Insights
Long-form articles on the craft of modern data work — machine learning models, data visualisation and statistical analysis. Practical, opinionated, and free to read.
From linear baselines to gradient boosting and deep neural networks — how to pick, train and trust the model that actually fits your problem.
Charts are arguments, not decoration. A short guide to picking the right visual, removing noise and making numbers persuasive without distorting them.
Hypothesis tests, effect sizes, confidence intervals and the assumptions everyone forgets. A practical refresher for analysts who want defensible conclusions.
Algorithms get most of the credit, but features decide the ceiling. A field guide to encoding, scaling, interactions, leakage and the features AutoML still misses.
Most dashboards die from neglect within three months. A guide to building dashboards that drive decisions — hierarchy, defaults, alerts and the one chart you really need.
Priors, posteriors, credible intervals and why probabilistic answers communicate better than p-values in most business conversations.
SHAP, LIME, permutation importance and partial dependence — a practical guide to extracting trustworthy explanations from complex models.
Stacking, blending, boosting and bagging — how to combine models so the ensemble outperforms every member, and when it is worth the complexity.
The best visualisations do not just show data — they tell a story. A framework for structuring analytical narratives that persuade and endure.
Animation can guide attention, reveal change over time, or waste it. Principles for using motion responsibly in charts, dashboards and presentations.
When randomised experiments are impossible, how do you still estimate causal effects? A tour of difference-in-differences, regression discontinuity and instrumental variables.
When the event has not happened yet for everyone, ordinary regression fails. A guide to Kaplan-Meier curves, Cox models and handling censoring with confidence.
A tour of the eleven dashboard layouts inside ML Studio — what each one is optimised for, where it fails, and how to pick the grid that matches your audience.
Version control, reproducible training, model registries and monitoring — the minimum viable MLOps stack for teams of one to five.
Colour-blind palettes, screen-reader tables, keyboard navigation and the small design choices that widen the audience for your analysis.
Autocorrelation, stationarity, ARIMA vs Prophet vs state-space models — how to forecast without insulting the statistics.
Golden datasets, LLM-as-judge, red-teaming and cost-quality trade-offs — a disciplined evaluation loop for language-model features.
Choropleths, dot-density, hex bins and cartograms — how to choose a map that reveals rather than distorts spatial patterns.
Effect sizes, minimum detectable effects and the calculations that separate a well-designed experiment from an expensive false-negative.
Turn point predictions into calibrated intervals, understand what coverage guarantees actually mean, and recognise when changing data invalidates them.
Evaluate fraud, failure and other rare outcomes using precision, recall, realistic validation and thresholds that reflect the cost of mistakes.
Use interval plots, forecast bands and thoughtful labels to communicate what the data supports—and what remains uncertain.
Keep analytical dashboards responsive using meaningful aggregation, honest sampling and chart choices that preserve the important patterns.
Understand MCAR, MAR and MNAR, choose defensible missing-data treatments, and check how sensitive your conclusions are to the assumptions.
Distinguish exploratory screening from confirmatory testing, choose family-wise or false-discovery control, and report findings without cherry-picking.
Spot the leakage, duplication and scaling mistakes that make validation scores look better than reality — and build a split you can actually trust.
A disciplined search beats a thousand random fits: choose the parameters that matter, tune inside cross-validation, and stop before you overfit the validation set.
Use colour as information rather than decoration — ordered scales for ordered data, diverging scales for deviations, and accessible palettes that survive printing.
Read intervals as ranges of plausible effects, understand what the 95% actually promises, and report uncertainty in a way decision-makers can use.
Detect when a deployed model stops matching the world it serves, distinguish data drift from concept drift, and build a monitoring routine that catches both.
Every model is capped by the quality of its training labels. Measure agreement between annotators, write the guideline before the first batch, and treat disagreements as data rather than noise.
When a single chart with twenty lines becomes unreadable, split it into a grid of identical mini-charts. Shared axes, one highlight at a time, and a fixed order turn clutter into comparison.
Checking a running A/B test and stopping the moment it looks significant quietly multiplies false positives. Decide the sample size first, or use a stopping rule built for repeated looks.