Anomaly Detection in Time Series
Outlier detection asks whether a value is unusual. Time series anomaly detection asks whether it is unusual now, which is a different and harder question.
Recent writing
Probabilistic Forecasting: Beyond the Point Estimate
A point forecast answers the wrong question. Most decisions depend on how bad things could plausibly get, which is a statement about the whole distribution.
Missing Data and Irregular Sampling in Time Series
A missing row in a table is a nuisance. A missing interval in a time series changes the meaning of every lag, window and seasonal index computed from it.
Conformal Prediction for Operational Risk Decisions
Conformal prediction helps teams express model uncertainty as calibrated intervals or prediction sets that can be used in operational risk decisions.
Hierarchical Forecasting: Making Forecasts Add Up
Forecast every store separately and the total will not match the forecast you made for the company. Reconciliation is how you make a hierarchy of forecasts coherent.
Weak Supervision for Better Machine Learning Labels
Weak supervision helps teams scale labeling by combining imperfect rules, heuristics, and external signals instead of hand-labeling every example.
Feature Engineering for Time Series Without Leaking the Future
Turning a time series into a tabular problem unlocks powerful models and introduces a specific failure: features that quietly contain information from the future.
Multiple Seasonality: MSTL, TBATS, and Fourier Terms
Hourly and daily data rarely has one season. Electricity demand cycles daily, weekly and annually at the same time, and a single seasonal period cannot represent that.
Multilevel Models for Operational Analytics
Multilevel models help analysts estimate group-level performance without overreacting to small samples or ignoring real differences between sites.
Forecasting Baselines That Are Hard to Beat
An RMSE of 4.2 means nothing on its own. Without a baseline you cannot tell whether a model is skilful or merely arithmetic.