I'm digging into the idea of a 'local series' as a way to model subsets of a larger time‑series dataset. Unlike global models that try to capture the whole pattern, local series focus on specific windows or regimes, often adapting more quickly to changing dynamics. I'm curious how practitioners define the boundaries of these local segments, what criteria are common for switching between them, and which statistical or machine learning techniques work best in practice. Any papers, tutorials, or personal experiences you can share would help a lot. How do you approach this in your projects? 😊
Exploring the Mechanics Behind Local Series in Time‑Series Forecasting
👁️ 73 görüntüleme💬 1 cevap❤️ 0 beğeni
1 Cevap
In my recent projects I’ve treated a “local series” as a rolling window that’s re‑trained whenever a statistical drift test flags a change. A cheap but surprisingly reliable way to set the boundaries is to run a sequential Chow test or a rolling‑window ADF test; when the p‑value crosses a preset threshold you slice the series and start a fresh model on the next segment. In practice I combine that with a simple change‑point detector like Ruptures (PELT or binary segmentation) – it gives you concrete breakpoints without over‑fitting, and you can tune the penalty term to control how sensitive you want the switches to be.
For the actual modeling I prefer lightweight models that can be fit quickly on each segment: a seasonal ARIMA for short, well‑behaved windows, and a LightGBM or Temporal Fusion Transformer when the regime is more volatile. The key is to keep the retraining pipeline automated: detect a change, re‑estimate the window length (often 30–90 days for daily data), and spin up the next model. A handy reference is the paper “Detecting Regime Shifts in Time Series with Bayesian Online Change‑Point Detection” (Adams & MacKay, 2007) – it walks through the Bayesian formulation that integrates nicely with probabilistic forecasts. If you need a quick start, the “tslearn” library’s sliding‑window segmentation tutorial is a good practical example.