Time-Series ML

Read these first

1. 01Direct vs Recursive Multi-Step Forecasting

   Forecasting five days ahead can mean feeding a one-step model its own output five times in a row, or training five separate models, one per horizon, each with its own tradeoff between compounding error and inconsistency.

2. 02Global vs Local Forecasting Models

   When you have thousands of related time series to forecast, one per stock, one per store, you can fit a separate model to each one, or pool them all into a single shared model, and the right choice depends on how much data each individual series actually has.

3. 03Hierarchical Forecasting and Reconciliation

   Forecast a company's revenue and forecast each of its business segments separately, and the segment forecasts almost never add up to the company total, reconciliation is the discipline of adjusting them so a hierarchy of forecasts is internally consistent.

4. 04Reframing Forecasting as Supervised Learning

   Classical time-series models treat a series as one continuous object to model directly; machine learning instead slices it into a table of past-windows-predicting-future-values, turning forecasting into an ordinary supervised learning problem, with real tradeoffs from making that switch.

Then the rest

1. 05Anomaly Detection in Time Series
2. 06Choosing the Forecast Horizon
3. 07CRPS for Probabilistic Forecast Evaluation
4. 08Designing the Lookback Window
5. 09Dynamic Time Warping and Shape Similarity
6. 10Exogenous and Known-Future Covariates
7. 11Foundation Models for Time-Series Forecasting
8. 12Lag Features and Leakage Boundaries
9. 13Multi-Horizon Joint Forecasting
10. 14Per-Series Target Scaling
11. 15Preprocessing Leakage in Time-Series Pipelines
12. 16Recency-Weighted Training Samples
13. 17Sequence-to-Sequence Forecasting Architectures