◉ HASSELT · BEBook →
← CASE STUDIES·E-COMMERCE · NL

Veloport stopped guessing stock levels and recovered €190k in twelve months.

A gradient boosting model trained on 3 years of sales history, seasonal signals, and external events. Weekly per-SKU forecasts replace the manual Excel reorder process entirely.

-31%
STOCKOUT INCIDENTS
€190k
OVERSTOCK RECOVERED
14
DAYS TO PRODUCTION
/ THE SITUATION

The buying team was ordering on instinct. Both ends of the problem showed up in the P&L.

Veloport's buying team managed reorders through a shared spreadsheet updated weekly by hand. Fast-moving SKUs ran out ahead of peak periods. Slow-movers piled up in the warehouse. The same team was simultaneously fire-fighting stockouts and writing down overstock, a problem that was getting worse as the catalog grew past 4,000 active lines.

/ WHAT WAS BUILT

A model that reads the signals before demand moves.

Sales history alone was not enough, the model needed to see what was coming. External signals turned unpredictable spikes into forecastable patterns. The buying team now reviews a ranked list of reorder recommendations, not a raw spreadsheet.

01

Historical data extracted

Three years of per-SKU daily sales pulled from Snowflake alongside promotional calendars, holiday flags, and competitor pricing signals.

02

External signals added

Weather data, regional events, and Google Trends indices joined to the training set. Demand spikes that looked like noise turned out to be predictable patterns.

03

XGBoost model trained

Gradient boosting model trained per product category with shared feature engineering. Validated on a 12-week holdout before any production run.

04

Weekly forecast pipeline

FastAPI endpoint called every Monday morning by the buying team's planning tool. Forecast horizon of 4 weeks per SKU, refreshed with the latest sales data each run.

05

Reorder trigger integrated

Model output piped directly into the reorder workflow, purchase orders drafted automatically when forecast minus current stock drops below safety threshold.

06

Accuracy tracked in-dashboard

Forecast vs. actual tracked weekly in Looker. Model retrained quarterly or whenever MAPE drifts above the agreed threshold.

/ RESULTS

Less stock out. Less stock sitting. Better margin.

-31%

Stockout incidents

Reduction in out-of-stock events across the top 500 SKUs in the first full quarter post-deployment.

€190k

Overstock recovered

Value of inventory held above safety stock that was brought back within target range within 12 months.

4 wks

Forecast horizon

Per-SKU weekly forecast updated every Monday. Buying team acts on a rolling 4-week view, not a gut feel.

14

Days to production

From first data handoff to a live FastAPI endpoint serving the planning tool.

/ STACK

The smallest stack that solved the problem.

XGBoostPythonSnowflakeFastAPILooker StudioGoogle Trends API
/ YOUR SYSTEM

Replace gut feel
with a model.

30-minute call. We'll audit your data and tell you exactly what is and isn't ready for a custom ML model, free, no deck.