AI & Analytics

Vertex AI Forecasting: A Step-by-Step Guide

A step-by-step walkthrough for building production-ready forecasting models using Vertex AI’s forecasting tools and BigQuery ML.

Why Use Vertex AI for Forecasting

Vertex AI provides managed tools specifically designed for time-series forecasting, including AutoML Forecasting and custom training options built on top of Google’s internal forecasting research. For enterprises forecasting demand, revenue, or resource utilization, Vertex AI removes much of the infrastructure burden of building, tuning, and deploying forecasting models compared to maintaining custom pipelines from scratch.

Step 1: Prepare and Structure Your Data

Forecasting models require historical data structured with a clear time column, a target value to predict, and a series identifier if you are forecasting multiple entities such as products or regions. Data should typically live in BigQuery, since Vertex AI integrates natively with BigQuery for both training data ingestion and batch prediction output. At this stage, teams should also decide on the forecast horizon and granularity, such as daily sales forecasts thirty days out versus monthly demand forecasts twelve months out, since this shapes feature engineering decisions.

Step 2: Engineer Features and Add Covariates

Accuracy improves significantly when models have access to covariates beyond the historical target values, such as promotional calendars, pricing changes, holidays, or weather data. Vertex AI Forecasting supports both known future covariates, such as a planned promotion, and unknown future covariates that are only available historically. Spending time identifying which external signals correlate with the target variable is often the highest-leverage step in the entire process.

Step 3: Configure and Train the Model

Within Vertex AI, teams can choose AutoML Forecasting for a managed, lower-effort approach that automatically searches across model architectures, or they can use Vertex AI’s custom training jobs with open-source frameworks such as Prophet, temporal fusion transformers, or custom TensorFlow models for more control. AutoML Forecasting is generally the right starting point, since it provides strong baseline accuracy without requiring deep forecasting expertise, and teams can benchmark custom approaches against it later.

Step 4: Evaluate Model Performance

Vertex AI provides standard forecasting evaluation metrics including mean absolute percentage error, root mean squared error, and quantile loss for probabilistic forecasts. Rather than evaluating only on an aggregate metric, teams should review performance by segment, since a model with strong overall accuracy may still perform poorly for specific product categories or regions that matter most to the business. Backtesting across multiple historical windows, not just a single holdout period, gives a more reliable picture of real-world performance.

Step 5: Deploy and Operationalize

Once a model meets accuracy thresholds, it can be deployed for batch prediction, which is typical for periodic demand planning cycles, or for online prediction if forecasts need to be generated on demand. Operationalizing the model also means building monitoring for data drift and forecast accuracy decay over time, since demand patterns shift and a model trained on pre-pandemic or pre-disruption data can degrade silently without active monitoring.

Step 6: Integrate Forecasts Into Planning Processes

The forecast itself only creates value once it is consumed by planning processes, whether that means feeding into an ERP demand planning module, a supply chain planning tool, or a finance forecasting workbook. Building a clear feedback loop, where planners can flag forecast errors and that feedback informs model retraining, closes the gap between a technically accurate model and a forecast that the business actually trusts and uses.

Symhas helps enterprises design and deploy forecasting pipelines on Vertex AI and BigQuery, from data architecture through planning integration. Contact us to scope your forecasting use case.

Schedule a Briefing →

Frequently Asked Questions

Do I need machine learning expertise to use Vertex AI Forecasting?

AutoML Forecasting is designed to be accessible without deep ML expertise, though data preparation and business context knowledge are still essential for good results.

What data is required to start forecasting with Vertex AI?

You need historical time-series data with a clear timestamp, target value, and optionally covariates such as pricing or promotions, typically stored in BigQuery.

How often should forecasting models be retrained?

Most organizations retrain monthly or quarterly, or trigger retraining automatically when monitoring detects accuracy degradation or data drift.