Regime-Switching Models for Time Series
A single model fitted across a recession and an expansion describes neither. Regime-switching models allow the dynamics themselves to change, with the regime inferred rather than assumed.
Recent writing
Nowcasting with Mixed-Frequency Data
The quantity you care about arrives quarterly and two months late. Related indicators arrive daily. Nowcasting is the problem of estimating the present from what has already been published.
Interrupted Time Series and Causal Impact
A intervention happened at a known date and you need its effect. There is no control group, only the series itself before and after, and the counterfactual has to be constructed.
Forecast Value Added: Is Your Process Helping?
Forecasting processes accumulate steps: a statistical model, a planner override, a consensus meeting. Each is assumed to improve the number. FVA is how you find out.
Temporal Hierarchies: Reconciling Across Time Granularities
A hierarchy does not have to be geographic. Aggregating a series over time produces the same coherence problem, and the same machinery solves it.
Label Noise in Supervised Learning: When the Target Cannot Be Trusted
Label noise is one of the most damaging data quality problems in supervised learning because it corrupts the target the model is trained to imitate.
Neural Forecasting: What the Architectures Actually Do
Neural forecasting has produced genuinely useful architectures and a great deal of noise. The differences between them are more interesting than their benchmark scores.
Modelling Count Time Series
Daily incident counts are integers, non-negative, often small, and correlated with yesterday. ARIMA assumes none of that and Poisson regression assumes independence.
Long Memory and Fractional Integration in Time Series
Standard practice offers two options: the series is stationary, or you difference it. Some series are genuinely in between, and forcing them either way loses information.
Dynamic Time Warping and Time Series Clustering
Two series can trace an identical shape while one runs slightly ahead of the other. Point-by-point distance calls them dissimilar; dynamic time warping does not.