Advanced Forecasting and MLOps Implementation Workshop

Forecasting Fundamentals and Hierarchical Data

• Overview of forecasting methodologies
  • Traditional statistical methods (ARIMA, ETS, exponential smoothing)
  • Machine-learning/time-series ML methods (Gradient Boosting, Random Forests)
  • Survey deep-learning alternatives: LSTM, GRU and Transformer-based models; implement a selected approach rather than every architecture.
• Selecting the right technique for mid-term horizons
  • Definition of mid-term horizon in export/import/trade context
  • Trade-offs: interpretability vs accuracy vs operational cost
• Data requirements and quality considerations
  • Data ingestion: sources, frequency, granularity, missing values
  • Data quality: completeness, consistency, hierarchy issues
  • Handling hierarchical time‐series (multi-level) data structures — drill-down/roll-up issues; ensuring coherence across levels.
• Performance metrics and evaluation frameworks
  • Forecast accuracy metrics (MAE, RMSE, MAPE, sMAPE)
  • Check aggregate consistency of hierarchical forecasts alongside predictive error.
  • Baseline vs benchmark models; cross-validation for time-series
• Introduction to hierarchical forecasting frameworks in Python
  • HierarchicalForecast reconciliation examples: Bottom-Up, Top-Down and MinTrace; check the requirements of the chosen method.
  • Use of sktime for grouped/hierarchical forecasting.

Models and Feature Engineering

• Time series deep-learning approaches
  • LSTM and GRU architectures: when to use, how to configure
  • Transformer approaches for time-series forecasting: compare suitability, data needs and computational cost without assuming superiority.
• Ensemble approaches: evaluate whether combinations improve on the baseline.
  • Hybrid statistical + ML + deep learning ensembles
  • Bagging, stacking, blending in forecasting context
• Feature engineering for forecasting
  • Deriving time features (lags, rolling windows, seasonal indicators)
  • External factors: trade data, commodity flows, economic indicators
  • Encoding hierarchical identifiers (levels, country, commodity)
• Handling seasonality, trends, external factors
  • Decomposition (trend/seasonality/residual)
  • Incorporating external regressors (economic indicators, sentiment)
  • Dealing with hierarchical trends and cross-level interactions
• Practical hands-on: building a forecasting model in Python
  • From raw/hierarchical CSV or SQL data to model input
  • Training, tuning, validation, forecasting
  • Reconciling hierarchical forecasts to ensure coherence

Iterative Proportional Fitting and Constrained Flow Balancing

• Understanding the flow-match methodology and implementation challenges
  • Iterative proportional fitting adjusts a non-negative multidimensional table towards specified target margins; distinguish this balancing task from forecasting or generative flow-matching models.
  • Illustrative multi-level country/commodity trade flows: dimensions, compatible margins and hierarchy constraints.
• Optimization techniques for flow-matching algorithms
  • Specify constraints, target margins, convergence criteria and computational considerations.
  • Python techniques: matrices, sparse representations, reconciliation methods.
• Troubleshooting common flow-match issues
  • Data sparsity, missing levels, misalignment of hierarchies
  • Convergence issues, stability of reconciliation, scalability concerns
• Hands-on: Implementing flow-match optimization
  • Build a small IPF-based trade-flow balancing exercise in Python.
  • Compare the balanced flow table with forecast outputs and check the aggregation constraints explicitly.
  • Evaluate margin agreement, convergence and forecast error separately; improved coherence does not guarantee improved predictive accuracy.

Economic Indicators and Trade-Data Analysis

• Identifying key economic indicators for forecasting models
  • Macro indicators (GDP, PMI, Exchange rates, Commodity prices)
  • Sentiment indicators, trade policy indicators, export/import volumes
• Cross‐correlation analysis between economic indicators and historical trade data
  • Correlation matrices, lag analysis and Granger predictive tests; distinguish predictive association from causal conclusions.
  • Feature engineering: creating lagged indicator features, interaction terms
• Understanding market volatility patterns and their impact on forecasting
  • Volatility bursts, regime changes, structural breaks in time-series
  • Implications for forecasting; methods to detect and model volatility shifts
• Integration of economic sentiment and market indicators into forecasting pipelines
  • Design workflow to ingest indicator data (SQL/CSV), merge with trade time-series, feed into model
• Illustrative scenarios: economic shifts, trade patterns and forecasting error.
  • Use approved published or synthetic examples of disruptions; do not claim an unidentified client’s results.
  • Discuss how comparable patterns could be evaluated in a participant’s own setting.

Tracking, Deployment and Monitoring

• End-to-end prototype pipeline in a training stack: Python, SQL Server and Azure DevOps.
  • Architecture: data ingestion → feature engineering → model training → forecasting → deployment → monitoring
  • Connect hierarchical data and balancing/reconciliation steps; discuss feature management and experiment tracking.
• Model versioning and experiment tracking
  • Tracking hyper-parameters, model artifacts, dataset versions
  • Tools and best practices: MLflow, Git, Azure DevOps pipelines
• Automated retraining and model drift detection
  • Scheduling retraining workflows; detecting performance degradation over time
  • Handling concept drift, data drift in forecasting systems

• Production deployment strategies
  • Illustrative VM-based deployment with Azure DevOps; discuss container-free patterns where a target environment requires them.
  • Run the Python forecasting workflow, store outputs in SQL Server and review REST-service/Power BI integration options; do not assume model execution inside SQL Server.
• Data engineering & infrastructure
  • Architecture overview of real-time ingestion and Azure or custom ETL; the guided prototype uses a manageable training dataset.
  • Data validation and monitoring: schema checks, quality checks in pipeline
  • Discuss scaling patterns for on-premises or VM deployments rather than building a full scalable platform in the workshop.
  • Compare cloud-native patterns with environments that constrain containers.
• Monitoring & maintenance
  • Review performance alerts, dashboards and drift signals using prototype monitoring examples.
  • Compare a challenger forecast with a baseline using suitable time-aware evaluation; discuss controlled operational experiments where appropriate.
  • Alerting and incident response workflows: when forecast performance drops, when reconciliation fails
  • Continuous improvement workflows: how to collect feedback, update models, versioning, maintain pipeline health

Guided Build, Test Deployment and Handoff

• Assemble a guided forecasting pipeline prototype for illustrative hierarchical trade data.
  • CSV/SQL ingestion, selected features and model, forecast generation and output storage.
• Integrate and test the selected balancing or reconciliation step.
• Develop economic indicator correlation models and integrate into forecasting pipeline
• Deploy or demonstrate the prototype in an approved test environment using the Python/Azure DevOps/SQL Server/Power BI stack.
• Create selected monitoring examples for forecast error, drift or margin/reconciliation failures.
• Troubleshooting scenarios: pipeline failures, drift detection alarms, hierarchy reconciliation issues
• Review the prototype, known limitations and further engineering required before production.