# Demand forecasting software that reaches the store order

_For food retailers with 50+ stores_

A forecast for every product, store and day that becomes the order quantity in the system you already use. Fresh, bakery and food-to-go included.

A few stores with steady demand? Your ERP's auto-ordering is enough

**In short.** Demand forecasting software predicts how much of each product each store will sell on each day and turns that into how much to order or produce. Large grocers buy platforms such as RELEX or Afresh. Chains with 50 to 300 stores often already have auto-ordering in their ERP that works from average sales and reacts late to holidays, promotions and differences between stores. dev.family builds a machine-learning forecast that replaces that one number inside your existing ordering, tests it against your current system on your own history, and runs it every night.

## What's inside a forecast that reaches the order

- **Product × store × day**: Gradient boosting on sales, prices, promotions, holidays and store attributes. New items get a forecast, not a blank.
- **The quantity, not just a forecast**: Pack sizes, delivery schedules and minimum stock stay your rules.
- **The shelf, not just accuracy**: We simulate stock day by day for both systems on the same history.
- **Inside your infrastructure**: Sales and stock arrive after midnight. Orders reach your ERP before the morning deadline.
- **Your current ordering as backup**: If a run is late or incomplete, your existing auto-ordering places the order.
- **Every order explained**: What the model proposed, what staff changed, what was sent and what arrived.

## What demand forecasting costs and how long it takes

Forecasting platforms mostly quote on request and replace your ordering process. A forecast inside your current ordering starts with a fixed-scope pilot.

- Forecasting platforms (RELEX, Afresh, ToolsGroup, Logile): Pricing on request. Source: Vendor sites, checked November 9, 2026, checked 2026-11-09
- Inventory and demand planning (Netstock): From $900 a month, billed annually. Source: [netstock.com/pricing, checked November 9, 2026](https://netstock.com/pricing), checked 2026-11-09
- Forecasting service (Lokad): A flat monthly fee that starts in the low thousands of dollars or euros. Source: [lokad.com/contractual-agreement, checked November 9, 2026](https://lokad.com/contractual-agreement), checked 2026-11-09
- What empty shelves and overstock cost retail: $1.73 trillion a year, 6.5% of global retail sales. Source: IHL Group via Path to Purchase Institute, September 26, 2025, checked 2026-11-09
- Forecasting Pilot: 4–6 weeks : a data audit, a forecast for one category, accuracy before and after, ordering at one store. Source: dev.family, checked 2026-11-09
- Rollout across the chain: 8–16 weeks after the pilot, fixed estimate after the pilot. Source: dev.family, checked 2026-11-09

## One number changes. The process stays

The model sits between your data and your ordering rules.

Sales, Stock and in transit, Prices and promotions, Calendar → dev.family model: forecast and order quantity for every product, store and day → Staging database checks, Your ERP rules, Orders to DC and suppliers

Built on the same forecast:

## Works with the ERP and data you already have

The model reads from your ERP or warehouse and hands orders back the same way.

SAP · Oracle Retail · Microsoft Dynamics 365 · NetSuite · Snowflake · BigQuery · Your data warehouse

Another ERP? Orders reach it through a staging database it can read.

## Change one number in your ordering, not the whole process

Most auto-ordering works like this: days of stock times average daily sales, minus stock on hand and in transit, rounded to the supplier's pack. The weak spot is the average. We replace it with a forecast and leave everything else alone: pack sizes, schedules, minimum stock and warehouse checks stay your rules, and your current system stays as a fallback.

Daily sales, Stock on hand, In transit → Forecast model (replaces only the average; your current auto-order stays as the fallback) → Your ERP rules, Store orders

## ERP auto-ordering, a forecasting platform or a forecast inside your ordering

| Criterion | ERP auto-ordering | Forecasting platform | Data science consultancy | dev.family |
| --- | --- | --- | --- | --- |
| Pricing | Included in the ERP | On request, enterprise contracts | Project fee | Fixed price per scope |
| Promotions, holidays, store differences | Limited | Yes | Yes | Yes |
| Works inside your current ordering | Yes | Replaces it | Model only | Yes |
| Runs every night after launch | Yes | Yes | Project ends | Run under an SLA |

## When a platform is the better call

If you're a large grocer replacing your whole replenishment process, RELEX, Afresh or a similar platform fits. If you have a few stores and steady demand, your ERP's auto-ordering is enough. We fit in between: chains with working auto-ordering that need a better number inside it.

## A forecast that beat the chain's own auto-ordering

We built a machine-learning model that forecasts sales of every product in every store for every day and turns them into order quantities. We ran the chain's real sales history through both systems, our model and its own auto-ordering, on the same stores, covering about 90% of its automatic orders. The model ordered closer to actual sales, 93.8% vs. 88.8%, and inventory turnover didn't move. Preliminary results, before final sign-off:

## Where it usually hurts

- **Average availability hides best sellers.**: A store with 30 slow items on the shelf and one best seller sold out can still show about 97%. We report lost units next to availability.
- **The model lives with pack sizes.**: A forecast of 12 is worth nothing if the supplier ships 20 to a case. Pack sizes and minimum stock belong in the test, or the model looks better on paper than in the store.
- **Simulate the shelf, not the forecast.**: Forecast accuracy and what happens on the shelf are different questions. Only a day-by-day stock simulation answers the second.

From forecasting for a 650-store retail chain. Only what we've seen ourselves.

## How we run the forecast after launch

- **Nightly run**: Input checks every night: dropped products, stale stock, unusually large orders.
- **Retraining**: Retrained on fresh data, with one version frozen during tests.
- **Errors**: Reviewed with your team, order by order.

## Questions

### What is demand forecasting?

Demand forecasting estimates how much of each product customers will buy, where and when. In retail, that means a number for every product, store and day, which then drives orders to the distribution center and suppliers, or a production plan for fresh food.

### How do you do demand forecasting for a retail chain?

Start from daily sales by product and store, then add what moves demand: prices, promotions, holidays, weather and store attributes. A machine-learning model learns those effects from history. The forecast only pays off when it changes the order, so the last step is turning it into an order quantity within your pack sizes, delivery schedules and minimum stock.

### How does forecasting help a business avoid losing money?

Order too little and you lose sales on empty shelves. Order too much and fresh food is written off or marked down. A better forecast moves both at once: for a 650-store retail chain, our model cut lost sales by 37% at the same inventory turnover (preliminary, before final sign-off).

### How do demand forecasts integrate with ERP and planning tools?

The forecast runs at night after sales and stock arrive, and the order quantities go to your ERP through its API or a staging database before the ordering deadline. Pack sizes, schedules and minimum stock stay rules in your ERP.

### Can machine learning run on top of our existing auto-replenishment?

Yes. In the 650-store chain, the model replaced one number in the chain's own ordering process: the store's expected demand. Pack sizes, delivery schedules and minimum stock stayed the chain's rules, and staff kept working the way they did before.

### How do you test the model before it places real orders?

In two steps. First, both systems order on your real sales history while we simulate the shelf day by day. Then the model orders for real in a few test stores, compared with matched control stores that keep your current system, and any late or incomplete run falls back to the old ordering automatically.

### What does a one-category pilot prove?

Whether your data can carry a forecast and how much it beats your current ordering on your own history, before you commit to a rollout. The Forecasting Pilot takes 4–6 weeks: a data audit, a forecast for one category, accuracy before and after, and ordering at one store.

### What data does the model need?

Daily sales, stock and goods in transit by product and store, prices, promotions, store attributes, pack sizes, delivery schedules and minimum stock levels. Two years of history is ideal; less still works for a pilot.

## Let's check whether your data can carry a forecast

Tell us your stores, categories and ERP. On the call, we'll tell you what a pilot on one category would show.

## Who this page is for

- **[Fresh food retail: bakery, food-to-go, grocery](https://dev.family/foodtech/food-retail)**
- **[Food manufacturers](https://dev.family/foodtech/food-manufacturers)**
- **Fresh forecasting benchmark**

---

Last updated: 2026-11-09
Page: https://dev.family/foodtech/demand-forecasting
Contact: https://dev.family/project-evaluation