A common way to forecast on Shopify goes like this: pull the last 90 days of sales, divide by 90, and order against the result. It feels rigorous. It has data in it. And for a lot of multi-location merchants, it quietly produces the same two problems every month — a stockout at the busy location, a growing pile at the slow one.

The math has one flaw. A flat average treats every day and every location as identical, and neither is true. This post covers what a weighted, per-location forecast does differently, plus a short checklist for interrogating whichever forecasting tool you use. Ours included.

Why the flat average fails multi-location stores

If you run one location and demand barely moves, a flat average works fine. Most multi-location merchants have neither condition. Three things break:

• Trends get flattened. A product that went from two sales a day to six averages out to four, so you under-stock the location where it’s accelerating and find out at the worst possible time.
• Locations get blended. Aggregate demand can look healthy while one store starves and another hoards. The average has no opinion on where stock should sit.
• Stockout days read as zero demand. If a product couldn’t sell for ten days, the average learns that nobody wanted it. The forecast drops, you order less, and the cycle repeats.

None of these are exotic edge cases. They’re the default failure modes of averaging, and they’re expensive. IHL Group, which has tracked the cost of inventory distortion for nearly two decades, puts the global bill for out-of-stocks and overstocks at $1.73 trillion a year. Your slice of that number is smaller, obviously. But it’s made of the same ingredients: stock in the wrong place, ordered off a forecast that wasn’t paying attention.

A flat average treats every sales day as equal. Your customers don’t.

What weighted history actually means

Weighted forecasting starts from a plain idea. Recent sales say more about next month than old sales do, so you let last week count for more than last quarter instead of giving every day an equal vote.

A worked example. Say a product sold 90 units over the past 90 days.

• A flat average reads that as one unit a day, so it forecasts 30 units for the next 30 days.
• Now look at the shape of those sales. The first 60 days moved 30 units, about half a unit a day. The last 30 days moved 60, two a day.
• A recency-weighted forecast reads that acceleration and lands closer to 50 or 60 units for the next month.

Same data, same product, and the two methods produce reorder decisions nearly double apart. Only one of them noticed the product is taking off.

The same logic protects you on the way down. When a product cools, a weighted forecast cools with it, instead of telling you to keep ordering at July volumes through October.

Last week tells you more about next week than last quarter does.

Forecast by location, not just by product

For a multi-location store, the useful unit of forecasting is the product at the location. Not the product overall.

A combined forecast can be exactly right while the distribution is exactly wrong. Say you hold 100 units across two stores and expect to sell 100 next month. Looks perfect on paper. But if 80 of those sales will happen downtown while 80 of the units sit at the warehouse store, you’ll spend the month losing sales in one place and dusting shelves in the other. The total never sees the problem.

This distinction also changes the fix. When demand genuinely exceeds your total stock, you buy more. When stock is simply in the wrong place, a transfer solves it for the cost of a van ride. Knowing which situation you’re in is most of the job, and a blended number can’t tell you.

The new-product problem: forecasting with no history

Weighting history fails honestly in one case: there isn’t any. A product launched last Tuesday has no 90-day weight.

Merchants handle this a few ways:

1. Borrow a comparable product’s early sales curve and adjust from there.
2. Treat the first sales as a strong signal, then update the estimate quickly as real data arrives.
3. Hold a small buffer at each location until a pattern shows.

A good tool should do some of that thinking for you. BeeBalanced combines recent and historical sales patterns with new-arrival logic, so a product with two weeks of history still gets a workable per-location estimate instead of a blank.

Pick a forecast period that matches how you operate

A forecast horizon is a planning decision, not a default you accept. If you move stock between locations weekly, a short window keeps the forecast responsive. If you rebalance monthly, a longer window smooths out noise you’d otherwise chase.

Some rules of thumb:

• Shorter periods react faster but run noisier. Better for fast movers and frequent transfers.
• Longer periods are steadier but slower to catch a turn. Better for stable catalogue items.
• Whatever you pick, match it to your transfer or reorder cadence. A 60-day forecast feeding a weekly transfer routine is answering a question nobody asked.

In BeeBalanced, the forecast window is a setting, measured in days, and the estimate ignores out-of-stock days so a stretch at zero doesn’t drag demand down. The rest of the settings panel gets its own walkthrough soon.

How to sanity-check any forecast, including ours

Run these five questions against whatever tool you’re evaluating:

1. Does it weight recent sales, or does every day get an equal vote?
2. Does it forecast per location, or only per product?
3. What does it do with a product that has no sales history?
4. Does it account for days a product couldn’t sell?
5. Can you change the forecast window to match how often you actually move stock?

A tool doesn’t need a perfect answer to all five. But a vendor who can’t answer them at all is selling you a number, not a forecast. Anyway, the questions cost nothing to ask.

A forecast you can’t interrogate is a guess with a dashboard.

Where this leaves you

Weight recent history. Forecast at the location level. Have a plan for products that haven’t sold yet. Get those three right and the flat average stops costing you money, whether you build the spreadsheet yourself or let software carry it.

BeeBalanced runs this kind of forecast across every product at every location, then turns the result into draft transfers you can review, edit, or delete before anything moves.

There’s a 14-day free trial if you want to see your own numbers in it.