What is demand forecasting from historical sales data?

Demand forecasting from historical sales data is the process of analyzing past unit sales — typically at the SKU level — to predict future customer demand. Before applying any forecasting model, practitioners must clean the data by identifying and flagging anomalies such as stockouts, promotional spikes, and receiving errors that distort the true demand signal.

Why should demand forecasting be done at the SKU level instead of category level?

Forecasting at the category level averages out the individual signals of each SKU, hiding demand patterns that matter for replenishment decisions. SKU-level forecasting preserves velocity differences, seasonal behavior, and anomaly events that would otherwise be lost in an aggregate view, leading to more accurate and actionable inventory plans.

How do you clean historical sales data before building a forecast?

Cleaning historical sales data involves going through each period and flagging events that distorted real demand — including stockouts that suppressed sales, promotions that inflated volume, one-time bulk orders from customers who will not reorder, and receiving errors that logged inventory incorrectly. Each flagged period is then adjusted or excluded so the model trains on genuine demand rather than operational noise.

How much historical data do you need for seasonal demand forecasting?

Seasonal decomposition requires a minimum of 24 months of clean historical sales data before it produces results worth trusting. With fewer than two full years of history, the model cannot reliably separate repeating seasonal patterns from random variation or one-time events.

What is an acceptable MAPE for a demand forecast?

For core SKUs, a mean absolute percentage error (MAPE) above 30 percent indicates the forecast model is not ready for operational use. If validation against a holdout period of two to three months produces a MAPE at or above that threshold, the correct response is to return to the data cleaning step rather than switching forecasting algorithms.

Building a Forecasting Model with Historical Sales Data: A Walkthrough

Yuneva Stock Count
Apr 30
2 min read

Analyst reviewing historical sales data on a laptop to build a demand forecast

Building a demand forecast from historical sales data sounds like a data science project. It's not. It's mostly just cleaning up your own mess.

Here's what I mean. When you pull 12 months of sales history to start a forecast model, the first thing you'll find is that three of those months are lying to you. A stockout in February made it look like demand dropped. A promotional push in Q3 inflated one SKU by 40%. A receiving error in October logged 200 units that never actually moved. If you feed that raw data into any model without flagging those events, your forecast inherits every mistake your operation made over the past year. Garbage in, slightly more confident garbage out.

The actual walkthrough goes something like this. Start by pulling unit sales at the SKU level, not category level. Aggregate too early and you'll average away the signals you actually need. Then go period by period and mark the anomalies — promotions, stockouts, one-time bulk orders from a customer who will never reorder. Most practitioners call this "cleaning" but what you're really doing is deciding which past demand was real and which was noise. That judgment call matters more than which forecasting method you pick later.

Once the data is clean, you choose a method that fits the pattern. Stable, predictable movers? Weighted moving average with more weight on recent periods works fine. Anything seasonal? You need at least 24 months of history before seasonal decomposition gives you something worth trusting. Sporadic, low-velocity items are a different problem entirely — most standard models handle them poorly, and sometimes a simple average with a manual buffer beats a complex algorithm that confidently predicts 1.3 units.

The last thing people skip is validation. Hold back the most recent two or three months before you build the model, then run your forecast against what actually happened. If your mean absolute percentage error is sitting above 30% on your core SKUs, the model isn't ready — go back to the cleaning step, not the algorithm.

A forecast isn't a magic number. It's a structured guess you can defend, adjust, and improve over time. That's worth building correctly.

Yuneva builds tools for the operations side of inventory — take a look at what they've put together at www.yuneva.com. If you're counting stock on the floor, www.count-inventory.com is where CountIt lives.

#Yuneva #InventoryManagement #SupplyChain #DemandForecasting #WarehouseOps

Building a Forecasting Model with Historical Sales Data: A Walkthrough

Recent Posts

Comments