The situation
The retailer (publicly listed, ~5,000 stores worldwide) loses about EUR 80M per year to mid-season markdowns; a 10 percent MASE improvement on the top-500-SKU panel would materially reduce that.
Your task
Pick the demand-forecasting approach that gives the best MASE improvement over the SARIMA baseline at acceptable training cost.
The brief
You start with
You receive 24 months of daily sales for 500 SKUs across 200 stores, plus calendar features (holidays, promotions, weather codes).
What you do
- Forecast 14 days out per SKU-store.
- Benchmark SARIMA, LightGBM (with lag + calendar features), and a small N-BEATS or Temporal Fusion Transformer (TFT) on a rolling-origin evaluation.
- Report Mean Absolute Scaled Error (MASE) and bias by horizon.
You deliver
The benchmark notebook, a results table, and a 3-page recommendation memo.
What you will practice
- Apply SARIMA, tree-based, and neural forecasters to real retail data
- Evaluate forecasts with rolling-origin folds (not random split)
- Quantify bias and error by forecast horizon
- Communicate a forecasting recommendation to a supply-chain audience
A suggested way in
- Rolling-origin evaluation: never use a random split for time series
- MASE normalizes against a naive baseline; report it instead of MAPE on intermittent SKUs
- Always slice errors by horizon (1-3, 4-7, 8-14 days); average hides the long-horizon story
Skills you will practice
Time series forecasting, Sarima, Gradient boosting, Neural forecasting, Rolling evaluation, Python
Where it leads
Data Scientist, Machine Learning Engineer, Applied AI Scientist
How it is assessed
- Evaluation correctness30%
Rolling-origin evaluation done without leakage.
- Model comparison rigor25%
All three families tuned with comparable budgets.
- Horizon analysis20%
MASE + bias reported per horizon slice.
- Recommendation clarity25%
Supply-chain lead can decide direction from the memo.
Resources
- PythonTool
Standard for the benchmark harness.
- statsmodelsTool
SARIMA implementation.
- LightGBMTool
Tree-based forecaster with lag features.
- PyTorchTool
Implement N-BEATS or TFT baseline.
- MatplotlibTool
Plot horizon-wise errors and failure modes.
- NotionTool
Compose the recommendation memo.
Finish it. Earn this.
A verified certificate with your name. Anyone can check it.
Time to practice.
Solve this challengeMore like this.
- Data ScientistExplain a Credit-Risk Model with SHAP for a Fintech
- Senior AccountantIFRS 16 Lease Accounting for a Renewable Energy Firm
- Software EngineerODE Integrator Benchmark for a Pharma PK/PD Pipeline
- Data ScientistInterpretable-by-Design GAM for an Insurer's Claims Triage
-
Explain a Credit-Risk Model with SHAP for a Fintech Data Scientist You receive a trained XGBoost credit-risk model (binary default prediction), the training feature schema (38 features), and a held-out 10,000-sample t shapinterpretabilityfairness analysis Open challenge → -
IFRS 16 Lease Accounting for a Renewable Energy Firm Senior Accountant You are a senior accountant at SunPower Renewables. The company is adopting IFRS 16 for the first time. You are given a sample lease for a wind turbin ifrs 16lease accountingright of use asset Open challenge → -
ODE Integrator Benchmark for a Pharma PK/PD Pipeline Software Engineer Use SciPy's solve_ivp + SUNDIALS (via scikit-sundae or diffeqpy) to benchmark: RK45 (baseline), LSODA, BDF (CVODE), Radau, and Rosenbrock. Run on thre numerical odesstiff integratorsscientific python Open challenge → -
Interpretable-by-Design GAM for an Insurer's Claims Triage Data Scientist You receive an anonymized claims dataset (around 60,000 claims, target: log reserve), a feature schema (22 features), and an existing LightGBM baselin generalized additive modelsebminterpretability Open challenge →