Foresight AI is an AutoML forecasting platform for multi-series time series prediction across industries. It serves retail, energy, finance, and manufacturing sectors. The platform handles hierarchical data structures, grouping levels, and exogenous variables.
This article details the construction of a machine learning pipeline on Vertex AI. It addresses the decision between Google Cloud AutoML Forecasting and custom Vertex AI models. The text provides practical guidance for defining a forecasting strategy for rapid implementation or long-term scale.
Vertex AI Forecasting: AutoML or Custom Models?
ML teams using Vertex AI for time series forecasting face a core decision: AutoML or manual modeling. AutoML suits quick, low-effort forecasts with minimal infrastructure. Custom models suit complex, hierarchical, large-scale forecasting. A hybrid pipeline combines the strengths of both.
Google Cloud AutoML Forecasting offers a low-effort solution. Users upload data, set parameters, and deploy models. However, Foresight AI requires domain-agnostic intelligence and hierarchical time series forecasting. This need complicates the decision.
The system must:
- Scale to thousands of time series
- Adapt to retail, finance, and energy sectors
- Ingest external signals, such as weather, calendar events, and macroeconomic indicators
- Maintain flexibility and interpretability at scale
We evaluated AutoML and custom models, then designed a hybrid pipeline that combines the strengths of both.
Google Cloud AutoML Forecasting Strengths
Google Cloud AutoML performs well in:
- Zero infrastructure overhead: No operations required. Deploy with a single action.
- Automated machine learning for forecasting: Provides basic lags, trends, and validation by default.
- Rapid iteration: Suited for retail prototyping and standard business forecasting strategies.
- Standard backtesting: Includes built-in metrics such as RMSE and MAPE.
AutoML supports early prototyping and benchmarking. However, customization, scalability, and domain control requirements exceed its capabilities.
Vertex AI AutoML Forecasting Limitations
Constraints in Complex Forecasting
AutoML simplicity provides both primary benefits and primary constraints.
- AutoML lacks robust hierarchical data handling. It struggles with multi-level identifiers, such as product, region, store, category, and time.
- Rigid feature engineering blocks the integration of custom seasonality, business calendars, and domain-specific lags.
- AutoML limits model selection. You cannot apply domain knowledge, such as ARIMA for short-term retail sales or Prophet for long-horizon planning.
- AutoML lacks robust support for exogenous variables. It cannot easily incorporate external signals like weather, economic indicators, or promotions. These signals are essential for energy and finance forecasting.
Building a Custom Forecasting Pipeline on Vertex AI
We built a custom ML pipeline to address these gaps. The goal combines automation with full control using Vertex AI Pipelines, Nixtla libraries, and modular components.
Architecture Overview
The pipeline consists of the following components:
- Data Validation Framework
- Automated EDA Engine
- Time-Series Feature Engineering
- Multi-Model Training System
- Vertex AI Implementation Architecture
This section defines each component.
Data Validation: Ensure Input Accuracy
Standardize and validate all time series prior to modeling:
- Date Parsing: Enforce ISO standards for consistent input
- Missing data strategy: Apply forward fill, then backward fill, then drop rows.
- Duplicate handling: Aggregate data using logic defined for the use case.
- Outlier detection: Use the IQR method with manual overrides.
A robust validation layer is essential. It ensures reliability and reproducibility across all domains.
Automated EDA Drives Forecasting Strategy Selection
The EDA engine extracts key patterns from your data:
- Detect seasonality with STL and ACF
- Analyze trends using regression and the Mann-Kendall test
- Test stationarity with ADF and KPSS
- Identify cross-series correlations
Model Selection Criteria
- Strong seasonality. Use AutoETS, Prophet, or Seasonal Naive.
- Multiple exogenous variables. Use machine learning. Select XGBoost or Random Forest.
- High correlation across series. Use global models. Add series_id as a feature.
Time Series Feature Engineering
See "Bridging Two Worlds of AI: A New Approach to Time Series Forecasting" for integration methods.
Custom feature engineering achieved cross-vertical accuracy.
Lag Features
- Daily: lags 1, 7, 14, 30
- Weekly: lags 1, 4, 12, 52
- Monthly: 1, 3, 6, 12
Rolling Features:
- Mean, std, min, max over 3, 7, 30 periods
- Capture trend, volatility, and uncertainty
Seasonal Features:
- Time: Sine/cosine encoding
- Calendar events: Holidays, quarter ends, business hours
Exogenous variables:
- Numeric features: StandardScaler
- Categorical features: Encoding strategies
- Lead-lag correlation analysis to align predictor timing
Multi-Model Training: Balance AutoML Speed with Custom Accuracy
We selected Nixtla libraries over Darts to improve performance in multi-series forecasting.
Model Portfolio (5–8 models per run):
Statistical Models: AutoARIMA, AutoETS, Prophet, Seasonal Naive
Models: XGBoost, LightGBM, Random Forest, Linear Regression
Validation Method:
- 3-fold temporal cross-validation
- Primary Metric: MAPE
Ensemble Strategy:
- Equal-weight blend of the top three models
- Improved forecast accuracy by up to 15% over naive baselines
Vertex AI Implementation: Scaling Custom Models
This process used Vertex AI forecasting tools for both AutoML and custom model runs.
Custom Training:
- Model selection via exploratory data analysis
- Precise feature engineering control
- Total flexibility to optimize forecast accuracy
Vertex Pipelines
- Coordinate training, retraining, and batch predictions
- Register models with version control
- Deploy models using A/B testing and performance monitoring
Cost and Performance: AutoML vs Custom Models
| Criteria | Google AutoML Forecasting | Vertex AI Custom Models |
|---|---|---|
| Setup Time | ✅ Rapid | ⏳ Longer |
| Forecast Accuracy | ❌ Limited in complex cases | ✅ Higher with feature tuning |
| Infrastructure Overhead | ✅ None | ⚠️ Requires setup |
| Scalability | ⚠️ Limited for 1000+ series | ✅ Built for scale |
| Flexibility | ❌ Black-box | ✅ Fully customizable |
Conclusion: Hybrid Forecasting Models
Effective forecasting requires combining automation with domain-specific intelligence. Select the appropriate approach for each use case.
Consult our data engineering experts for tailored recommendations.
Contact Us
Q: When should you use Vertex AI AutoML Forecasting instead of custom models?
- Use AutoML for quick, low-effort forecasts with minimal infrastructure.
- It suits standard time series with limited hierarchy and few external variables.
- It works best when historical patterns are clear.
- Teams use it for prototyping, benchmarking, or simple retail demand forecasts.
- Switch to custom models as complexity increases.
Q: Why do custom models on Vertex AI improve forecasting accuracy for complex use cases?
- Custom models allow tailored feature engineering, model selection, and validation.
- They capture data behaviors found in exploratory analysis.
- Combining statistical, machine learning, and ensemble methods improves accuracy.
- This approach handles seasonality, volatility, and cross-series patterns better.
- It provides higher accuracy and interpretable forecasts in complex enterprise environments.
Q: What does a hybrid forecasting approach on Vertex AI look like in practice?
- Use AutoML for fast experimentation and baseline benchmarks.
- Use custom models for production forecasting with advanced features and scalability.
- Vertex AI Pipelines orchestrate both workflows.
- This enables automated retraining, version control, and performance monitoring.
- The hybrid model offers AutoML speed and custom ML precision without single-strategy lock-in.
Q: How does SquareShift help organizations choose between AutoML and custom forecasting models on Vertex AI?
- SquareShift performs a forecasting readiness and data maturity assessment.
- It reviews business objectives, data complexity, and scale requirements.
- We recommend AutoML, custom models, or a hybrid approach for the best ROI.
- Our guidance relies on production experience with Vertex AI, not theoretical comparisons.
Q: Can SquareShift build and operate enterprise-scale custom forecasting pipelines on Vertex AI?
- Yes. SquareShift builds end-to-end forecasting pipelines on Vertex AI.
- This includes data validation, feature engineering, multi-model training, and deployment via Vertex AI Pipelines.
- Solutions are scalable, modular, and MLOps-ready.
- We support automated retraining, monitoring, and version control for thousands of time series.
Q: How does SquareShift ensure long-term value and accuracy from forecasting solutions?
- SquareShift focuses on continuous improvement, not one-time delivery.
- We implement robust performance monitoring and periodic retraining.
- Ensemble-based optimization maintains accuracy as data patterns change.
- We combine domain expertise, custom modeling, and Vertex AI best practices.
- This sustains forecasting performance at enterprise scale.
Q: What is the difference between AutoML and custom models for Vertex AI forecasting?
AutoML is rapid to set up and needs no infrastructure, but accuracy is limited in complex cases and it is a black box. Custom models take longer to set up and need infrastructure, but they give higher accuracy with feature tuning, are built for scale and are fully customizable.
Q: Can Vertex AI AutoML handle hierarchical time series?
AutoML lacks robust hierarchical data handling. It struggles with multi-level identifiers, such as product, region, store, category, and time. It is also limited for 1000+ series. The article builds a custom pipeline on Vertex AI Pipelines to address these gaps, with Nixtla libraries and modular components.
Q: Which models does the custom Vertex AI forecasting pipeline use?
Each run trains a portfolio of five to eight models. Statistical models include AutoARIMA, AutoETS, Prophet and Seasonal Naive. Machine learning models include XGBoost, LightGBM, Random Forest and Linear Regression. Validation uses 3-fold temporal cross-validation with MAPE, and an equal-weight blend of the top three models forms the ensemble.