From customer discovery to operational decisions.

Business problem

Forecast SKU-store demand at retail scale and turn those forecasts into optimized price recommendations that account for competitor position, inventory and commercial signals.

Data & scale

More than 200 stores, over 20,000 SKUs, competitor prices collected every two hours and more than 10 billion model-training records processed through the feature pipeline.

Two-stage approach

• Scalable ETL and leakage-aware feature engineering across SKU-store histories.
• Demand forecasting with temporal, lag, rolling, promotion, inventory and competitor-price features.
• Baseline comparison, automated model tuning and explainability for forecast quality.
• An optimization layer that converts forecast outputs and commercial constraints into dynamic price recommendations.

Measured forecasting result

The tuned forecasting model reduced RMSE by 21% versus the baseline. The stronger demand signal then feeds the price-optimization stage, separating model accuracy from the quality of the final pricing decision.

Production lens

The system is designed as two connected, independently evaluable stages: scalable demand forecasting followed by dynamic-price optimization across 200+ stores and 20,000+ SKUs.

Business problem

Capture ordered purchasing behavior that unordered basket analysis misses, then convert those patterns into better cross-sell, discovery and recommendation decisions.

Data pipeline

Hadoop/Spark ETL, SKU-to-taxonomy enrichment, anonymization, session and basket construction, then controlled execution through Spark MLlib and SPMF.

Approach

• Compare candidate-generation and projection-based algorithms.
• Evaluate main category, subcategory, brand and product granularity.
• Track runtime, memory, output size, pattern length and temporal stability.
• Version parameter configurations and auditable outputs.

Measured result

PrefixSpan completed within seconds while GSP required tens of minutes in the reported tests. PrefixSpan also retained much broader pattern coverage as granularity increased.

Production lens

The value was not only faster mining. The pipeline included threshold calibration, stability monitoring, resource envelopes and business-readable sequence outputs.

Business problem

Group warehouse orders to reduce spatial dispersion and operating effort while respecting batch capacity and SKU-level stock feasibility.

Model

A mixed-integer nonlinear formulation with deterministic allocation, stock constraints and a shared proximity-based objective evaluator.

Experiment design

• Three dataset sizes under a common evaluator.
• 600 objective evaluations per run.
• Ten independent runs for GA and PSO on the large instance.
• Best, average, worst, standard deviation and runtime comparison.

Measured result

GA delivered the strongest best-case solution quality; Local Search showed the lowest variability and strongest repeatability; PSO remained feasible but weaker under the shared budget.

Decision takeaway

Use GA where best-case quality matters and Local Search where stable, repeatable performance is the stronger operational requirement.