Data Science with Python Training Course

Participants can configure and use a reproducible Python workspace for data science work.

• Jupyter notebooks organised with narrative, code, outputs and reproducibility notes
• Virtual environments isolate package versions for repeatable project execution
• NumPy arrays support vectorised numerical operations and matrix calculations
• pandas Series and DataFrames model labelled tabular business data
• File paths, CSV imports and Excel exports handled reproducibly
• Functions package repeated transformation logic into testable analytical steps

Participants can inspect data sources and establish whether their structure supports the intended analysis.

• Data dictionaries document fields, units, source systems and refresh frequency
• Schema inspection identifies data types, ranges, keys and duplicate records
• Primary keys and foreign keys tested before joining relational extracts
• Granularity mismatches identified between transaction, customer and monthly datasets
• Sampling strategies distinguish operational extracts from analytical populations
• Data lineage records transformations from source extract to model input

Participants can clean common data defects and create validated analysis-ready tables.

• Missing-value patterns assessed using counts, proportions and visual summaries
• Imputation methods selected from median, mode, forward fill and model-based approaches
• Duplicate records resolved using business keys and survivorship rules
• String standardisation corrects casing, whitespace, codes and category labels
• Date parsing creates valid timestamps, periods and elapsed-time measures
• Validation checks flag impossible values, outliers and inconsistent records

Participants can transform raw fields into features suitable for analysis and modelling.

• Merge, concat and groupby operations assemble analytical datasets from source tables
• Pivot tables calculate customer, product and period-level summary features
• Categorical variables encoded using one-hot, ordinal and target encoding
• Numerical scaling applies standardisation, min-max scaling and logarithmic transformation
• Lagged variables capture prior behaviour for forecasting and retention models
• Feature definitions documented with formulas, timing rules and source fields

Participants can identify patterns, anomalies and candidate relationships through structured exploratory analysis.

• Distribution shapes examined using histograms, density plots and box plots
• Summary statistics compare central tendency, spread, skewness and missingness
• Segment analysis contrasts outcomes by customer, channel, region and product
• Correlation matrices identify linear associations and multicollinearity warnings
• Scatter plots reveal nonlinear patterns, clusters and influential observations
• Outlier investigations separate data errors from rare but valid cases

Participants can create charts that communicate analytical evidence clearly and accurately.

• matplotlib figures control axes, labels, annotations and export settings
• seaborn charts compare distributions, categories, relationships and statistical estimates
• Chart selection matches comparison, composition, distribution and trend questions
• Colour scales distinguish categories without obscuring important values or uncertainty
• Faceting exposes differences across segments, periods and geographic areas
• Analytical narratives connect visual evidence to decisions and recommended actions

Participants can use probability concepts to reason about variation, samples and uncertain outcomes.

• Probability rules describe joint, conditional and independent events
• Bayes' theorem updates prior beliefs using observed evidence
• Random variables model discrete counts and continuous measurements
• Binomial, normal and Poisson distributions represent common business processes
• Sampling distributions explain variation in estimates from repeated samples
• Confidence intervals express plausible ranges around estimated population values

Participants can test analytical claims and interpret experimental or observational evidence responsibly.

• Null and alternative hypotheses express measurable claims about populations
• t-tests compare means between independent and paired samples
• Chi-square tests assess associations between categorical variables
• p-values, effect sizes and confidence intervals interpreted together
• A/B test design defines randomisation, metrics, duration and guardrail measures
• Multiple-testing risks controlled through pre-registration and adjusted thresholds

Participants can build and diagnose regression models for continuous-value prediction.

• Linear regression estimates continuous outcomes from numerical and categorical predictors
• Ordinary least squares coefficients interpreted with units and reference categories
• Residual plots diagnose nonlinearity, heteroscedasticity and influential observations
• Regularisation uses Ridge and Lasso penalties to control overfitting
• Mean absolute error and root mean squared error compare predictions
• Polynomial and interaction terms represent conditional and nonlinear relationships

Participants can develop classification models and assess their suitability for decision use.

• Logistic regression estimates probabilities for binary business outcomes
• Decision trees split records using interpretable feature thresholds
• Random forests aggregate trees to improve predictive stability
• Confusion matrices classify true positives, false positives and missed cases
• Precision, recall, F1 score and ROC-AUC assess classifier performance
• Threshold selection reflects intervention capacity and unequal error costs

Participants can compare candidate models using appropriate validation designs and performance measures.

• Train, validation and test splits preserve unbiased final performance estimates
• K-fold cross-validation estimates model performance across multiple data partitions
• Stratified sampling preserves target proportions within classification datasets
• Data leakage prevented by fitting transformations within training partitions only
• Grid search tunes hyperparameters against predefined validation metrics
• Baseline comparisons test whether complexity delivers material predictive improvement

Participants can address rare outcomes and investigate where predictive models fail.

• Class imbalance diagnosed through target prevalence and confusion matrix patterns
• Random oversampling and undersampling adjust training-set class representation
• Class weights penalise errors on underrepresented target outcomes
• Precision-recall curves assess rare-event models more meaningfully than accuracy
• Error segmentation compares performance across customer, channel and risk groups
• Misclassified records investigated for data defects and missing predictors

Participants can create and assess meaningful segments where no target outcome exists.

• K-means clustering groups records around iteratively calculated centroids
• Feature scaling prevents high-range variables dominating distance calculations
• Elbow plots and silhouette scores guide cluster-number selection
• Principal component analysis reduces correlated features into component scores
• Cluster profiles describe behavioural, value and service-use differences
• Segmentation actions distinguish useful groups from statistically convenient clusters

Participants can prepare time-indexed data and evaluate practical forecasting approaches.

• Time indexes, frequency and missing periods checked before forecasting
• Trend, seasonality and calendar effects identified through decomposition plots
• Lag features and rolling statistics create forecasting predictors
• Naive and seasonal-naive forecasts establish essential benchmark performance
• Time-based validation preserves chronological order between training and test periods
• Forecast error measured using MAE, RMSE and mean absolute percentage error

Participants can explain model behaviour, identify material risks and document responsible use conditions.

• Feature importance ranks variables contributing to tree-based model predictions
• Partial dependence plots show average predicted effects across feature values
• Residual and error analysis identifies groups receiving poorer predictions
• Protected characteristics and proxy variables assessed for potential unfair treatment
• Model cards document purpose, data, metrics, limitations and intended users
• Monitoring plans define drift indicators, review frequency and escalation owners

Participants can package notebooks, code and findings for review and reuse by colleagues.

• Notebook sections separate business context, methods, results and recommendations
• Reusable functions replace copied code and undocumented transformation sequences
• Version control commits record meaningful changes to scripts and notebooks
• Requirements files specify package versions for consistent project setup
• Code review checks readability, assumptions, outputs and failure handling
• Executive summaries state findings, confidence, limitations and decision implications

Participants can build, review and present a complete Python data science project addressing a defined business decision.

• Business problem framed with target, users, decisions and success criteria
• Source data cleaned, joined and validated using documented transformation rules
• Exploratory findings translated into justified feature engineering decisions
• Candidate models compared against baselines using appropriate validation methods
• Performance, limitations, fairness risks and monitoring needs documented clearly
• Final notebook and stakeholder presentation communicate recommendations and next actions