Multi-Step Forecasting

In real business problems, predicting just the next time step is rarely enough. Most decisions require knowing what will happen over a range of future points.

This is where multi-step forecasting comes in.

---

What Is Multi-Step Forecasting?

Multi-step forecasting means predicting multiple future values at once.

Examples:

• Forecasting sales for the next 30 days
• Predicting energy demand for the next week
• Estimating stock levels for the next quarter

Instead of answering:

“What happens next?”

We answer:

“What happens over the next N steps?”

---

Why Single-Step Forecasting Is Not Enough

Single-step forecasts work well for short-term control problems. But they fail when decisions depend on longer horizons.

Consider inventory planning. Ordering too much or too little depends on expected demand over many days — not just tomorrow.

---

Main Strategies for Multi-Step Forecasting

There are three commonly used approaches:

• Recursive forecasting
• Direct forecasting
• Multi-output forecasting

---

1. Recursive Forecasting

Recursive forecasting predicts one step ahead, then feeds that prediction back into the model to predict the next step.

This repeats until all future points are generated.

Real-World Interpretation

Think of walking in fog. Each step is based on your previous position — but small mistakes accumulate over time.

Visual Example

What to notice:

• Forecast starts accurately
• Error grows as steps increase
• Uncertainty compounds

---

2. Direct Forecasting

Direct forecasting builds a separate model for each future step.

One model predicts t+1, another predicts t+2, and so on.

Real-World Interpretation

Instead of guessing step by step, you plan each future point independently.

Strengths and Weaknesses

• Less error accumulation
• More stable long-term forecasts
• Higher computational cost

---

3. Multi-Output Forecasting

Multi-output forecasting trains a single model to predict all future steps at once.

This is common in deep learning models.

Real-World Interpretation

It’s like planning the entire route before starting the journey.

Where It’s Used

• LSTM and GRU networks
• Transformer models
• Sequence-to-sequence forecasting

---

Choosing the Right Approach

There is no universally best method. The choice depends on the problem.

Scenario	Recommended Approach
Short horizons	Recursive
Medium horizons	Direct
Long horizons	Multi-output

---

Business Impact of Multi-Step Errors

Errors grow with horizon length.

That’s why companies:

• Re-forecast frequently
• Use confidence intervals
• Combine models

Good forecasting is not about perfect accuracy — it’s about managing uncertainty.

---

Practice Questions

---

Next lesson: Probabilistic forecasting — predicting uncertainty, not just point values.