Forecasts that don't stop at a dashboard - they order the stock.
A national retailer was forecasting in spreadsheets and paying for it twice: stockouts on the fast movers, working capital buried in the slow ones. Welzin built a demand-forecasting system and wired it straight into replenishment, so the prediction becomes the purchase order.
An empty shelf and an overstuffed warehouse are the same mistake.
Buying decisions for hundreds of stores ran on last month's sales, gut feel, and a spreadsheet that broke every promotion week. The fast movers sold out by Friday while slow movers sat on the balance sheet for months. Nobody could say, SKU by SKU and store by store, what next week actually looked like.
Chronic stockouts
Best sellers went dark for days at a time. Every empty facing was a sale handed to the store across the road.
lost salesCapital on the shelf
Over-ordering the slow movers buried cash in inventory and ended in markdowns nobody had budgeted for.
frozen cashPromotions broke everything
A festival week or a price cut made history useless. The spreadsheet had one answer for every week: last month, again.
One forecast for everyone
A metro flagship and a small-town store got the same buying logic, though their demand had nothing in common.
A forecast for every SKU, every store, every day.
Six pieces work together: clean features in, calibrated probabilistic forecasts out, and a replenishment integration that turns the numbers into orders - with the monitoring to keep it honest month after month.
A feature store, not a spreadsheet
Sales history, promotions, price changes, holidays, weather, and store attributes land in one governed feature store - computed once, reused by every model.
Hierarchical forecasting
Gradient-boosted models forecast at SKU x store, then reconcile up through category, region, and network so the numbers agree at every level of the business.
Uncertainty you can order against
Every forecast ships as quantiles, not a single point - so safety stock reflects real demand risk instead of a blanket buffer.
Straight into replenishment
Forecasts flow into the ERP's replenishment run as suggested orders. No export, no re-keying - the prediction becomes the purchase order.
An exception workbench
Planners review only the forecasts that need judgment - new SKUs, big swings, low confidence - each with the drivers behind the number, not a black box.
Backtests, drift watch, retraining
Rolling backtests gate every model release, and drift monitoring triggers retraining before accuracy decays - the same rigor we apply to any system we run.
A nightly run and a weekly loop, both boringly reliable.
The system does its work while the stores sleep. By the time buyers sit down with their coffee, the orders are drafted, the exceptions are queued, and yesterday's accuracy is already on the scoreboard.
Every night
From yesterday's sales to today's suggested orders.
- IngestPull the day's sales, stock positions, and promo calendar from source systems.
- ScoreRefresh features and generate forecasts for every SKU-store pair in the network.
- ReconcileAlign the forecasts up the hierarchy so store, region, and network totals agree.
- Draft ordersConvert forecasts plus lead times and safety stock into suggested replenishment.
- Flag exceptionsRoute low-confidence and high-impact lines to the planner workbench.
Every week
Keeping the model honest as the business moves.
- MeasureScore last week's forecasts against actuals, by category and by store.
- Watch for driftCompare live accuracy and feature distributions against the backtest baseline.
- Retrain when it paysRefresh models when drift or new patterns justify it - not on a superstition schedule.
- Review togetherWalk the metric with the planning team: what moved, why, and what to tune next.
Built on the systems the retailer already runs.
No rip-and-replace. The forecasting system reads from the existing data estate and writes into the existing buying workflow, so adoption was a login, not a migration.
Data sources
Warehouse & pipelines
Modeling
Downstream
MLOps
Fuller shelves, lighter balance sheet.
The metric we signed up for was availability without inventory bloat - and both moved. Figures below reflect observed outcomes in production and are directional, not guarantees.
The buyers stopped arguing about whose number was right and started arguing about what to do with it. That's when we knew the forecast had become the system.
