Machine Learning and Predictive Modelling Training Course

Participants can configure a reproducible Python environment for predictive modelling work.

• Python environments managed with virtual environments and dependency files
• Project structure separating raw data, code, outputs and documentation
• Pandas data frames for loading, joining and profiling tabular datasets
• Scikit-learn pipelines for repeatable preprocessing and model fitting
• Random seeds, version control and reproducible experiment settings
• Notebook outputs converted into reviewable modelling reports

Participants can assess whether a dataset is suitable for a stated prediction task.

• Data dictionaries linking fields to operational definitions and source systems
• Profile missingness patterns, duplicate records and invalid value ranges
• Assess target quality, label delay and historical outcome completeness
• Identify selection bias from excluded populations and process changes
• Timestamp audits preventing post-outcome information entering training data
• Data quality findings recorded in a modelling readiness assessment

Participants can create leakage-resistant preprocessing and feature engineering pipelines.

• Imputation strategies selected by variable type and missingness mechanism
• One-hot encoding, ordinal encoding and target encoding trade-offs
• Scaling numerical variables for regularised and distance-based algorithms
• Feature transformations using log, ratio, lag and rolling-window measures
• Pipeline fit stages restricted to training partitions only
• Feature lineage documented from source field to model input

Participants can identify useful predictive signals without contaminating validation evidence.

• Univariate distributions assessed using quantiles, skewness and outlier rates
• Segment target rates by customer, product, geography and time
• Correlation analysis checked for multicollinearity and proxy variables
• Visualise nonlinear relationships using bins, splines and partial plots
• Compare training and holdout distributions before model fitting
• Candidate features prioritised by availability, stability and decision relevance

Participants can build and evaluate regression models for continuous business outcomes.

• Linear regression assumptions tested with residual and influence diagnostics
• Regularisation using ridge, lasso and elastic net penalties
• Mean absolute error, RMSE and R-squared compared appropriately
• Transform skewed targets using logarithmic and Box-Cox approaches
• Prediction intervals distinguished from confidence intervals for coefficients
• Residual patterns translated into feature and segmentation improvements

Participants can build classification models and convert probabilities into decision scores.

• Logistic regression estimates interpreted through odds ratios and marginal effects
• Classification thresholds chosen using costs, capacity and risk appetite
• Confusion matrices analysed by false positive and false negative consequences
• ROC curves and precision-recall curves compared for class prevalence
• Probability outputs converted into ranked queues and treatment bands
• Scorecard-style features reviewed for stability and stakeholder acceptance

Participants can train tree-based models and control their complexity.

• Decision tree splits selected using Gini impurity and entropy
• Pruning controls using depth, leaf size and complexity parameters
• Random forests reduce variance through bootstrap samples and feature subsampling
• Out-of-bag error used for preliminary random forest assessment
• Permutation importance compared against impurity-based feature importance
• Tree models inspected for unstable splits and fragmented segments

Participants can build gradient boosting models for high-performing tabular predictions.

• Gradient boosting fits successive trees to residual prediction errors
• XGBoost parameters control learning rate, depth and subsampling
• Early stopping prevents unnecessary boosting rounds and overfitting
• Handle missing values and nonlinear interactions within boosting models
• Compare XGBoost results with linear and random forest baselines
• Model complexity balanced against latency, explainability and maintenance

Participants can select validation designs that provide credible estimates of model performance.

• Train-validation-test splits preserve independence between performance estimates
• Stratified k-fold cross-validation maintains minority class representation
• Time-based splits prevent future observations contaminating historical predictions
• Grouped cross-validation protects against repeated customer or account records
• Nested cross-validation separates tuning decisions from final performance estimates
• Performance uncertainty reported using fold variation and confidence intervals

Participants can improve classification decisions under class imbalance and uncertain probabilities.

• Class imbalance diagnosed through prevalence, event counts and sampling history
• Class weights and resampling methods compared against business costs
• SMOTE limitations assessed for rare events and temporal datasets
• Calibration curves reveal gaps between predicted and observed probabilities
• Platt scaling and isotonic regression calibrate classification scores
• Lift charts quantify prioritisation value for constrained intervention capacity

Participants can design and assess forecasting models for time-indexed operational data.

• Forecast horizons aligned with replenishment, staffing and planning decisions
• Trend, seasonality and calendar effects extracted from historical series
• Lag features and rolling statistics constructed without future leakage
• ARIMA benchmarks compared with supervised machine learning forecasts
• Walk-forward validation estimates performance across successive forecast origins
• Forecast bias and interval coverage monitored by product segment

Participants can apply unsupervised methods to segment populations and flag unusual records.

• K-means clustering evaluated using inertia, silhouette scores and stability
• Hierarchical clustering supports interpretable segment structure and profiling
• Standardise variables before distance-based clustering and anomaly methods
• Isolation Forest identifies unusual observations without labelled outcomes
• Anomaly thresholds set using investigation capacity and false alert costs
• Cluster profiles converted into operational hypotheses and treatment options

Participants can explain model behaviour and conduct an initial fairness review.

• Global feature importance identifies dominant drivers of model predictions
• Partial dependence plots show average nonlinear feature-response relationships
• SHAP values explain individual predictions and local feature contributions
• Counterfactual explanations test actionable changes under model constraints
• Fairness metrics compare selection rates and error rates across groups
• Proxy discrimination risks documented with mitigation and approval actions

Participants can tune model parameters systematically and assess ensemble alternatives.

• Grid search, random search and Bayesian optimisation compared
• Search spaces bounded by computational budget and model behaviour
• Cross-validated tuning scores tracked with parameter configurations
• Stacking combines diverse base models through meta-learners
• Blending and voting ensembles assessed against simpler alternatives
• Champion-challenger comparisons support controlled model replacement decisions

Participants can specify an appropriate deployment pattern for a predictive model.

• Batch scoring supports periodic prioritisation, planning and campaign decisions
• Real-time APIs support transaction decisions with defined latency requirements
• Feature stores reduce training-serving inconsistency across model releases
• Input validation checks schemas, ranges and required predictor fields
• Model artefacts packaged with preprocessing logic and version identifiers
• Release criteria define approval evidence, rollback triggers and ownership

Participants can design monitoring controls for model performance after release.

• Data drift monitored through population stability index and distribution tests
• Concept drift identified through declining outcome accuracy and calibration
• Operational metrics track latency, failures, coverage and scoring volumes
• Delayed labels incorporated into scheduled performance review cycles
• Retraining triggers distinguish data changes from genuine model deterioration
• Monitoring dashboards assign thresholds, escalation routes and accountable owners

Participants can document model risk controls for review, approval and audit.

• Model cards record purpose, users, limitations and intended decisions
• Datasheets document dataset provenance, collection methods and known gaps
• Validation reports capture methodology, results, exceptions and approvals
• Access controls protect sensitive data, features and model outputs
• Human review checkpoints address high-impact or contested decisions
• Model inventories track versions, owners, review dates and dependencies

Participants can plan delivery of a predictive modelling solution from discovery through operation.

• Use-case prioritisation scores value, feasibility, risk and time-to-impact
• Delivery roadmaps sequence data access, modelling, testing and release activities
• RACI assignments clarify business, data, technology and risk responsibilities
• Acceptance criteria combine model metrics with operational decision measures
• Pilot designs define population, duration, comparison group and success metrics
• Implementation risks logged with mitigations, dependencies and escalation owners

Participants can build, review and present a complete predictive modelling package for a defined business decision.

• Frame a decision-focused problem using targets, costs and intervention constraints
• Prepare a reproducible feature pipeline from documented source data
• Train and compare baseline, linear and gradient boosting models
• Validate performance using appropriate splits, metrics and calibration evidence
• Explain key predictions, fairness findings and operational model limitations
• Present deployment, monitoring, governance and implementation recommendations