Domain-driven Data Science
summary
Argues domain know-how matters more than extra model complexity for supply‑chain forecasting, showing concrete techniques: probabilistic forecasts to size safety stock, modeling lead‑time variability, and choosing granularity (SKU × market × cadence). Uses exponential smoothing, Croston, LightGBM and cites the M5 competition; mentions Python time‑series tooling (Nixla) and a Streamlit UI used to demo an auto‑initializer that is provably better than Excel but fails adoption until redesigned for domain workflows.
AI-written summary · about
chapters (22)
- 0:00Introduction to domain-driven data science
- 3:20Why domain expertise matters
- 4:46Domain-driven design inspiration
- 5:50Logistics and supply chain overview
- 8:20Modern supply chain challenges
- 10:29Demand forecasting fundamentals
- 12:17Uncertainty and safety stock
- 15:27Time series forecasting basics
- 18:43Exponential smoothing models
- 20:09Evaluation metrics in forecasting
- 21:31M5 competition insights
- 22:37Domain-specific forecasting effects
- 25:27Making data science domain-driven
- 26:52User trust in forecasting models
- 27:33Domain insights in data exploration
- 29:20Simple heuristics over ML complexity
- 30:44Learning from domain experts
- 32:09Documenting business domain essentials
- 33:53Resources for domain learning
- 35:19Inspiration from domain practices
- 36:22Summary and call to action
- 37:30Q&A on domain knowledge limits
related talks
22mRonald Menich, Chief Data Scientist, Predictix, LLC
37mForecasting using machine learning methods
28mUse case in demand forecasting
28mMastering Demand Forecasting: Lessons from Europe's Largest Retailer
59mPredictability Check: Demand Forecasting
29mBecoming a data-driven enterprise
02h 00mData Science life cycle from a technical perspective
43mData Modelling for Startups: Domain Models & the Modern Data Stack