Inventory Forecasting Methods for Shopify Merchants

Ask five merchants how they forecast inventory and you'll get five different answers, most of them half right for whichever SKU the merchant happened to be thinking of. A candle that sells five units a day, steady as a metronome, and a niche accessory that sells four units on a good month and zero the rest of the time are not the same forecasting problem. The method that works for one will actively mislead you on the other.

This post surveys the methods worth knowing, organized around what shape a SKU's demand actually takes. Get that right first, and picking a method stops being a guess.

Quantitative vs. qualitative

Quantitative forecasting builds a number from your own sales history: a formula applied to real data. Qualitative, or judgmental, forecasting substitutes experience and comparison when there isn't enough history to compute from, covered in its own section below.

Before picking between quantitative methods, it's worth knowing the classification that actually explains why one method works well on one SKU and badly on another. A demand-forecasting classification credited to Syntetos, Boylan and colleagues splits demand into four patterns, based on two things: how regularly a SKU sells at all, and how much the quantity swings when it does.

Smooth — sells often, in roughly the same quantity each time.

Erratic — sells often, but the quantity swings wildly.

Intermittent — sells rarely, but roughly the same quantity when it does.

Lumpy — sells rarely, and the quantity is unpredictable too.

The paper draws the line between "regular" and "irregular" selling at an average gap of 1.32 periods between sales, a figure taken directly from the paper's own published cutoff. A companion cutoff of 0.49, for how much quantity is allowed to vary before demand counts as "high variation," is repeated everywhere this classification shows up, but it couldn't be traced back to the primary paper's own text. Treat it as a convention, not a fixed line.

One more wrinkle: which bucket a SKU falls into isn't fixed. A SKU that looks intermittent counted day by day can look smooth counted week by week, simply because a wider bucket has fewer zero-sales periods to show. Before deciding a SKU is "just unpredictable," check whether you're looking at it at a sensible level of aggregation.

The rest of this post is organized around that classification. First, though, get your own sales data into shape. Every method below assumes it's already there.

Moving average

A simple moving average takes your last N periods of sales and averages them; the next forecast is just that average, and the window slides forward one period at a time. It's the method behind the reorder-point formula's "average daily sales" figure, and it's genuinely the right tool for smooth demand: regular, low-variation selling with no real trend.

Its documented limit is trend: it's only a useful estimate when the underlying data has none, and explicitly not a good estimator once a trend appears. Every period in the window carries equal weight, including the stale ones, so it always lags a genuine shift up or down. A weighted variant, giving recent periods more influence, softens this without solving it. For the full formula and a worked example, see our moving average deep dive.

Exponential smoothing

Exponential smoothing solves the "stale data still counts fully" problem differently: instead of a hard cutoff at N periods, every past observation gets a weight that decays the further back it sits.

Next smoothed value = α × (last actual value) + (1 − α) × (last smoothed value)

Alpha, somewhere above 0 and up to 1, controls how much weight recent data gets: higher reacts faster, lower smooths harder and holds onto history longer. There's no single correct alpha. The standard guidance is to pick whichever value minimizes your forecast error on your own data, typically by testing between 0.1 and 0.9, not by copying a number some blog post calls "industry standard." Like a plain moving average, single smoothing doesn't excel once a trend shows up, which is why extended versions exist: double smoothing adds a trend term, and triple smoothing ("Holt-Winters") adds a seasonal term on top of that.

Intermittent demand, sells rarely, consistent quantity when it does, breaks both methods above, because most periods have nothing to smooth. Croston's method was built for this: it keeps two smoothed series, one for sale size and one for the gap between sales, updating only when something actually sold.

Forecast rate = (smoothed demand size) ÷ (smoothed interval between sales)

It has a documented weakness of its own: research on the method finds a consistent positive bias, meaning it tends to over-forecast. A commonly proposed correction can just as easily overcorrect the other way, so neither version is neutral by default. Worth knowing if a few of your slow movers would use it, without expecting a specific number attached to the bias, because no traceable source puts one on it.

Reorder-point-based forecasting

The reorder point formula isn't usually filed under "forecasting," but it functions as one: average daily sales, projected across lead time, plus a buffer. The same Cedar & Fig, 250g example from that guide (5 units a day, a 12-day lead time, 30 units of safety stock) reaches a reorder point of 90 units, and that 90 is really a short-horizon demand forecast wearing a trigger-point's name.

It fits smooth-to-moderately-erratic demand well, for the same reason a moving average does: it assumes recent velocity is a fair stand-in for near-future velocity. It stops making sense once a SKU is intermittent or lumpy. "Average daily sales" barely means anything for a product that sells twice a month in wildly different quantities, and a reorder point built on that average triggers close to the wrong moment either way.

Judgmental forecasting

Judgmental methods turn experience and comparison into a number when there isn't enough sales history for a formula to work from: a genuinely lumpy SKU, or one with no history at all. Worth knowing: analog forecasting (finding an existing, similar product and using its early pattern as a template), expert opinion, and the Delphi method, a structured process of anonymous, multi-round estimation designed to reduce the groupthink a single opinion or an open discussion tends to produce.

None of these are neutral. Judgmental estimates carry the estimator's own bias by construction, anchoring on an initial number or an agenda unrelated to the product. Expert opinion is weakest exactly where it matters most: something genuinely new, with no real precedent to reason from. The most consistent finding in the research is that combining more than one method (an analog estimate, a judgmental estimate, and the first real weeks of sales once they arrive) outperforms leaning on any single one.

Lumpy demand deserves a mention here, since it's the pattern none of the quantitative methods above handle well: irregular timing and unpredictable size, together, among the hardest forecasting problems there is, per the research, with no single-formula fix. For a brand-new product specifically, with no sales history yet, we cover that case in full: analog comparison, early signals, and sizing a first order under real uncertainty.

Choosing by catalog size

Under roughly 30 SKUs, almost any of the above works by hand. Pick a moving average or the reorder-point formula for your smooth sellers, note which SKUs are intermittent or lumpy and treat those with more judgment and a wider buffer, and recalculate monthly.

Past that, the real constraint isn't which formula is "best." It's that a real catalog is never one demand pattern. A typical store has a handful of smooth bestsellers, a wider band of erratic mid-tier products, and a long tail of intermittent and lumpy SKUs that barely sell often enough to average anything from. Applying one method to all of them isn't a simplification, it's a mismatch for most of the catalog.

Reclassifying which pattern every SKU belongs to, and re-running the right formula for each one, is manageable at a few dozen SKUs and genuinely unsustainable past a few hundred done by hand every month. StockCue doesn't ask you to pick one method and apply it to everything. It blends a recent and a baseline sales window per SKU, adds trend sensitivity and seasonal indices on top, and excludes outliers automatically, on every plan including Free. For the complete process this fits into, see our complete guide to Shopify inventory forecasting.

Frequently Asked Questions

What is the simplest inventory forecasting method for a small Shopify store?

For a small catalog of steady sellers, a simple moving average or the existing reorder-point formula is usually enough — both just need an average of recent sales, a lead time, and a buffer. They're the right starting point for smooth demand specifically; a SKU that sells rarely or in wildly different quantities needs a different approach.

What's the difference between moving average and exponential smoothing?

A moving average gives every period inside a fixed window equal weight and drops anything older entirely. Exponential smoothing keeps all of history but lets its influence decay the further back it goes, controlled by a smoothing constant, alpha. Both share the same core limitation: neither handles a trend well on its own, which is why trend- and season-adjusted variants of exponential smoothing exist.

When should a Shopify store use qualitative forecasting instead of a formula?

When there isn't enough sales history for a formula to compute from — a brand-new product, or an existing SKU whose demand is genuinely lumpy (irregular timing and unpredictable quantity together). The strongest approach in that situation is combining more than one judgmental method, or a judgmental estimate with early real sales data, rather than trusting a single opinion.

Does the best forecasting method change as a store grows?

The right method for a given SKU is really about its demand pattern, not the store's size directly — but a bigger catalog has more SKUs sitting in every pattern at once, and manually classifying and recalculating each one becomes the real constraint. That's less a "better formula" problem and more a "who has time to redo this every month" problem.