Ecommerce Inventory Forecasting for Shopify, Amazon and DTC Brands

This page sits under demand forecasting for apparel brands and is for an ops or planning lead running a Shopify store, an Amazon channel, or both. It covers what makes DTC and marketplace forecasting different from wholesale planning, the arithmetic of censored demand, what the platforms constrain, and where the exercise collides with a factory calendar.

What makes ecommerce inventory forecasting different from retail planning

Four differences, each of which changes the maths rather than just the tooling.

Granularity is daily or weekly, not monthly. A wholesale plan works in months because orders arrive in months. A DTC channel produces real weekly and daily signal — and equally real weekly and daily noise. Forecast weekly at minimum; forecast daily only if you have genuine day-of-week and promotional structure to model.

A share of demand is a decision you made, not a pattern you observed. A paid-media push is not seasonality. If you spent $40,000 on Meta in week 5 and $8,000 in week 6, the demand gap between those weeks is an output of your marketing budget. A model without the spend calendar reads it as random variation; a model with it separates baseline from promoted demand. This is the highest-return data addition for most DTC brands, and it costs nothing but a join.

Launches and drops have no history at all. New colourways, collaborations and limited drops need attribute-based analogues, which is the one place machine learning genuinely beats a planner: see what AI demand forecasting genuinely adds.

Returns are a real inventory line, not a revenue reversal. The National Retail Federation estimated that 19.3% of US online sales would be returned in 2025, on total returns of $849.9 billion, with 9% of all returns fraudulent (published 15 October 2025). Apparel sits above that all-category average because size and fit dominate return reasons — treat the apparel-specific rate as an industry range of roughly 20-30% rather than a sourced figure, and measure your own. Units that pass inspection re-enter sellable stock, so they are inbound supply with a lag of typically three to six weeks. If 25% of units are returned and 80% of those are resellable, 100 gross orders consume only 80 units of net inventory, so a buy planned on gross orders is 25% larger than the stock the channel consumes — but because returns arrive late they do not help your first cover period.

The censored-demand problem: a sold-out SKU records zero demand

This is the error that costs DTC brands the most money and appears in the fewest tools. Your sales table records units sold, not demand. When a variant is out of stock it sells zero, and a model trained on that history learns your best seller is a mediocre seller — so you buy less of it, which causes the next stockout, which reinforces the lesson. The FreshRetailNet-50K study of 50,000 store-product series calls this a self-reinforcing cycle and quantifies it: training on raw censored sales produced a −7.37% weighted percentage error, while recovering latent demand cut bias to about 2.58% and improved WAPE from 31.75% to 29.02%.

Worked example. One core cotton crew tee, black, size M, eight weeks of Shopify sales. The variant sold out on the Thursday of week 4 and was not restocked until the Monday of week 7.

• Raw weekly mean = 1,193 ÷ 8 weeks = 149.1 units/week.
• Availability-corrected rate = 1,193 ÷ 39 in-stock days = 30.59 units/day = 214.1 units/week.
• The raw figure understates true demand by (214.1 − 149.1) ÷ 214.1 = 30.4%.

Now push it into the buy. At a 20-week cover — the low end of a realistic import replenishment cycle — the raw number orders 149.1 × 20 = 2,982 units; the corrected number orders 214.1 × 20 = 4,282 units. The gap is 1,300 units: 351 kg of fabric at 0.27 kg per tee, comfortably more than a 300 kg custom dye-lot minimum and therefore a change to what the mill runs; $7,904 of inventory at $6.08 landed; and $36,296 of full-price gross margin at a $34 ticket and $27.92 contribution, if that demand was real and you failed to hold it.

The correction is arithmetic, not machine learning: divide units by in-stock days, not calendar days, before you fit anything. Ask every tool you evaluate whether it does this, using the checklist in how to evaluate demand planning software.

Shopify inventory forecasting: what the native reports do and do not do

Shopify's native reporting is competent and entirely backward-looking. Per Shopify's own Help Center, the default inventory reports include a month-end inventory snapshot and value, inventory sold daily by product, products by percentage sold, ABC product analysis grading products by revenue contribution over the last 28 days, products by sell-through rate, inventory remaining per product, and inventory adjustment reports.

The table below shows what each channel gives you and what it withholds. As of August 2026.

What Shopify does not include: a demand forecast, safety stock, a reorder point, a supplier lead time, a purchase order, or any adjustment for days a variant was out of stock. The "inventory remaining" report extrapolates your average sales rate — exactly the censored figure the previous section warns about. Shopify inventory forecasting therefore needs an app on the platform or an external planning layer that can read order and inventory data and write purchase orders back.

Amazon inventory forecasting: capacity limits and inbound lead times

Amazon adds a constraint Shopify does not: you cannot send in as much as you want.

FBA capacity limits moved to a monthly cycle from 1 January 2026, measured in cubic feet rather than units, with sellers receiving one confirmed limit for the coming month and two estimated limits beyond it, according to Spreetail's tracking of the change. Allocation is driven by inventory health signals including the Inventory Performance Index, sell-through velocity and historical and forecast sales; trade guidance commonly cites an IPI threshold of 400, and Seller Labs describes low scores as leading to reduced storage space and restricted inbound shipments. Verify your own limit in Seller Central rather than relying on any published figure.

Two consequences. Your Amazon forecast has to be feasible, not just accurate — a plan implying 4,000 units into a limit permitting 1,800 is not a plan. And inbound is not instant: units are not sellable until received and checked in at the fulfilment centre, which adds days to weeks on top of freight.

This is also where the phrase "amazon inventory forecasting agency" comes from. Agencies genuinely help with Amazon-specific mechanics — capacity headroom, IPI recovery, restock sequencing, placement fees. They cannot help with the constraint most importing apparel brands actually face, which is two sections down.

CPG demand forecasting and CPG demand planning: what carries over

CPG demand forecasting predicts consumer offtake for packaged goods and is the most mature discipline in this field, which is why DTC teams borrow its vocabulary. Three practices carry over cleanly.

Separate shipments from consumption. In CPG the manufacturer ships to a retailer, who then sells to a consumer, and shipments include the retailer's inventory build — so forecasting demand from shipments confuses restocking with buying. The DTC equivalent is forecasting from units sent to Amazon rather than units sold on Amazon.

Reconcile several demand signals. CPG planners blend retailer point-of-sale data, syndicated market data from providers such as Circana, and internal shipments; Circana describes current practice as combining retailer POS and loyalty data, manufacturer trade plans, category data and real-time signals. The DTC equivalent set is site sessions and conversion, marketplace POS, wholesale sell-through and paid-media spend.

Model promotions as a separate process. CPG demand planning treats trade-promotion lift as its own model rather than as noise in the baseline — exactly the discipline a DTC brand needs for paid-media pushes and drop launches.

What does not carry over is the tolerance. CPG staples run 10-25% MAPE at SKU level (25-35% during promotions) per Umbrex's compilation; apparel runs 35-60%. Importing CPG accuracy targets into an apparel plan will make you conclude your forecast is broken when it is merely apparel.

Tooling: where Anaplan and the enterprise platforms fit

Anaplan is an enterprise connected-planning platform used across finance, sales and supply chain, including demand planning. Gartner placed Anaplan as a Challenger in both 2026 Magic Quadrant reports for Supply Chain Planning Solutions, published 18 March 2026, and lists it among the vendors reviewed in that market on Peer Insights. Anaplan forecasting is a modelling environment rather than a packaged apparel engine: it will represent nearly any planning logic you can specify, and it needs someone to build and maintain the model. That makes Anaplan demand planning a reasonable fit for a company with a planning function and a poor fit for a five-person DTC team, which is better served by a commerce-platform app. The full category comparison, including pricing models, is in how to evaluate demand planning software.

The real problem: a daily demand signal against a 15-23 week PO-to-DC pipeline

Everything above improves the demand number, and for a DTC apparel brand importing from Asia the demand number was never the binding constraint. A Shopify or Amazon brand reads demand in days and replenishes in months. A first order runs 10-16 weeks PO to FOB, which becomes 15-20 weeks PO to a US DC from India and 18-23 weeks from Bangladesh once ocean transit and clearance are added; repeat orders compress to 8-12 weeks PO to FOB. The replenishment lead time is an order of magnitude longer than the demand signal that triggers it.

Three consequences, all of which show up as cash.

The reorder decision is made on a forecast, not on a read. By the time four weeks of sell-through tell you a colourway is working, the replenishment you place lands 15-20 weeks later PO to DC — most of a season away. Lead-time variability, not demand variability, then dominates the safety-stock calculation: reorder points when your lead time is 90 days.

The order quantity is not continuous. A custom colour has a dye-vessel minimum of roughly 300-500 kg of fabric, about 1,110-1,850 tees at 0.27 kg per tee, so a forecast recommending 640 units of a colour is not orderable at all: how fabric and dye-lot minimums set apparel MOQs.

Marketing moves demand faster than sourcing can answer. A paid push that doubles weekly velocity has a five-month supply response. The only fast levers are air freight on a partial quantity, pre-positioned fabric, or capacity reserved in advance against a forecast.

That is the honest summary for an importing brand: the forecast is necessary and not sufficient, and getting the number right earns an earlier warning rather than a faster supply chain. Turning that number into a stock position is covered in inventory forecasting for apparel brands; the vocabulary is defined in the Yarnstick glossary of apparel sourcing terms.

Frequently asked questions