Multi-step Forecasting

Predicting tomorrow is one thing; predicting the next two weeks is a different kind of challenge. Multi-step forecasting is about producing a sequence of future predictions from a time series, not just the next point.

What it means

In multi-step forecasting, you predict values at horizons 1, 2, …, H steps ahead (for example, sales for each of the next 14 days). Formally, given observations up to time t, you want estimates of y_t+1, y_t+2, …, y_t+H. The key difficulty is that uncertainty grows with the horizon: small errors at early steps can snowball into larger errors later.

Common strategies

• Recursive (iterated): train a one-step model and feed its own predictions back in to reach farther horizons. Simple, but can accumulate error.
• Direct: train a separate model for each horizon (one model for t+1, another for t+7, etc.). Less error propagation, more models to manage.
• Multi-output (joint): one model predicts the whole future vector at once. Can learn relationships across horizons.
• Hybrid: combinations like “direct-recursive” to balance bias, variance, and complexity.

Practical examples

• Forecasting hourly electricity demand for the next 48 hours to schedule generation.
• Predicting hospital admissions for each day next week to plan staffing.
• Projecting product sales for the next 12 months for inventory decisions.

Why it matters in ML

Many AI systems need decisions over a horizon (pricing, staffing, supply chain), so a single-step forecast isn’t enough. Multi-step setups also affect how you evaluate models: you typically report errors by horizon (e.g., MAE at 1-step vs 14-step). You’ll see multi-step forecasting in ARIMA/ETS baselines, tree-based regressors with lag features, and deep models like RNNs, Temporal Convolutional Networks, and Transformers (often via sequence-to-sequence training). Libraries such as statsmodels, scikit-learn (with feature engineering), and PyTorch Forecasting support these patterns.