We small warehouses cannot afford that kind of money. Last year a vendor came to pitch us an AI demand forecast system: tens of thousands a year, plus three months of "data cleansing" before we could even start. I laughed. Our warehouse is about 3,000 square meters, two staff plus me, the owner, and all our data already lives in Excel. Three months of cleansing, and my shelves would have stocked out twice over.
So we took the scrappy route: no new system, just Excel and a simple model to get AI demand forecast running. After half a year of it, stockouts dropped and overstock shrank. Here is exactly how we did it, so other small operators can copy it.
The small-warehouse problem: demand forecasting sounds great, but the enterprise version does not fit us
Some background. In the summer of 2025, our warehouse carried under 200 SKUs, mostly pet supplies and small home goods, shipping 4,000 to 6,000 orders a month. Small, but the problems were real: bestsellers stocked out without warning, while slow movers gathered dust on the shelves.
The worst case was July 2025. A cat scratcher went viral in short videos and sold a month's volume in three days. We reordered by gut feel, the replenishment was still in transit, and the shelf sat empty for four days. Complaints piled up and the store rating dipped. That week still stings. In the opposite direction, we once brought in 800 pet beds that took six months to sell through half of — rent and cash tied up the whole time.
That is the small-warehouse reality: the enterprise version of AI demand forecast talks about data platforms, algorithm teams, and live integrations. We have none of that. But forecasting itself does not need heavy equipment.
Step one: get your sales history into one readable table
Do not start with algorithms. The first step of forecasting is getting data into a shape a human can read. Our method is embarrassingly low-tech, and it works:
First, pull the last 12 months of shipments by SKU and by week — one row per SKU per week. No fancy system needed; a pivot table over the order records does it in an afternoon.
Second, flag the weird weeks: promotion weeks, stockout weeks, and weeks where sales were zero because there was nothing to sell. A stockout week's sales are fake demand — not nobody buying, but nothing to buy. Feed that into a model and it learns the wrong lesson.
Third, model only the top 20% of SKUs. In our warehouse, 40 of 200 SKUs drive about 80% of sales. For the long tail, experience plus a safety buffer is plenty. Do not build models for products that barely move.
Here is the counterintuitive part: small warehouses usually forecast badly not because the model is too simple, but because the data is too dirty. Stockout weeks, one-off bulk orders, return reversals — leave those in, and even an expensive system produces garbage. Just flagging the abnormal weeks improved our forecasts more than any formula did.
The Excel model that runs our AI demand forecast: moving average times a seasonal factor
Our model is almost too simple to brag about, but it works, and every formula lives in Excel:
Step one, take a 4-week moving average: add the last four weeks of sales for a SKU and divide by four. That smooths out single-week noise.
Step two, multiply by a seasonal factor: last year's sales in the same week divided by last year's average weekly sales. Our cat scratcher runs about 1.4 in summer and 0.7 in winter, so the base demand times the factor gives next week's forecast.
Step three, set safety stock: safety stock = daily average sales × replenishment lead time in days × 1.5. We arrived at 1.5 by trial; raise it if your supplier's lead time is shaky, lower it if it is steady.
Finally, suggested replenishment = forecasted demand + safety stock − stock on hand − stock in transit. One formula dragged down the sheet, and every Monday morning the whole warehouse's reorder list is ready in half an hour.
| Approach | Cost | Learning curve | What we actually saw |
|---|---|---|---|
| Gut feel | Zero | Zero | Stockouts in peak season, overstock in slow season |
| Excel moving average plus seasonality | Zero | Half a day | Stockouts roughly halved, overstock down about a third |
| Buying a forecasting system | Tens of thousands a year | Data cleansing and training | Marginal gain at our volume |
Those results are our own experience, not lab numbers — but for a warehouse our size, they are enough.
The neighbor's lesson: a model nobody reviews dies fast
A model alone is not enough. A friend running a nearby home-goods warehouse copied our sheet in September 2025 and abandoned it within a month. I went to look: the problem was not the model, it was ownership. He built the sheet, then handed weekly updates to a new hire who did not know stockout weeks had to be flagged. Zero-sales weeks went straight into the average, the model drifted further off, and everyone stopped trusting it.
Our fix is equally low-tech: every Monday morning I spend 20 minutes on the sheet myself, checking only three groups — SKUs whose forecast jumped more than 30% from last week, SKUs below safety stock, and SKUs where forecast missed actuals by more than 40% two weeks running. Twenty minutes covers it. A model is dead code; a weekly human review keeps it alive.
One trade-off to be honest about: this simple model handles steady demand and seasonal swings well, but it cannot predict a viral spike. When a short video blows up, no model sees it a week early. Our answer is a "quick-response budget": each month we hold back roughly one week's worth of sales value in flexible cash, so when something starts trending we can place a small fast reorder immediately instead of waiting for the model to catch up. The money sits idle most of the time and saves us when it matters.
Advice for warehouses like ours: get forecasting running first, upgrade later
The right order for a small warehouse is this: get sales history into a clean table, run the moving-average-plus-seasonality model, review it yourself once a week, and judge it after three months. After three months, if your SKU count has passed 500 or volatility has outgrown the simple model, then go talk to vendors — and you will walk in with clean data and a tested logic, which makes for a much stronger negotiation.
The enterprise version is not bad; it is just wrong for our stage. Small warehouses have their own way of surviving: start running first, upgrade when the numbers say so, and spend money where it actually hurts.
By Xiao Ma, Small Warehouse Owner
