Predictive Analytics Training Course

Participants can assess analytical data quality and prepare a valid modelling dataset.

• Profile data types, ranges, missingness, duplicates and temporal coverage
• Build a data dictionary covering source, owner, grain and refresh frequency
• Identify target leakage using data availability dates and process knowledge
• Treat missing values with deletion, imputation indicators and domain rules
• Encode categorical variables using one-hot, ordinal and target encoding
• Scale numerical features where distance-based and regularised models require it
• Split data by time, entity and outcome availability

Participants can derive and document predictive features that are available at prediction time.

• Inspect distributions with histograms, box plots and quantile summaries
• Measure associations using correlation, contingency tables and mutual information
• Engineer recency, frequency and monetary features from transaction histories
• Create lag, rolling-window and rate-of-change predictors from dated records
• Handle skewed variables through logarithmic transformations and winsorisation thresholds
• Document feature lineage, definitions and permissible prediction-time inputs

Participants can build, diagnose and compare regression models for continuous outcomes.

• Specify linear regression assumptions for continuous operational outcome prediction
• Fit multiple linear regression models with Python and scikit-learn
• Interpret coefficients, confidence intervals and residual diagnostic plots
• Compare MAE, RMSE and MAPE against business error tolerance
• Apply Ridge and Lasso regularisation to correlated predictor sets
• Generate prediction intervals and communicate expected estimation uncertainty
• Prioritise forecasting accuracy without confusing prediction with explanation

Participants can develop classification models and translate probability scores into decisions.

• Formulate binary and multiclass targets from operational decision criteria
• Estimate probabilities with logistic regression and calibrated classification models
• Read confusion matrices, precision, recall, specificity and F1 scores
• Plot ROC and precision-recall curves for imbalanced event detection
• Set decision thresholds using false-positive and false-negative costs
• Address class imbalance through weighting, resampling and stratified splitting
• Assess calibration with reliability curves and Brier scores

Participants can compare tree-based models and explain their predictions to stakeholders.

• Build decision trees using split criteria, depth limits and pruning
• Compare random forests and gradient boosting for nonlinear prediction
• Tune hyperparameters with grid search, random search and cross-validation
• Control overfitting through learning curves, validation curves and early stopping
• Rank predictors using permutation importance rather than impurity alone
• Explain individual predictions with SHAP values and partial dependence plots

Participants can select a model using valid tests and operational selection criteria.

• Design holdout, k-fold and time-series cross-validation schemes for candidate models
• Keep preprocessing inside training folds to prevent information leakage
• Use nested cross-validation when comparing tuned model candidates
• Benchmark candidate models against naive baselines and existing rules
• Quantify uncertainty with bootstrap confidence intervals for performance metrics
• Record experiments, parameters, data versions and evaluation results
• Select models using accuracy, interpretability, latency and maintenance criteria

Participants can produce and evaluate forecasts for time-dependent operational decisions.

• Distinguish trend, seasonality, cycles and irregular variation in time series
• Create chronological training, validation and test partitions before feature generation
• Construct autoregressive and seasonal features from historical observations
• Fit exponential smoothing and ARIMA models for short-term forecasts
• Evaluate forecasts with rolling-origin validation and horizon-specific error measures
• Reconcile forecast outputs with staffing, inventory and budget decisions

Participants can document and govern predictive models for accountable operational use.

• Identify fairness risks from proxy variables, sampling and label bias
• Assess subgroup performance using selection rates, error rates and calibration
• Apply data minimisation, purpose limitation and retention requirements
• Document model purpose, inputs, limitations, approvals and accountable owners
• Establish human review points for consequential automated decisions
• Prepare explanations appropriate for executives, operators, auditors and affected individuals

Participants can specify how a predictive model will be deployed, monitored and reviewed.

• Define deployment requirements for batch scoring, real-time scoring and alerts
• Specify input validation, output actions, service levels and escalation paths
• Monitor data drift, concept drift, performance decay and missing inputs
• Set retraining triggers, review cadence and model retirement criteria
• Create a model card with intended use and known limitations
• Draft a 90-day predictive analytics application plan for one decision