix and the misclassification error. Re-run KNN with ROC as the performance metric and compare the two confusion matrices. Judge the model by how it handles leavers, not by overall accuracy alone.
7. 7Read variable importance and map the top drivers back to your influence diagram. Drivers you can act on (supervisor, commute) go to managers. Drivers you cannot act on inform hiring.
8. 8Convert prediction into prescription. Nielsen's model prescribes retention actions by tenure. For first-year hires, managers are reminded to make contact at critical contact points. Attach one concrete action and one owner to each risk segment, then track turnover over the following year as Experian did.

Experian (2020) predicted employee attrition by computing employee risk factors with predictive algorithms. The prediction used about 200 attributes, including the performance of the supervisor and commuting distance. After one year of implementation the firm reduced turnover by 2–3%. Nielsen (2020) took the next step with prescriptive modelling of turnover and retention, set out by years of tenure. For first-year new hires the model prescribes establishing critical contact points, so managers of new hires are reminded to contact them during that first year. The pattern to copy: a wide attribute set feeds the prediction, the drivers are named, and each tenure band gets a specific manager action. Success is measured as a turnover reduction over a defined period.

• If you only need to know who is likely to leave, use predictive analytics (Experian-style). If you must tell managers what to do about it, extend to prescriptive analytics by segment (Nielsen's tenure-based actions).
• If a driver such as supervisor performance or commuting distance ranks high in variable importance, then route it to the owner who can change it. If it cannot be changed, use it for hiring and screening instead.
• If the confusion matrix shows the model misses actual leavers, then change the performance metric (e.g. ROC for KNN) or revisit the variables before deploying. Do not rely on overall accuracy.
• If the employee database lacks the attributes in your influence diagram, then fix data collection first. A model on thin data produces confident noise.
• If the target mixes voluntary and involuntary exits, then separate them. The cost and the retention levers differ.

• Modelling before defining the outcome. The tell: the target variable mixes voluntary and involuntary exits and nobody can say which leavers the model is meant to catch.
• Judging the model on training data. The tell: strong results that vanish on the held-out split, or no train/test split in the script.
• Reading accuracy and ignoring the confusion matrix. The tell: high overall accuracy while many actual leavers are classified as stayers.
• Stopping at a risk score. The tell: a ranked list of at-risk employees with no named action, owner or tenure-based intervention, so turnover does not move.
• Collecting data but not retaining or quantifying it. The tell: HR holds large amounts of data but cannot link it to business outcomes, so the effect of turnover on unit performance stays unproven.

• ▸Outcome definition: state exactly what counts as a leaver (voluntary only or all exits) and the time window.
• ▸Dependent / independent / moderator variables: list each variable from your influence diagram and its role.
• ▸Data source and attributes: name the employee database fields (e.g. supervisor performance, commuting distance, tenure) and who owns them.
• ▸Tool and preparation: ANN (normalised data) or KNN (target as factor), and why, per the tool-selection basis.
• ▸Train/test split: record how the dataset was split and the size of each part.
• ▸Evaluation: record the confusion matrix, misclassification error and the metric used (e.g. ROC), with how many actual leavers were caught.
• ▸Top drivers: list the highest variable-importance items and mark each actionable or not.
• ▸Segment action: for each risk or tenure segment, state the intervention (e.g. manager contact in year one) and the owner.
• ▸Success measure: state the turnover change targeted and the review date.

1. 1Map hiring as a system before touching data. Write the applicant-screening process as inputs (applicant data), transformation (screening/selection), outputs (hire or reject) and a feedback loop. The source applies this process logic to applicant screening (Table 1.1). Put the performance of people once hired into the feedback loop. Without it the system never learns whether it selected well.
2. 2Fix the outcome variable before the predictors. Decide what 'successful performer' means in your data, such as a performance ranking or productivity measure. The passages treat the target as a factor variable for classification, so code it as a category (successful / not successful). Applicant attributes are the independent variables.
3. 3Build a conceptual model from logic, not from whatever columns exist. Choose which attributes to collect using theory, experience, and conversations with experts or seniors. The source's analogy: if sales fall, ask whether employee attitudes or product quality is the cause, then collect only data that tests that hypothesis. Apply the same discipline to which applicant attributes plausibly drive performance.
4. 4Audit the data foundation (the 'D' in DELTA): data must be accessible, consistent and high quality. Check whether recruitment records and later performance records sit in disparate HR systems. Disparate systems create redundancy and inconsistency, and they don't track data over time, which blocks longitudinal analysis and causal links between HR variables. Join applicant data to later performance records, and fix inconsistencies before modelling.
5. 5Classify the problem, then pick the tool. Predicting who will be a successful performer is a classification problem. The source demonstrates K-nearest neighbour (KNN) and artificial neural networks (ANN) in R for applicant screening. For ANN, normalise the dataset. For KNN, set the target as a factor, split the data into training and test sets, and fit the model.
6. 6Judge the model on held-out data. Read the confusion matrix and the misclassification error. Optionally re-run KNN with ROC as the performance metric, which plots sensitivity against specificity. Check variable importance to see which applicant attributes carry the prediction, and compare that to your conceptual model.
7. 7Use the model on new applicants, and don't wait for perfect data. Feinzig (IBM) advises that no dataset is perfect, that data size matters less than quality, and that you should proceed once the data fundamentals are assessed. Feed each hired cohort's actual performance back into the dataset so the model is refreshed.
8. 8Report the result as hROI. Tie the model to business objectives (the vertical integration step in DELTA's 'Enterprise') and state what better selection or performance management returned, validated with numbers.

The passages do not give a numeric quality-of-hire case, so this combines their real examples with their method. A global restaurant chain wanted to raise customer satisfaction. Correlation analysis of historical frontline-employee data against customer satisfaction rates showed a significant relationship, so it chose to improve frontline employee retention to raise revenue. It then built a predictive model on individual employee details, shift details and restaurant details. The pattern is the one to copy for hiring. First, run a correlation check to confirm the people variable moves the business outcome. Second, define the target. Third, model with individual-level and context data. Applied to selection: code 'successful performer' as the target factor, split the applicant dataset into training and test sets, and fit KNN. Read the confusion matrix and misclassification error, then re-fit with ROC as the metric to compare. The source's retirement-forecast firm shows why to go beyond the obvious predictors. It added role change, pay level, rate of change in pay and likelihood of incentives to age and tenure. Likewise, test attributes beyond the standard screening fields.

• If recruitment and performance data live in separate systems with no time stamps, then integrate and time-track them first, because without that you cannot establish causal links between selection variables and performance.
• If the outcome is a category (successful vs not), then treat it as classification, set the target as a factor, and evaluate with a confusion matrix and misclassification error. Don't use correlation alone.
• If the data is imperfect but its fundamentals check out (accessible, consistent, high quality), then proceed. Feinzig's guidance is not to stall for perfect data and to prioritise quality over size.
• If you cannot explain why an attribute should drive performance, then check it against theory, experience or expert input before including it. Variable importance should confirm the conceptual model, not replace it.
• If HR objectives are not aligned to business objectives and other functions, then fix that first. Vertical and horizontal integration is what makes the analytics yield results, and hROI needs the business outcome to measure against.

• Modelling without a defined performance outcome. The tell: you can't state what 'successful' means or which column holds it, and the model has no clear target variable.
• Building on fragmented HR systems. The tell: the same employee shows inconsistent values across systems, and you can't follow a hire's performance over time.
• Throwing every available field into the model with no logic. The tell: no one can explain the variable importance output in terms of a role-performance hypothesis.
• Waiting for a perfect dataset, or trusting a huge messy one. The tell: months of delay with no model, or a large dataset whose quality nobody has assessed. Feinzig says size doesn't matter and quality does.
• Never closing the loop or proving value. The tell: the screening model is built once, post-hire performance never flows back, and there is no hROI statement validating the HR success story with numbers.

• ▸Role / population: the job or employee group the model covers.
• ▸Process map: inputs, transformation (screening steps), outputs and feedback loop for this hiring process.
• ▸Target variable: definition of 'successful performer' (performance ranking, productivity measure) and how it is coded as a category.
• ▸Hypothesised drivers: applicant attributes you expect to predict performance, and the theory, expert or experience basis for each.
• ▸Data sources and consistency check: which systems hold applicant and performance data, whether they join, and whether they are time-tracked.
• ▸Problem type and tool: classification; KNN or ANN; normalisation and train/test split decisions.
• ▸Evaluation results: confusion matrix, misclassification error, ROC result, top variables by importance.
• ▸Business-objective link: the business objective this serves and the HR/business/functional alignment (vertical and horizontal).
• ▸Feedback and hROI plan: how new-hire performance feeds back into the dataset, and how the return will be stated in numbers.

1. 1Start with the business outcome, not the HR activity. Name the outcome the investment must move (revenue, hiring and training cost, retention rate, strategic alignment) before you collect any data. The source warns that firms quantify training's effect on individual performance but never connect it to overall business performance.
2. 2Check strategic alignment first. Ask whether the work, organisational strategy/process, structures and jobs are aligned, and whether the firm's human capital can accomplish them. If the answer is no, an ROI figure for any single programme is premature.
3. 3Build the data foundation: an employee database that tracks key employee metrics in order (the source cites Google's database of attitudes, behaviours and on-job performance). Retain the HR data you collect. The source notes that many firms collect huge amounts of HR data but few retain and quantify it.
4. 4Quantify subjective variables before linking them to outcomes. Decide whether the variable can be measured directly, list its attributes (for job satisfaction: supervisor, salary, benefits, career growth), classify each as objective or subjective, then scale subjective ones (nominal, ordinal, interval or ratio) using a rating or ranking technique.
5. 5Run the HR intervention, then track a people metric and a business metric side by side, as Harrah's does with wellness programmes: engagement metrics first, then how engagement changes correspond to firm revenue.
6. 6Express the link as a rate of change in a business figure. Best Buy's example is a 0.1% rise in engagement tied to US$100,000 more revenue per store. Sysco's is retention rising from 65% to 85%, which maps to savings in hiring and training costs.
7. 7Move from descriptive reporting to prediction. Use statistical and analytical tools to model the causes of an HR problem (e.g. turnover) and the relationship between HR practices and outcomes, so you can forecast the return and act before the loss occurs.
8. 8Report the result as tangible ROI in the unit leadership already uses: money, retention percentage, or revenue per unit. Pair each figure with a stated target on an HR scorecard so the contribution is traceable to strategy.

Sysco, a US food-products distributor, targeted immediate retention strategies at its delivery associates using analytics. Retention rose from 65% to 85%. The source states this corresponds to savings in hiring and training costs. To make the impact explicit, convert the 20-point gain into avoided exits. Multiply the delivery-associate headcount by 0.20, then multiply that count by the firm's per-hire hiring and training cost. The headcount and cost inputs are not in the source, so they must come from your own data. Harrah's follows the same chain for engagement: wellness programme, then engagement metrics, then firm revenue. Best Buy sets the revenue yardstick: 0.1% engagement is worth US$100,000 per store. Each case names one HR action, one people metric and one business figure.

• If you can only show an effect on individual performance (e.g. training scores), then do not claim business ROI. Find the store-, unit- or firm-level metric that the performance change should move, and measure that.
• If the variable is subjective (engagement, satisfaction), then break it into attributes and scale them before correlating with revenue or cost. An unquantified variable cannot carry an ROI claim.
• If retention or turnover is the problem, then express the benefit as avoided hiring and training cost, and use prediction to target the highest-risk group first, as Sysco did with delivery associates.
• If the HR function has not retained historical people data, then fix the database and tracking before promising predictive ROI. Prediction needs retained data.
• If a result links an HR metric to revenue but alignment between jobs and strategy is unclear, then report it as correlation to investigate, not as proven contribution.

• Measuring activity instead of impact: reports show training hours or programme participation but no business metric. The tell is that no figure in the report is in money, retention or revenue.
• Stopping at the individual level: performance gains are quantified per employee but never rolled up to business performance. The tell is a missing link between the HR metric and a firm or store outcome.
• Collecting but not retaining data: the firm has huge HR data volumes yet cannot reproduce last year's baseline. The tell is that a before-and-after comparison cannot be built.
• Reactive, static reporting: analysis describes past turnover rather than modelling the future, so the firm stays in the cycle of hiring and firing. The tell is a report that explains exits only after they happen.
• Unscaled subjective inputs: engagement or satisfaction is cited as an ROI driver without defined attributes or scales. The tell is that no one can state how the metric was measured.

• ▸HR investment: name the programme or action and its total cost.
• ▸Strategic objective: state the business goal it serves and confirm that jobs, structures and processes align to it.
• ▸People metric: define the metric the investment should move (e.g. retention rate, engagement score) and its attributes and scale if subjective.
• ▸Baseline value: record the people metric before the intervention, from retained employee data.
• ▸Post-intervention value: record the metric after the intervention (e.g. 65% to 85%).
• ▸Business metric: name the firm-level outcome tied to it (revenue per store, hiring and training cost, firm revenue).
• ▸Conversion rate: state the quantified link (e.g. 0.1% engagement = US$100,000 per store) or the cost per avoided exit.
• ▸Net ROI statement: compare benefit with the Section 1 cost and write a one-sentence result in money or percentage.
• ▸Predictive next step: name the model or forecast that will identify where to act next, and the target on the HR scorecard.

1. 1Start from the decision, not the dataset. Write the HR question as a business problem (e.g., 'which employees are likely to leave?' or 'which applicants should be screened in?'). Then name the outcome variable and the action the answer would change. The source treats modelling the business problem as its own step before tools: draw an influence diagram and label each variable dependent, independent, moderator or mediator.
2. 2Classify the ambition level before choosing tools. Descriptive analytics narrates what happened. Predictive forecasts future events. Prescriptive adds a recommendation. Also name the outcome type: classification, clustering or association. This fixes how much data, skill and tooling you need.
3. 3Audit the data foundation. Identify where HR data lives (a spreadsheet, an HRIS or an ERP system), since the basic premise is collecting, storing and disseminating HRM data. List the attributes you hold for the outcome. Experian's attrition model used about 200 attributes, including supervisor performance and commuting distance, so look beyond the HR master record. Decide how missing data will be collected and which sampling method applies.
4. 4Define HR metrics tied to efficiency, efficacy, impact and service. These are what HR professionals use to solve problems and assess HR. Fix definitions up front (e.g., turnover = individual movement across the organisation's membership boundary; turnover intent = probability of changing jobs within a period) so every dashboard uses the same terms.
5. 5Match technique to skill. The source demonstrates artificial neural networks (ANN) and K-nearest neighbour (KNN) in R/RStudio for turnover prediction and applicant screening. Choose one technique your team can build, explain and maintain. Decision trees are also covered. Pair an analytically trained person with an HR domain owner for each model.
6. 6Evaluate before acting. Use the evaluation chapter's tools (e.g., a lift chart) to test whether the model outperforms the current approach. Define in advance what result justifies deployment.
7. 7Wire the output into a decision and a named owner. Nielsen's model prescribes retention actions by tenure and reminds managers of new hires to make contact. The insight reaches the person who acts on it, and the action is tied to the employee life cycle.
8. 8Extend scope in stages: from a single model to workforce planning (future business requirements vs current capability, forecasting staffing gaps). Then to external stakeholders such as the employee supply chain. Secure leadership backing at each stage, because the passages tie high performance to having the right tools and technologies and move HR from back-end administration to strategic partner.

Experian (as reported in the source) predicts employee attrition by computing employee risk factors with predictive algorithms. Read it through the five components of capability. Data: about 200 attributes per employee, such as the performance of the supervisor and commuting distance, which shows integration beyond core HR fields. Technique: predictive algorithms that score each employee's risk of leaving. Outcome type: forecasting a future event, which makes it predictive rather than descriptive. Action: the firm reduced turnover by 2–3% over one year of implementation, and the source ties the value to the result, not the model. Compare Dow Chemicals, which mines historical data on about 40,000 employees to forecast promotion rates, internal transfers and labour count required. Nielsen goes a step further into prescriptive use. For first-year hires, its model says critical contact points matter for retention, and managers are reminded to reach out. The source excerpt does not give Experian's algorithm, tooling or team, so do not assume them. Copy the structure: a rich attribute set, a defined forecast, and an action linked to the score.

• If you can only describe what happened, call it descriptive and fix the data and metric definitions first. Do not label it predictive until the model forecasts a future event.
• If a prediction has no named action and owner (as with Nielsen's tenure-based manager reminders), treat it as unfinished and move to a prescriptive output.
• If your data lives only in a spreadsheet, build the metric set and collection discipline first. Move to an HRIS/ERP with decision-support and BI features once the questions repeat.
• If the team cannot explain the technique (ANN, KNN, decision tree), choose the one it can build and maintain in R and document it before extending.
• If a model does not beat current practice on the evaluation (e.g., lift chart), do not deploy it. Add attributes (e.g., supervisor or commute data) or revisit the problem model.

• Starting with a tool or dashboard instead of a business problem. The tell: no dependent variable is named and no decision changes after the report.
• Modelling only on narrow core HR fields. The tell: a thin attribute set, whereas Experian's turnover model used about 200 attributes.
• Stopping at prediction. The tell: a risk score exists but managers receive no prompt or prescribed action, unlike Nielsen's tenure-based reminders.
• Treating HRIS as storage rather than a decision-support asset. The tell: data is collected and stored but no HR metrics, DSS or BI feed problem-solving, so HR stays a back-end administrative function.
• Applying one action to every employee. The tell: uniform interventions despite segment and life-cycle differences. Convergys found that paying increments every six months rather than annually improved retention, which shows timing and segment matter.

• ▸Business problem: one sentence stating the HR question and the decision it informs.
• ▸Analytics level: descriptive, predictive or prescriptive, and outcome type (classification, clustering or association).
• ▸Variables: dependent variable plus independent, moderator and mediator variables from the influence diagram.
• ▸Data sources and system: where each attribute lives (spreadsheet, HRIS, ERP) and what is missing.
• ▸HR metric definitions: how each metric is calculated (e.g., turnover, turnover intent) and who owns it.
• ▸Technique and tool: e.g., ANN, KNN or decision tree in R/RStudio, and who builds and maintains it.
• ▸Evaluation test: the measure (e.g., lift chart) and the threshold for deployment.
• ▸Action and owner: the intervention, the manager or role who receives the insight, and its timing in the employee life cycle.
• ▸Leadership sponsor and next scope step: the sponsor, and whether the use case extends to workforce planning or the external supply chain.

1. 1Start from the business problem, not the dataset. Define the problem first, then list the variables the model needs. The passages say to understand and define the problem thoroughly and only then model it. Source those variables from experience, expert conversation, or theory.
2. 2Map each HR function as a system with defined inputs, processes and outputs. For payroll, that means work data (absent days, performance rate), then employee record (base pay, benefits), then salary computation, then payroll generation. Record which system holds each data element, so you can see where gaps and overlaps sit.
3. 3Audit the DELTA 'data' standard on each source: accessible, consistent, high quality. Test the source against the six barriers named in the passages: inadequate, incomplete, inaccessible, not integrated, outdated, inconsistent. Mark each field pass or fail on each test.
4. 4Check for redundancy and conflict across disparate HR IT systems. Compare the same employee attributes in each system and log the mismatches. The passages state that disparate systems multiply redundancy and inconsistency.
5. 5Check time-stamping. Confirm each system keeps dated history of values (role, pay level, pay change). Without time tracking you cannot run longitudinal analysis or test causal relationships between HR variables, so flag those fields as non-modelable until fixed.
6. 6Identify inter-functional data you need (for example sales or customer satisfaction) and secure access early. Build the collaboration with the owning function, because HR managers need it to reach that data.
7. 7Quantify variables so they can be modelled. Convert qualitative constructs (such as employee attitudes) into measured data, and label each variable as dependent or independent.
8. 8Assess data fundamentals, then move. Take expert advice on quality, fix the critical gaps, and proceed. No dataset is perfect and size matters less than quality, so do not hold the analysis for perfection.

A firm builds a retirement-prediction model. The conventional inputs are age and tenure. The firm adds changes in employee role, pay level, rate of change in pay, and likelihood of receiving incentives. Those extra factors are only usable if the data meets the quality tests above. Role and pay need dated history, so the change and its rate can be computed. That requires a system that tracks values over time, which disparate systems often do not. Pay, role and incentive records must also agree across the systems that hold them. Retirement behaviour is shifting toward part-time and contractual work, so employment-type records must be complete and current. With these checks passed, the firm could forecast who is likely to retire and manage the retirement cycle effectively. If role or pay history were missing, the model would fall back to age and tenure alone.

• If the same attribute differs across two HR systems, then fix or designate a system of record before modelling, since inconsistency grows with the number of disparate systems.
• If a variable has no time-stamped history, then exclude it from longitudinal or causal analysis and flag it for system change.
• If the data is imperfect but the fundamentals (accuracy, completeness, accessibility) check out, then proceed, because no dataset is perfect.
• If the model needs data owned by another function and you lack access, then secure that access and collaboration before building.
• If you are tempted to collect more data to compensate for quality problems, then spend the effort on quality instead, since size matters less than quality.

• Trusting the HRIS because its purpose is to store data cleanly. The tell: data is in fact inadequate, incomplete, inaccessible, not integrated, or outdated when you inspect it.
• Merging records from disparate systems without reconciling them. The tell: duplicate or conflicting values for the same employee and unexplained differences between reports.
• Attempting causal claims on snapshot data. The tell: no historical values exist, so you cannot establish how HR variables changed relative to outcomes.
• Waiting for perfect data. The tell: the project stalls on cleaning while no model is built, contradicting the guidance that analysts should keep moving.
• Collecting data before defining the problem. The tell: you hold many fields but cannot say which attitudes or variables link to the outcome.

• ▸Business problem: state the decision or question the model must support.
• ▸Variables required: list each dependent and independent variable needed.
• ▸Source system: name the system or function holding each variable.
• ▸Accessible? (Y/N): note who owns it and what access is needed, including inter-functional data.
• ▸Accurate and consistent? (Y/N): compare against other systems and list the conflicts found.
• ▸Complete and current? (Y/N): record missing, outdated or inadequate fields.
• ▸Time-stamped history? (Y/N): confirm dated values exist for longitudinal analysis.
• ▸Quantified? (Y/N): state how each qualitative construct is measured.
• ▸Fix or proceed: record the action, the owner, and an expert consulted on data quality.

1. 1Start from the business problem, not the practice. State it as one measurable outcome, such as turnover in a unit, low store sales, or customer satisfaction. The source's retailer example moves from low sales to employee performance only after product pricing is ruled out as a constant.
2. 2Map the candidate drivers with an influence diagram before building any model. List the HR practices you currently run (compensation, training for client handling, career planning, engagement activities, retention measures) beside non-practice factors such as store ambience, supervisor rapport and organisational climate. Draw how each is thought to affect the outcome.
3. 3Pull the employee records you already hold: tenure, performance ratings, skill levels, previous experience, promotion history, supervisor, and the interventions that occurred in each area. The source's Indian IT turnover work goes beyond exit interviews and ties each of these fields to turnover intent.
4. 4Test the practices against the outcome with correlation or a predictive model, not intuition. Intuition says hire; a model says how many. Cross-check the output against your conceptual (influence) model before you trust it.
5. 5Rank the practices by measured effect. Do not assume the costliest one matters most. In the restaurant case, performance depended on career development opportunities and cultural norms more than on the variable compensation being offered.
6. 6Convert findings into an action plan that names who does what. Involve all stakeholders in writing it, and build resistance and change-management steps into the plan. A diagnosis has no value until treatment is prescribed.
7. 7Implement, then measure the same outcome again, tracking it as hROI: the investment in the HR practice against the result. Record any new question the result raises, such as pay dissatisfaction in one region, as the next analysis.

A firm had high turnover even though it offered the best compensation and benefits package. Pay was therefore not the explanation, and the firm could not name a cause. It correlated all the organisational and employee data it held and found three patterns. Leavers had not been promoted for a long time. Most had worked under a few common supervisors. Few organisational interventions had reached the areas where turnover was highest. Each pattern points to a different HR practice: career planning, supervisor development, and engagement or retention activity. Compensation was already maxed out and was not the lever. The source notes that intuition or judgement would never have surfaced these concerns. The sensible next move is to target promotion paths and the identified supervisors' units, then track turnover in those areas.

• If an outcome stays poor despite strong spending on one practice (for example, top pay with high turnover), then stop adding to that practice and test other drivers: promotion history, supervisor, climate, interventions.
• If the data shows the outcome depends on career development and culture more than on variable pay, then shift budget toward those, as the restaurant chain did, and re-measure.
• If analysis traces a problem to a single factor like pay dissatisfaction in one region, then treat that as the next analysis question rather than a final answer.
• If the question is how many people to hire or train, then use a quantitative model; use intuition only for the direction (hire versus train).
• If stakeholders resist the plan, then add communication and change-management steps; the source names lack of communication as the main reason adoption fails.

• Spending more on the visible practice (pay, benefits) as the default retention answer. The tell: turnover or performance does not move after raising it, and no other drivers have been tested.
• Choosing interventions from managerial intuition alone. The tell: the rationale is 'it feels like' and no data on tenure, promotions or supervisors was examined.
• Stopping at the analysis result with no owner or action. The tell: a report with findings but no statement of who does what and by when.
• Writing the action plan without the people who must execute it. The tell: resistance appears at rollout and communication gaps are cited afterwards.
• Measuring activity rather than outcome. The tell: you can report training or engagement sessions delivered but not their effect on revenue, customer satisfaction or attrition (hROI).

• ▸Target outcome: one measurable problem (e.g., turnover in unit X, low store sales).
• ▸Ruled-out factors: items held constant or eliminated (e.g., pricing, pay level) and why.
• ▸HR practices in play: compensation, training and development, career planning, engagement, retention measures currently running.
• ▸Other candidate drivers: supervisor, climate, promotion history, interventions received, tenure, skill level.
• ▸Data available: records and fields you can pull, and any gaps.
• ▸Finding: what the correlation or model shows, ranked by strength.
• ▸Practice to change: the intervention chosen and the evidence behind it.
• ▸Action owner and stakeholders: who does what, plus resistance and communication steps.
• ▸Re-measure metric and date: the outcome or hROI figure you will check, and any follow-up question raised.

1. 1Define the outcome first. State the business problem, then name the target variable you can measure. In the passages' transportation firm, the goal is lower road-accident cost, but the dependent variable is the likelihood of a driver getting into an accident.
2. 2Define 'right people' for this specific role and organisation. The passages say the definition varies by job profile, organisation and industry, and can be framed as KSAs, psychometric traits or prior experience. Write down which of these you are testing.
3. 3Split your existing employees into good and bad performers at each position. Use measurable outcomes such as job satisfaction, performance ranking and grades. This gives the model its criteria for classifying a good versus a bad hire.
4. 4Select individual-difference variables as candidate predictors on theoretical grounds. The passages cite theory that traits drive behaviour and predict job satisfaction, job performance and work behaviours. Include the traits you can defend with that logic rather than everything available.
5. 5Record each trait as an independent variable, with a defined scale. The driver example uses concentration, motor speed, reaction speed, stress tolerance and tracking deviation. Note whether each is nominal, ordinal, interval or ratio, since that determines which statistics are valid (means and standard deviations need at least interval data).
6. 6Check for moderators. A moderator is any variable that changes the trait-outcome relationship. If you ignore one, the model's implications can become largely irrelevant once it comes into play, so list candidates such as role type before fitting.
7. 7Fit the predictive model on current employees' traits and outcomes, then score applicants. Applicants' attitudes and outcomes cannot be observed yet, which is exactly why you predict them from traits.
8. 8Apply the causal sanity checks before acting. Control for other independent variables when interpreting one trait's effect, and confirm a theoretical framework supports each relationship. Drop any trait with no logical link to the outcome.

A transportation firm wanted to cut the cost of road traffic accidents among new drivers, so it redesigned its hiring assessment. The business goal was fewer accidents and better hires, but the dependent variable it modelled was the likelihood of a driver getting into a road accident. The independent variables were psychometric traits: concentration, motor speed, reaction speed, stress tolerance and tracking deviation. Each trait is a separate predictor, and each is justified by a plausible link to driving safety. The firm builds the model on drivers whose accident outcomes are known, then applies it to applicants, whose outcomes cannot yet be observed. Applicants are screened on predicted accident likelihood rather than on a hiring manager's impression. The passages do not report results, so treat this as the model design, not evidence of effect size.

• If the outcome cannot be measured for applicants, then predict it from traits and validate against current employees' known outcomes.
• If a trait has no theoretical link to the outcome (the passages' income-obesity example), then exclude it even if it correlates.
• If the relationship may change by context, then test for a moderator before generalising across roles.
• If the role or organisation differs, then rebuild the 'right person' profile instead of reusing another firm's trait set.
• If you want means or standard deviations of a trait, then confirm it is measured on at least an interval scale; use ordinal data only for rank claims.

• Modelling the business goal instead of a measurable dependent variable. The tell: your target reads 'better hires' rather than something like accident likelihood or performance ranking.
• Throwing in every available trait without theory. The tell: you cannot state in one sentence why a trait should affect the outcome.
• Ignoring moderators. The tell: the model works for one role or group and fails when applied to another.
• Treating Likert scores as having true magnitude. The tell: you claim a 5 versus 4 gap equals a 'real' difference; the passages note interval scales have no origin point.
• Importing a generic 'right person' profile. The tell: the profile was not derived from your own high performers in that job.

• ▸Business problem: state the organisational cost or goal in one line.
• ▸Dependent variable: the measurable outcome you will predict (e.g. accident likelihood, performance ranking).
• ▸Good/bad performer criterion: how existing employees at this position are classified.
• ▸Candidate traits (independent variables): list each trait you will measure.
• ▸Scale type per trait: nominal, ordinal, interval or ratio.
• ▸Theory link: one sentence justifying why each trait affects the outcome.
• ▸Possible moderators: variables that might change the trait-outcome relationship.
• ▸Controls: other variables held constant when interpreting each trait's effect.
• ▸Applicant scoring plan: how predicted outcomes will inform screening decisions.

1. 1Define the fit question as a business problem before touching data. Write one sentence naming the decision (for example, which applicants to shortlist for a role) and the outcome fit is supposed to improve (for example, accident proneness, turnover or performance). The source says a clearly defined problem removes half the confusion about method.
2. 2Map the hiring or retention process as a system: inputs (applicant data, test results, KSA), transformation (screening or selection), outputs (hire, performance, stay or leave), and the feedback loop. This shows where fit data enters and where the outcome is recorded.
3. 3Split fit into measurable components. Person-job fit uses knowledge, skills and abilities (KSA) plus job-relevant traits such as the psychometric scores in the driver example. Person-organisation fit uses values and personality (the source lists MBTI as a personality instrument). Record which variable measures which component.
4. 4Audit data quality before modelling. Check that predictor data (test scores, KSA, profile data) and the outcome data (performance, accidents, tenure) exist for the same people and are reliable. The source cites Deloitte 2017, in which only 39% of business people rated their people data 'good' or 'very good'.
5. 5Run descriptive analytics first. Profile the traits of people who perform well in the role, as Google does with its employee database, and test which trait scores relate to the outcome, as the transport company did with concentration and reactive stress tolerance. Keep only the traits that show a relationship.
6. 6Build a predictive screen on the surviving traits. The source demonstrates KNN and ANN in R for applicant screening and turnover. Use the descriptive findings to choose inputs, and fix the outcome label (fit versus non-fit) in advance.
7. 7Validate before use. Read the confusion matrix and accuracy. For classification, compute ROC and AUC (via twoClassSummary in caret), noting the source's rule that AUC near 1 is good and below 0.7 generally poor. For KNN, choose K using ROC as the criterion if needed.
8. 8Turn the result into a hiring rule and close the loop. Apply the validated traits in screening, then feed new hires' outcomes back into the dataset so the fit definition is re-tested against reality.

The source describes a local transport company that applied descriptive analytics to psychometric test results of its drivers. It found that concentration and reactive stress tolerance were related to proneness to accidents. That is person-job fit made operational: the job demand (safe driving) is linked to measurable personal attributes, and the finding has direct implications for hiring drivers. Following the steps above, the decision is which driver applicants to shortlist, and the outcome is accident proneness. The inputs are the psychometric scores and the outcome is each existing driver's accident record. Descriptive analysis shows which scores relate to accidents. A KNN or ANN screen could then be trained on those scores and checked with a confusion matrix and AUC, where a result below 0.7 would send the team back to the data. The source gives no figures for this case, so any thresholds for cut-off scores must come from the company's own validated results. A parallel example is Google, which describes the traits of key performers in each area and builds hiring plans on them.

• If you cannot name the outcome that fit is supposed to improve (accidents, turnover, performance), then do not build a fit model; define the problem first.
• If a trait shows no relationship to the outcome in descriptive analysis, then drop it from the screen, however plausible it seems.
• If the model's AUC is below about 0.7, then treat it as poor and do not use it for selection decisions; revisit the data and variables.
• If predictor or outcome data are missing or unreliable for the same people, then fix the data first; do not model around the gap.
• If you are screening on traits drawn from top performers in one area, then build separate profiles per role or area, as Google does, rather than one organisation-wide profile.

• Treating fit as a vague impression of 'cultural match'. The tell: nobody can state which variable measures it or which outcome it predicts.
• Using psychometric or personality scores (e.g. MBTI) without testing them against a performance or safety outcome. The tell: the tests are in the hiring file but no analysis ties them to results.
• Skipping the data-quality check. The tell: the model is built on partial or inconsistent people records, with no confirmation that outcome data exist for the scored individuals.
• Judging a screening model by accuracy alone. The tell: no confusion matrix, ROC or AUC is reported, so you cannot see where non-fit hires are misclassified.
• Never feeding hire outcomes back into the dataset. The tell: the fit profile is unchanged year after year while the job and workforce change.

• ▸Decision: the specific selection or retention decision this fit analysis supports.
• ▸Outcome variable: the measurable result fit should improve (e.g. accidents, turnover, performance) and its data source.
• ▸Person-job fit variables: KSA and job-relevant trait scores, with the instrument for each.
• ▸Person-organisation fit variables: values and personality measures, with the instrument for each.
• ▸Data quality check: are predictors and outcomes available and reliable for the same individuals? Note gaps.
• ▸Descriptive finding: which variables relate to the outcome, and which are dropped.
• ▸Model and validation: method used (e.g. KNN, ANN), confusion matrix result, AUC, and K if relevant.
• ▸Go/no-go: apply to screening only if AUC is at or above 0.7; record the decision and reason.
• ▸Feedback plan: how and when new hire outcomes are added to the dataset for re-testing.