SKU × store forecasts
Daily or weekly forecasts for every reference in every point of sale, refreshed with each ERP import.
Forecast
AI demand forecasting per SKU and per store
The right model for every reference, without a data scientist
IRIS trains and compares several models per SKU × store (Prophet, LightGBM, ARIMA, ETS, XGBoost) and keeps the most accurate. Seasonality, Ramadan and Eid, launches and cannibalisation are built in.
Who uses it
- Supply chain director
- Demand planner
- Brand manager
What is inside
Automatic model selection
Each series gets the model that back-tests best. Accuracy (MAPE) and bias are tracked and shown per SKU.
Seasonal profiles and calendar
Ramadan, Eid, summer, back-to-school, Black Friday and your own events, with year-on-year comparison.
New-product forecasting
Comparable products and a launch ramp for references without history; switch to own history after a few weeks.
Cannibalisation
Launches and promotions reduce the forecast of the references they replace, so you do not over-order both.
Manual override
Planners adjust a forecast with a reason; overrides are logged and their accuracy is measured.
The module, screen by screen
01
Seasonal profiles per category
Fragrance, skincare and make-up each get their own curve, learned from history and shown month by month. The planner sees why the model expects a peak before it orders for it.
Profiles per category and per store
Ramadan and Eid aligned on the lunar calendar
Refresh and cleanup with one click
02
Event and promotion calendar
Launches, campaigns, holidays and store openings are entered as events with a start, an end and an expected uplift. The forecast integrates them before the order is proposed.
Events by brand, category or store
Expected uplift entered by the brand manager
Calendar, list and timeline views
03
Forecast method and horizon
The ensemble method compares Prophet, LightGBM, ARIMA, ETS and XGBoost per series and keeps the most accurate. Horizon and smoothing are set once per network.
Ensemble with automatic model selection
Forecast horizon in months
Settings applied to every new forecast run
04
Cannibalisation for launches
A new product is linked to the references it will replace, with a default cannibalisation rate. The old references' forecasts drop accordingly, so both are not over-ordered.
New product and impacted products
Adjustable cannibalisation percentage
Paste the list of impacted SKUs from Excel
Key benefits
Fewer stock-outs on best-sellers
The references that sell the most get the model that fits them, refreshed with every import.
Less overstock on slow movers
Cannibalisation and launch ramps stop the double order on the old and the new reference.
Seasonality handled by data
Ramadan, Eid, summer and back-to-school profiles are learned per category, not guessed.
Accuracy you can show
MAPE and bias per SKU and store, so the forecast is a shared fact, not an opinion.
Planners stay in control
Overrides with a reason, logged and measured against the model.
No data science team needed
Model selection, retraining and monitoring run automatically.
What it changes
-40 to -50 %
stock-outs on best-sellers
MAPE tracked
per SKU and per store
12 to 24 months
of history is ideal, 6 is enough to start
How it works
Sales history and item master are imported from your ERP.
Models are trained and compared per SKU × store; the best is kept.
Forecasts feed the replenishment engine and every report.
Connected to your ERP
Fed by your ERP, several times a day
Invoiced sales, item master, stock by warehouse and open supplier orders are imported from Sage, Odoo, Dynamics 365, SAP or NetSuite. The forecast is retrained on a schedule and after every large import.
Implementation
Implementation
The same plan on every project. Your team spends a few hours a week; IRISYS does the rest.
- Week 1
Connect and check
ERP connector or CSV / SFTP export. Item master, stock by warehouse, sales history, open purchase orders. Data quality report.
- Week 2
Calibrate
Models trained on 12 to 24 months of history. Accuracy reviewed by category with the planner. Rules set: safety stock, service-level targets, case packs.
- Week 3
Run in parallel
IRIS proposals compared every morning with the current process. Differences explained; rules and events added.
- Week 4
Go live
Proposals validated in IRIS and created in the ERP. Thresholds opened category by category. On-site or remote training.
Questions about this module
Twelve to twenty-four months is ideal. Six months is enough to start; references with less history use comparable products.
Yes. Overrides are entered with a reason, logged, and their accuracy is measured against the model.
Yes. Ramadan, Eid al-Fitr and Eid al-Adha move every year; IRIS aligns history on the lunar calendar before training.
By MAPE and bias per SKU and per store over a rolling window, shown in the accuracy report and used to decide when a model is replaced.
Yes, per SKU, store and period to CSV or Excel, and through the REST API for your BI tool.
Ready to see your own numbers in IRIS?
Send us an anonymised sales export. We show you your forecasts and order proposals within 10 days.