Predictive Modelling with Python Training Course

Participants can inspect and prepare source data for predictive modelling.

• CSV, Excel, JSON and database extracts loaded with pandas
• Data dictionaries reviewed for grain, ownership, definitions and refresh frequency
• Schema checks identify invalid types, duplicates and impossible values
• Missingness patterns profiled using counts, proportions and visual inspection
• Outliers investigated with quantiles, boxplots and domain thresholds
• Analytical datasets assembled using documented joins and stable record keys

Participants can use exploratory analysis to identify predictive signals and data risks.

• Target distributions examined before selecting transformations or modelling approaches
• Numeric predictors profiled with histograms, quantiles and grouped summaries
• Categorical variables assessed through frequency tables and target rates
• Correlation matrices reveal redundancy, multicollinearity and unexpected relationships
• Segment analysis compares outcomes by customer, product and channel groups
• Temporal plots expose seasonality, drift and potential future-data contamination

Participants can create, transform and select features without contaminating model inputs.

• Date features derived from recency, duration, weekday and seasonal patterns
• Ratios, rates and aggregated measures constructed from transactional records
• Categorical encoding using one-hot, ordinal and target encoding methods
• Numeric transformations apply logarithms, binning, scaling and winsorisation appropriately
• Feature leakage identified through timing analysis and target-proxy inspection
• scikit-learn ColumnTransformer structures repeatable mixed-data preprocessing pipelines

Participants can build and interpret baseline regression models for continuous outcomes.

• Baseline forecasts establish minimum performance before complex model development
• Linear regression estimates continuous targets using interpretable coefficient structures
• Residual diagnostics assess bias, heteroscedasticity and influential observations
• Mean absolute error and root mean squared error compared
• R-squared interpreted alongside error measures and business tolerance ranges
• Regularisation with Ridge, Lasso and Elastic Net controls coefficient instability

Participants can construct binary classification models and interpret predicted probabilities.

• Binary targets defined using clear events, windows and observation cut-offs
• Logistic regression estimates event probabilities from structured predictor variables
• Odds ratios and marginal effects support stakeholder interpretation
• Classification thresholds selected from costs, capacity and risk appetite
• Confusion matrices distinguish true positives, false positives and missed cases
• Multiclass classification extends probability modelling to several outcome categories

Participants can design validation processes that produce credible estimates of model performance.

• Holdout samples protect final performance estimates from repeated model selection
• K-fold cross-validation estimates variability across alternative training partitions
• Stratified folds preserve event prevalence during classification model assessment
• TimeSeriesSplit respects chronological order for demand and risk models
• Nested cross-validation separates tuning decisions from performance estimation
• Validation protocols documented with sample dates, exclusions and random seeds

Participants can assess classification models using discrimination, calibration and decision metrics.

• ROC curves compare ranking performance across classification thresholds
• Precision-recall curves assess rare-event detection more informatively than accuracy
• AUC, precision, recall and F1 matched to decision priorities
• Lift charts quantify targeting value for marketing and collections campaigns
• Calibration curves compare predicted probabilities against observed event rates
• Brier scores measure probability quality beyond classification correctness

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

• Decision trees partition predictor space into interpretable decision rules
• Gini impurity and entropy guide classification split selection
• Tree depth, minimum samples and pruning limit overfitting
• Random forests reduce variance through bootstrap aggregation and feature sampling
• Out-of-bag estimates provide internal performance checks for forests
• Feature importance assessed critically for correlated predictor sets

Participants can tune gradient boosting models using controlled search strategies.

• Gradient boosting combines sequential weak learners to reduce residual error
• Learning rate, estimators and tree depth control boosting behaviour
• Grid search evaluates defined parameter combinations against validation metrics
• Randomised search samples wider parameter spaces with limited computing budget
• Early stopping prevents unnecessary iterations when validation performance plateaus
• Champion-challenger comparisons record gains, trade-offs and model selection rationale

Participants can build classification models for rare events without relying on misleading accuracy.

• Event prevalence quantified before selecting sampling and evaluation strategies
• Class weights adjust learning objectives for asymmetric misclassification costs
• Random undersampling reduces majority-class dominance while preserving evaluation samples
• SMOTE creates synthetic minority examples within training data only
• Precision at k supports capacity-constrained investigation and intervention queues
• Threshold policies align alert volumes with available operational resources

Participants can create and assess customer or operational segments using clustering methods.

• K-means clustering groups records around centroids using scaled inputs
• Elbow plots and silhouette scores inform plausible cluster counts
• Hierarchical clustering reveals nested relationships between customer segments
• Principal component analysis reduces dimensionality for visual diagnostic purposes
• Cluster profiles combine descriptive statistics, behaviours and business labels
• Segmentation stability checked across samples, periods and initialisation seeds

Participants can prepare time-based features and validation schemes for predictive forecasting tasks.

• Lag variables capture prior demand, balances, usage and event history
• Rolling windows calculate moving averages, counts and recent trend measures
• Calendar effects represent holidays, month-end cycles and reporting periods
• Walk-forward validation tests models using successive historical cut-off dates
• Forecast horizons aligned with ordering, staffing and intervention decisions
• Temporal leakage prevented through point-in-time feature construction rules

Participants can explain global and local model behaviour using appropriate interpretation methods.

• Permutation importance measures performance loss after predictor value disruption
• Partial dependence plots show average predictor effects on predictions
• Individual conditional expectation reveals variation hidden by average effects
• SHAP values attribute individual predictions to contributing feature values
• Local explanations reviewed for plausibility, stability and decision relevance
• Interpretation limits documented for correlated features and extrapolated observations

Participants can identify fairness concerns and document model risks before deployment.

• Protected characteristics and proxy variables identified during data review
• Selection bias assessed through population coverage and data collection processes
• Demographic parity compares positive prediction rates between defined groups
• Equal opportunity compares true positive rates across relevant groups
• Human review requirements defined for high-impact or uncertain predictions
• Model risk registers record assumptions, limitations, owners and mitigation actions

Participants can package preprocessing and modelling steps into reusable Python workflows.

• Pipeline objects combine transformations and estimators within one fitted workflow
• Custom transformers standardise repeatable business rules and feature logic
• Train-test separation maintained throughout preprocessing and parameter fitting
• Joblib serialisation stores fitted pipelines for controlled reuse
• Configuration files separate model settings from executable notebook code
• Code reviews check readability, assumptions, tests and reproducibility controls

Participants can define deployment controls and monitoring measures for predictive models.

• Batch scoring and real-time scoring matched to operational decision timing
• Input data checks detect schema changes, null spikes and invalid ranges
• Population stability index tracks shifts in predictor distributions over time
• Performance monitoring compares realised outcomes with expected model quality
• Retraining triggers specify drift thresholds, review dates and approval roles
• Model documentation records purpose, inputs, outputs, limitations and ownership

Participants can present predictive model evidence in terms relevant to decision makers.

• Value cases connect model outputs to revenue, cost, risk and capacity
• Intervention policies translate scores into actions, queues and service rules
• Model cards summarise intended use, performance, limitations and fairness findings
• Executive visuals communicate lift, calibration, uncertainty and key drivers
• Scenario analysis estimates outcomes under alternative thresholds and resource limits
• Stakeholder challenge sessions test assumptions, operational feasibility and accountability

Participants can build, review and present a complete predictive modelling solution with documented recommendations.

• Problem statement defines decision owner, target, intervention and success criteria
• Data preparation pipeline produces a documented analytical modelling dataset
• Feature set applies timing rules, transformations and leakage controls
• Candidate models compared using cross-validation and appropriate decision metrics
• Selected model assessed for calibration, explanation, fairness and operational risks
• Model card records assumptions, performance evidence and monitoring requirements
• Presentation defends recommendations, threshold policy and implementation plan to reviewers