Charting Multiple Routes

Opening Scene

A fleet of ships doesn’t navigate in complete isolation from each other. Weather reports from one vessel inform the routing of another; a current measured by one ship’s instruments applies to every ship crossing the same water. Treating each ship’s course as a totally independent problem throws away real, useful information the fleet collectively has access to. Many forecasting problems have this exact same structure.

In Plain English

Multivariate time-series forecasting models multiple related sequences together, rather than forecasting each one in isolation, explicitly capturing how they influence each other. A retailer forecasting demand for related products, an economist forecasting interest rates alongside inflation, or a utility forecasting demand alongside temperature are all multivariate problems — the sequences carry genuine, exploitable information about each other that univariate methods, covered in Articles 8 and 9, simply can’t access.

The Old Way

Before formal multivariate methods existed, people captured cross-series relationships more informally:

An economist noting that interest rates and inflation tend to move together, without a formal joint statistical model capturing exactly how.
A retailer noticing that ice cream and sunscreen sales rise and fall together, based on informal observation rather than a rigorous joint model.
A utility operator manually adjusting demand forecasts based on the weather report, without formally incorporating temperature as a model input.

The intuition that related sequences carry information about each other predates any formal multivariate technique.

What’s Changing (and Why AI Is the Reason)

Formal multivariate methods — vector autoregression (VAR), and multivariate extensions of machine learning approaches covered in Article 11 — now capture cross-series relationships explicitly and rigorously, rather than relying on informal, ad hoc adjustment.
As more related data streams have become available, connecting directly to the broader data ecosystem covered throughout this content library’s data engineering series, multivariate forecasting has become practical at a scale that simply wasn’t feasible before.
Modern deep learning forecasting methods, covered in Article 12, can incorporate related sequences and external variables naturally within the same model, further lowering the barrier to genuinely multivariate forecasting.

The Metaphor, Fully Extended

The Voyage	Multivariate Forecasting Concept
A fleet sharing weather and current information across ships	Related time series carrying information about each other
Charting one ship’s course while ignoring the rest of the fleet	Forecasting one sequence in isolation, ignoring related series
A shared weather report improving every ship’s routing	An external variable improving forecasts across related sequences
A coordinated fleet-wide navigation system	A formal multivariate model capturing cross-series relationships

For Beginners: What to Actually Do

Identify at least one related external variable for a forecasting problem you’re working on, and test whether including it actually improves forecast accuracy.
Learn the basics of vector autoregression as the classical entry point into formal multivariate forecasting.
Practice distinguishing genuine, exploitable cross-series relationships from coincidental correlation.

For Practitioners and Leaders: The Deeper Layer

Inventory related data streams that could plausibly improve a forecasting problem before committing to a purely univariate approach.
Weigh the real added complexity of multivariate modeling against the genuine accuracy gain it provides for your specific problem — it isn’t automatically worth it everywhere.
Connect forecasting teams with the broader data engineering and data platform work covered elsewhere in this content library, since multivariate forecasting depends directly on that data being genuinely available and well-integrated.

Quick Recap

Multivariate forecasting models multiple related sequences together, capturing information univariate methods can’t access.
The intuition behind it is old, but formal methods like vector autoregression made it rigorous and precise.
Growing data availability has made genuinely multivariate forecasting increasingly practical.
The added complexity of multivariate modeling should be weighed against its actual, measured accuracy benefit.

Where This Fits in the Series

Article 10 covered forecasting sequences together rather than in isolation. Article 11 turns to the modern machine learning fleet built to handle both univariate and multivariate forecasting at scale.