Plain-English introductions to the methods behind our models, no jargon required. Here are the core techniques we use in analytics and machine learning work.
Time-series forecasting predicts future values from historical patterns, capturing trend, seasonality and cycles, and crucially, how uncertain the forecast is. A good forecast comes with a range, not just a line.
Where we use it: demand, load, arrivals and price forecasts that feed planning and staffing.
Regression models how an outcome depends on one or more input variables. It does double duty: predicting the outcome, and quantifying which drivers matter and by how much, so you understand the "why", not just the "what".
Where we use it: driver analysis, pricing, and as a transparent baseline for prediction.
Bayesian methods combine prior knowledge with observed data to produce probabilities and honest uncertainty ranges. They're especially valuable when data is scarce, expensive, or arriving over time.
Where we use it: risk estimation and decisions where quantifying uncertainty is the whole point.
Gradient boosting builds many small models that each correct the errors of the last, combining into a strong predictor. On everyday tabular data it's one of the most accurate methods available.
Where we use it: prediction from structured operational and business data.
Neural networks are flexible models that learn complex patterns from large amounts of data, the engine behind deep learning for images, text and sequences. Powerful, but best reserved for problems where their flexibility earns its keep.
Where we use it: rich, high-volume data such as sensor streams, images and language.
Clustering groups similar records together, customers, assets, days, sites, without being told the answer in advance. It reveals the natural structure and segments hiding in your data.
Where we use it: segmentation, profiling and simplifying complex populations.
Anomaly detection learns what "normal" looks like and flags the unusual, faults, failures, fraud, often before they become expensive. It's the watchful eye on data too large to monitor by hand.
Where we use it: condition monitoring, quality and fraud or error detection.
Survival analysis models the time until an event, a failure, a churn, a breakdown, and handles the tricky reality that for many cases the event simply hasn't happened yet. That makes it the right tool for lifetime and reliability questions.
Where we use it: reliability, maintenance and retention modelling.