Demand forecasting on top of Picqer: what good looks like

Picqer is excellent at one thing — running a warehouse. It tracks stock movements, generates pick lists, talks to carriers, and gives you a clean, current picture of inventory. By design, it does not try to predict the future. That's a deliberate choice, and a sensible one. But it leaves a gap.

If you run on Picqer, every reorder decision still depends on a forecast — even if the “forecast” is a finger in the air or a spreadsheet that someone in purchasing keeps. The question is whether that layer should keep being a spreadsheet or become an actual system. This post is about what that system should do, and what it should pointedly avoid doing.

What a forecasting layer should do

Read everything in real time. Picqer publishes webhooks for every product change, order, and stock mutation. A forecasting tool that polls overnight is operating on yesterday's data — fine for monthly reports, useless for stockout prevention. Voorcast subscribes to the relevant Picqer webhooks and recalculates affected SKUs as events happen. The supplier cart you open in the morning reflects what happened ten minutes ago, not last night's batch.

Decompose demand from your own history. A forecast that gives you one number is doing it wrong. You want level, trend, and seasonality teased apart so a one-off promotion doesn't pollute your seasonal pattern, and a genuine trend isn't dismissed as noise. Statistical decomposition has been around for decades; it works.

Score confidence honestly. Specialty SKUs have thin sales history. Some products genuinely cannot be forecast. The right move is to surface that: confidence labels from insufficient through mature, transparent enough that the buyer knows when to trust the recommendation and when to use judgement.

Group risk by supplier. Reorder decisions are placed with suppliers, not with individual SKUs. The right unit of attention is the supplier card: every at-risk SKU under that supplier, the profit at stake, the recommended quantities, and a single button to send.

What it should not do

Demand configuration. A forecasting tool that asks you to set a dead-stock cutoff, a stockout alert level, and a best-sellers list is one that has pushed its analytical work onto you. Those numbers are derivable from your own data — percentiles of inter-sale gaps, sales velocity distributions, supplier lead-time variance. The tool should compute them, not ask you for them.

Replace Picqer. Stock movements, picking, shipping — that's Picqer's job, and it's good at it. A forecasting layer that tries to mirror Picqer's data model and become the source of truth for stock is duplicating effort and adding sync risk. It should treat Picqer as authoritative for warehouse state and add a layer of prediction on top.

Pretend AI is the answer. A lot of recent tools wrap a language model around their forecasting layer and call it AI. The forecasting itself should be statistical and explainable. AI is useful at the edges — anomaly explanation, plain-language summaries of why a SKU was flagged — but the forecast itself should be something you can audit.

What this looks like in Voorcast

When you connect Picqer, Voorcast pulls your full product, order, and stock-movement history through the API, then subscribes to webhooks for everything that follows. The forecasting engine decomposes seasonality from your own sales, derives every threshold from percentiles in your data, and continuously rescues those thresholds as your catalogue and seasonality shift.

You don't configure dead-stock cutoffs. You don't maintain a best-sellers list. You don't set alert thresholds per SKU. You see stockout risk grouped by supplier, with profit at risk and auto-filled supplier carts. You act on what's surfaced.

That's what a forecasting layer should be: a quiet system that takes the analytical work off your team's hands and surfaces only the decisions that need a human.