ML-Based Demand Forecasting for Retail Supply Chain
A 200-store retail chain was experiencing 12% average out-of-stock rates and high end-of-season markdown losses due to inaccurate demand forecasting. Their existing statistical models did not account for local events, weather patterns, or promotional calendars.
Client
Confidential, Retail Chain Client
Industry
Retail, D2C & E-CommerceTeam Lead
Sophie Müller
Date
Q1 2024
Business Challenge
A 200-store retail chain was experiencing 12% average out-of-stock rates and high end-of-season markdown losses due to inaccurate demand forecasting. Their existing statistical models did not account for local events, weather patterns, or promotional calendars.
Solution & Approach
We built an ensemble forecasting platform combining Facebook Prophet for trend/seasonality decomposition with LightGBM for feature-rich SKU-location prediction. The model incorporates external signals (weather, local events, competitor promotions) and is retrained weekly via Airflow.
Capabilities Delivered
- Ensemble demand forecasting using Prophet and LightGBM
- External signal integration (weather, events, promotions)
- SKU-location-level weekly retraining pipeline in Airflow
- Buyer-facing replenishment recommendation dashboard
- Forecast accuracy tracking and model performance monitoring
Business Outcomes
Average out-of-stock rate reduced from 12% to 4.3%, and end-of-season markdown losses decreased by 28%. The procurement team's forecast accuracy (measured by MAPE) improved from 31% error to 11% error.
4.3%
Out-of-stock rate (down from 12%)
28%
Reduction in markdown losses
11%
Forecast MAPE (down from 31%)
Related Projects
AI-Powered Assessment and Learning Analytics Platform for EdTech
Course completion rates improved from 38% to 61% within one semester. Learner satisfaction NPS increased from 31 to 58, and the platform reduced monthly subscriber churn by 2.8 percentage points.
61%
Course completion rate (up from 38%)
58
Learner NPS (up from 31)
Hospital Operations Digitization Platform
Patient discharge time reduced from 4.5 hours to 1.8 hours. Clinical documentation completeness improved from 67% to 96%, eliminating the data gaps that were triggering insurance claims rejections. Staff productivity scores improved by 34%.
4.5h→1.8h
Patient discharge time
96%
Clinical documentation completeness
Last-Mile Logistics Route Optimization Platform
Delivery SLA compliance improved from 76% to 94%, fuel costs per delivery dropped by 18%, and driver productivity reached 88% of theoretical capacity. The client expanded to 3 new cities within 6 months using the same platform.
94%
Delivery SLA compliance
18%
Reduction in fuel cost per delivery
AR-Based Training Simulation for Manufacturing Operations
Training time per operator reduced by 40%, and training-related safety incidents dropped to zero. Competency certification scores improved by an average of 22 percentage points. The client is expanding the platform to 2 additional facilities.
40%
Reduction in training time
0
Training-related safety incidents
ML-Based Demand Forecasting for Retail Supply Chain
Average out-of-stock rate reduced from 12% to 4.3%, and end-of-season markdown losses decreased by 28%. The procurement team's forecast accuracy (measured by MAPE) improved from 31% error to 11% error.
4.3%
Out-of-stock rate (down from 12%)
28%
Reduction in markdown losses
End-to-End Supply Chain Visibility and Control Tower
Supply chain disruption response time improved from 4 days (detection lag) to 6 hours. Stockouts attributable to supply chain visibility gaps were reduced by 71%. The procurement team avoided $1.8M in expedited freight costs in the first year.
4d→6h
Disruption detection response time
71%
Reduction in visibility-gap stockouts