Daniel Costa

Own store-level demand forecasting powering replenishment for 140 stores and roughly 30,000 SKUs.

• Replaced a legacy statistical pipeline with a gradient-boosted hierarchy model, improving SKU-level WAPE from 34% to 24%.
• Forecast improvements reduced perishable waste 18% and out-of-stocks 11%, worth an estimated €6M a year.
• Built drift monitoring and automated backtesting; silent model degradations detected in hours instead of weeks.
• Run the quarterly forecasting review with supply-chain leadership, translating model behaviour into buying decisions.

Short-term load and renewables forecasting for an energy retailer balancing a 400MW portfolio.

• Improved day-ahead load forecast MAPE from 4.1% to 2.9%, cutting balancing costs by roughly €1.2M a year.
• Built a probabilistic solar-production model used by the trading desk for intraday position sizing.
• Introduced feature-store patterns that cut new-model iteration time from weeks to days.

Customer and assortment analytics for retail clients.

• Built a churn-scoring model for a telecom client that lifted retention-campaign conversion 2.3x versus random targeting.
• Automated the weekly client reporting stack, freeing about one analyst-day per week.

Python library implementing MinT and bottom-up reconciliation with a scikit-learn-compatible API.

• 700+ GitHub stars; used in at least four production retail forecasting stacks per user reports.
• Benchmark suite reproduces published reconciliation results on three public datasets.

Hands-on evening workshop on production forecasting, run twice a year at a local data community.

• Taught 90+ practitioners across four cohorts, rated 4.8/5 on average.
• All materials open-sourced, including a realistic synthetic grocery dataset.

Thesis on hierarchical reconciliation methods for retail demand forecasting.

Certification covering ML pipeline design, deployment, and monitoring on managed cloud infrastructure.