Enterprise-Grade Time Series Forecasting with Extended Neural Models

With more than 120 locations in three countries and a product catalog of more than 40,000 items, the company must manage complex seasonal cycles, supplier-driven reordering and regional pricing. While the company’s BI and ERP systems provided basic dashboards for sales, strategic forecasting, particularly long-term forecasting at the category level, remained fragmented and prone to error.

The in-house analytics team experimented with basic statistical models and Python toolkits, but struggled with inconsistent accuracy, sparse data on low-volume items and an inability to capture event-driven fluctuations (e.g. promotions, weather effects, calendar shifts). In late 2023, the company engaged Devox Software under NDA to develop and deliver a production-ready, explainable ML engine for annual sales forecasting, capable of learning from historical behavior, integrating external signals and dynamically keeping pace with new sales patterns across the retail network.

• Machine Learning Architect (PyTorch, neural forecast, MLflow)
• 2 ML Engineers (Python, scikit-learn, PyTorch Forecasting, SHAP)
• Python Software Engineer (FastAPI, Pandas, PostgreSQL, ERP connectors)
• DevOps Specialist (Docker, AWS EC2, Airflow, GitHub Actions)
• Project Manager (Agile, Jira, stakeholder comms)

At project kickoff, the forecasting process combined legacy Excel workflows with partially scripted Python models. The client had several years of clean sales data and access to external signals, but forecast performance remained unstable, especially at the category and annual levels. During discovery, we mapped the key barriers:

• One-size-fits-all modeling: diverse time series patterns, including steady SKUs, erratic seasonal bursts, and promotion-driven items, were processed using identical statistical templates. This limited forecast adaptability across the portfolio.
• Fragmented signal architecture: marketing calendars, discounts, weather data and channel-specific prices were stored in isolated systems. These factors remained unused in the forecasting layer.
• No automation of retraining or version control: Forecast updates had to be made manually, without version control, backtesting or performance logging. This led to blind spots in planning and fluctuations between departments.
• Opaque forecasts: Department heads were given numerical forecasts without context, sources or confidence intervals. Decisions based on forecasts lacked support from explainable characteristics or a breakdown of factors.
• Inconsistent product hierarchy mapping: The training data frequently contained inconsistencies in store and SKU-level metadata, such as missing assortment identifiers or mismatched regional tags, which weakened signal quality before the models were even executed.

The combination of static models, unused signals, and fragmented ownership created operational gaps across forecasting, inventory, and replenishment teams.

• Machine Learning & Modeling: PyTorch, PyTorch Forecasting, neuralforecast, XGBoost, scikit-learn, SHAP, Optuna
• Data Engineering & Feature Pipeline: Pandas, NumPy, SQLAlchemy, Apache Airflow, custom calendar/holiday encoders, weather API integration
• API & Integration: FastAPI, PostgreSQL, ERP data connectors (ODBC + CSV pipelines), REST-based forecast publishing
• Infrastructure & DevOps: Docker, AWS EC2, S3, GitHub Actions, MLflow (for experiment tracking), Prometheus (basic monitoring)
• Visualization & Explainability: SHAP (global and local attribution), Plotly Dash (internal UI), Power BI (planner integration)

We delivered a production-grade forecasting platform deployed in a secure AWS environment, built to support long-horizon, high-granularity retail sales planning. Core components included:

• Multi-Model Forecasting Engine. The system used a hybrid architecture that combined N-BEATS, Temporal Fusion Transformer and XGBoost-LSTM ensembles and was optimized for annual forecasts for stable, volatile and ad-dependent product lines.
• Extended Feature Layer. Dynamic integration of calendar, weather, promotion, pricing, and regional metadata, enriched through lagged transformations and engineered covariates to enhance signal depth across 40,000+ SKUs.
• Model Training & Versioning Pipeline. Fully automated training DAGs orchestrated in Airflow, with cross-series validation, backtesting, and monthly retraining cycles. Drift triggers launched targeted retraining for volatile product segments.
• Explainable Forecast Outputs. Each prediction included SHAP-based driver attributions and attention heatmaps, enabling planners to visualize which factors influenced forecast curves at the SKU, category, and regional levels.
• Forecast API & Dashboard Connectors. The forecasting API and dashboard connectors included a FastAPI service layer that provided forecasts for downstream Power BI dashboards, with role-based access and override capabilities for business users.
• ML Monitoring & Observability. Integration of MLflow for experiment tracking, GitHub Actions for CI/CD and Prometheus-based uptime monitoring ensured full traceability and robust performance in both training and inference processes.
• Containerized Cloud Deployment. The system was fully containerized with Docker and deployed on AWS EC2, with an infrastructure-as-code setup, encrypted predictive storage and scalable retraining on demand.