Skip to content

6. Forecasting Pipeline

Workflow image

Machine Learning Forecasting Pipeline Diagram

The forecasting pipeline is built as a modular, scalable architecture to maximize reusability, flexibility, and robust deployment in production settings. Each step below corresponds to a key building block in the end-to-end ML process:

1. Data Ingestion

  • Objective: Aggregate data from multiple sources and standardize into a unified schema (typically columns like ds, y, plus regressors).
  • Why: Ensures compatibility across models and tools; enables seamless integration with upstream data sources.

2. Preprocessing

  • Objective: Clean and prepare raw data.
  • Key tasks: Date format normalization, handling missing values/outliers, adjusting data granularity (e.g., by store, product group).
  • Impact: Reduces noise, improves model robustness and data quality.

3. Feature Engineering

  • Automated (TSFresh): Extracts Fourier coefficients, entropy, statistical and temporal features.
  • Manual: Incorporates business-driven or domain-specific variables.
  • Hybrid (Prophet): Leverages decomposed trend, seasonality, special event features.
  • LLM-RAG: Uses generative AI for feature suggestion, automated reports, and advanced diagnostics.

4. Feature Selection

  • Objective: Identify and retain the most predictive features.
  • Tools: AutoML routines, SHAP for feature importance, LLM-RAG for diagnostic summaries and data quality checks.

5. Model Training

  • Objective: Fit the best forecasting model(s) to each data group or cluster.

  • Strategy: Clustering (e.g., KMeans) assigns time series to groups.

  • Prophet: For series with strong trend/seasonality.

  • XGBoost: For nonlinear, complex signals.
  • LightGBM: For noisy, large-scale or highly variable data.

6. Hyperparameter Tuning

  • Objective: Find the optimal model hyperparameters.
  • Method: Leverage Optuna or Grid Search per group/cluster, supports both global and per-model tuning.

7. Cross-Validation

  • Objective: Robustly estimate generalization error and prevent overfitting.
  • How: Time series cross-validation splits (respects data order), multiple train/test splits, supports metric-based model selection.

8. Model Evaluation

  • Objective: Quantify and compare model performance.
  • Metrics: MAE, RMSE, MAPE, SMAPE, Coverage, etc.
  • Tools: MLflow logs all metrics, model artifacts, and visualizations.

9. Forecast Reconciliation

  • Objective: Ensure coherence and consistency in multi-level (hierarchical) forecasts.
  • How: Use HierarchicalForecast with strategies like Bottom-up, Top-down, OLS, MinTrace to enforce aggregation constraints (e.g., system > region > store > SKU).

10. Model Diagnosis

  • Objective: Monitor training behavior and diagnose issues.
  • Tasks: Plot learning curves, inspect train/val loss, learning rate schedules, spot overfitting/underfitting, flag anomalies.

11. Deployment

  • Objective: Serve models for batch or real-time forecasting.
  • Tools: FastAPI provides REST API endpoints; supports both scheduled jobs and on-demand requests.

12. Prediction Output

  • Deliverables: Forecasts (yhat, yhat_lower, yhat_upper), actuals (y), and time index (ds).
  • Usage: Feeds into planning, downstream systems, and performance evaluation.

13. Visualization & Reporting

  • Objective: Enable business and technical users to interact with results.
  • Tools: Streamlit dashboard for trend analysis, forecast vs. actual comparison, error breakdowns, and auto-generated reports.

This modular pipeline enables the system to scale, adapt to changing business requirements, and ensure each step is traceable, testable, and easily maintained.

For a visual summary, see the Workflow Diagram above. For deeper technical detail on each module, refer to the Architecture and Feature Engineering sections.