A small store can forecast inventory with AI without a data team, as long as you use AI for the right job. Let it clean exports, write formulas and explain results. Do the forecast itself with a simple method you can check: same month last year times your growth rate, or exponential smoothing once you have two years of history. Then backtest it on months you already know, size a buffer from the error, and subtract the stock you already hold.
This guide covers what Shopify, Excel, Google Sheets and two forecasting apps actually offer, a step-by-step workflow, and a worked example for one product heading into the holiday peak. Product pages and help articles were checked on November 16, 2026.
What AI is good at here, and what it is not
Forecasting a few hundred SKUs is mostly arithmetic on monthly unit sales. The hard parts are getting clean data, picking a method that suits how much history you have, and knowing how wrong the forecast is likely to be. A general AI assistant helps with the first and last of those. It is less reliable at the arithmetic itself.
- Use AI for: reshaping an orders export into one row per SKU per month, writing the spreadsheet formula or a short script, spotting stockout months and one-off spikes you should exclude, and explaining a forecast in plain words.
- Do not use AI for: totalling thousands of rows inside a chat, or producing a forecast number with no method attached. If you cannot see how a number was made, you cannot test it.
Our guide to analyzing a Shopify sales export with ChatGPT or Claude covers how to get an assistant to write and run the calculation, and how to handle customer data safely.
What your tools already do
Before you pay for anything, check what you have. Some of it forecasts. Some of it only reports, which is still the starting point.
Shopify inventory reports (Analytics > Reports, category Inventory) Inventory sold daily by product, sell-through rate, ABC analysis, month-end snapshots and days of inventory remaining. These describe the past, they do not forecast. Days remaining uses the last 28 days of sales. Inventory history only goes back to October 1, 2023. Since October 1, 2026 the reports measure On hand instead of Available.
Stocky (Shopify POS app) Shopify’s older inventory and purchase order app. Shopify’s Stocky FAQ says the app is no longer available after August 31, 2026, and points merchants to Shopify’s own inventory features.
Excel FORECAST.ETS Exponential smoothing (the AAA version of ETS) with automatic seasonality detection. Needs a timeline with a constant step, such as the first of each month. Allows up to 30% missing points. Not available in Excel for the web, iOS or Android.
Google Sheets FORECAST A straight-line forecast based on linear regression. Google’s function page lists no seasonal or exponential smoothing option, so it will miss a holiday peak unless you model the season yourself.
Prediko (Shopify app) Lists AI demand forecasting, purchase orders, seasonality planning and replenishment alerts. Pricing is based on the last 12 months of sales: $49 a month under $100K, $119 for $100K to $500K. 14-day free trial.
Inventory Planner by Sage Essentials plan lists monthly forecasts by variant and reorder recommendations. The full product lists forecasting and purchase orders. Essentials is listed at $119.99 a month with a 14-day trial. The full product’s app listing gives no price and asks you to request a quote.
Features and prices as shown on Shopify Help, Microsoft Support, Google Docs Editors Help, the Shopify App Store and Prediko’s pricing page, checked November 16, 2026. Prices are in USD and change, so confirm before you buy. We have not tested these tools.Note what the app listings do not say. Neither Prediko’s pricing page nor the Inventory Planner listings explain how their forecasts are calculated. That does not make them bad. It means the only way to judge one is the same backtest you would run on your own spreadsheet, using your own data, during the free trial.
Pick a method that matches your history
The most common mistake is using a method your data cannot support. A model that looks for seasonality needs to see the season at least twice. A product with eight months of sales has not seen a December yet.
A simple rule of thumb. Whatever you pick, the backtest decides whether you keep it.
- Under 12 months: use a recent average and borrow the seasonal shape from a similar product you have sold longer. Expect large errors and keep orders small and frequent.
- 12 to 23 months: take the same month last year and multiply it by your growth rate. Growth is this year’s units divided by last year’s units over the same months.
- 24 months or more: exponential smoothing with a trend and a seasonal component (Holt-Winters). Excel’s FORECAST.ETS uses the AAA version of exponential smoothing, which belongs to this same additive family. Keep the simple method as a cross-check.
When we ran the worked example below in Python, the Holt-Winters model refused to fit on 22 months of data with the error that it needed at least two full seasonal cycles. That is a useful guardrail. A tool that returns a seasonal forecast from less history is guessing at the season.
The workflow, step by step
1. Export units sold by SKU and month
Export orders or sales by product variant and get it into one table: SKU, month, units sold. Ask your AI assistant to write the formula or script that does this reshaping, then check one SKU’s total against your admin. Use units, not revenue. Price changes and discounts distort revenue as a demand signal.
2. Clean the history
Flag any month where the SKU was out of stock for part of the month. Low sales in that month reflect missing stock, not low demand. Also flag one-off spikes such as a wholesale order or an influencer post you will not repeat. You can ask the AI to list candidate months for you, but you decide which ones to adjust.
3. Backtest before you trust anything
Hide the most recent 6 to 12 months, forecast them using only the older data, and compare against what actually sold. Measure two things:
- Error: total of the absolute misses divided by total units sold. This is often called WAPE.
- Bias: total of the signed misses divided by total units. A method that is always low will cause stockouts even if its error looks modest.
4. Forecast the next order cycle
Run the winning method forward over the period your next order has to cover: from today until the order after this one arrives.
5. Add a buffer, subtract what you have
Size a buffer from your backtest error, then subtract stock on hand and on order. If you use Shopify’s reports, remember that since October 1, 2026 they show On hand, which includes units already committed to open orders. Subtract committed units before you calculate what to buy. For a formula-based approach to buffers, see our guide to reorder points and safety stock.
6. Check the timing, not only the quantity
An order that covers January is no help if you run out in December while it is still on the water. Check that the stock you hold lasts until the new stock lands.
Worked example: one SKU heading into the holiday peak
This example is hypothetical. The product, its sales history, stock level and lead time are our assumptions, chosen to show the method. They are not benchmarks. All calculations were done in Python and checked a second time with exact fractions.
A store sells a gift item with a strong November and December. Here are its monthly units:
2024 98 83 107 115 98 98 103 106 114 118 185 209
2025 111 112 121 117 134 116 125 124 129 138 217 225
2026 121 114 138 144 152 145 152 136 143 154
Hypothetical monthly units sold for one SKU, January 2024 to October 2026.The backtest
We pretended it was October 31, 2025, forecast the next 12 months with four methods using only the data up to that point, and compared them with what happened.
Last month repeated 276 442 14.7% 10.0% low
3-month average 261 442 17.8% 15.0% low
Same month last year 394 442 12.0% 12.0% low
Same month last year x growth (1.18) 465 442 5.9% 3.9% high
Holt-Winters Would not fit: only 22 months of history
Hypothetical backtest by E-commerce Partners. Growth of 1.18 is January to October 2025 units divided by January to October 2024 units.The two methods that ignore the season would have left the store 166 to 181 units short over the peak. Same month last year was closer but still low, because the business had grown. Adding the growth factor brought the error under 6% and turned the bias slightly positive, which is the safer side to miss on for a peak.
The live forecast
Now it is October 31, 2026, with 34 months of history. Growth for January to October 2026 against the same months of 2025 is 1,399 units over 1,227, or 1.14.
- Same month last year x growth: 247 in November, 257 in December, 138 in January. Total 642.
- Holt-Winters (now possible, with nearly three years of data): 228, 244 and 147. Total 619.
The two methods land within 23 units, or about 4%, of each other. When two independent methods agree that closely, you can plan with more confidence. When they disagree by 20% or more, find out why before you order.
The order
Assume 420 units available (not committed), nothing on order, and a supplier lead time of six weeks. An order placed October 31 lands around December 12 and needs to carry the store through January.
Forecast demand, Nov to Jan 642 Backtest winner: same month last year x 1.14
Buffer 76 11.8% of forecast: the 5.9% backtest error, doubled for peak season (our judgment call)
Less stock available now 420 On hand minus units committed to open orders
Less stock on order 0
Order quantity 298 Round to your supplier’s case pack, so 300 if cases hold 25 or 50
Hypothetical reorder calculation. The doubling of the buffer for peak season is an assumption, not a rule.Then check timing. Forecast sales from November 1 to December 12 are about 347 units, which leaves about 73 units when the order lands. That is tight but workable. If the lead time slipped by two weeks, the store would run out around December 21, five days before the order landed.
Why the days-remaining number would have misled you
Shopify’s days of inventory remaining divides stock by average daily sales over the last 28 days. In this example October sold about 5 units a day, so 420 units looks like roughly 85 days of cover, lasting into late January. The seasonal forecast says the same 420 units run out around December 21 without a new order. The report is not wrong. It is measuring the past, and the past 28 days were not a peak month. For seasonal products, treat days remaining as a quick screen and use a forecast for the buying decision.
A prompt to get the AI to do the right part
Paste a sample of your export, with customer details removed, and ask for a method rather than an answer:
I sell physical products. Attached is an export of order lines.
1. Write a formula (or a Python script) that turns it into a
table of SKU | month | units sold, excluding refunds.
2. For each SKU, count how many months of history it has.
3. List months where units fell more than 40% below the
average of the months either side, so I can check
whether the product was out of stock.
4. For SKUs with 12 to 23 months, write the formula for
"same month last year x growth", where growth is units
in the latest 10 months divided by the same 10 months
a year earlier.
5. Write a backtest: hide the latest 12 months, forecast
them, and report WAPE and bias per SKU.
Show your formulas. Do not estimate or round any totals.
Run the formulas yourself in your spreadsheet or have the assistant run the script, then check one SKU by hand. If the hand check and the formula disagree, fix the formula before you look at the rest.
Common mistakes
- Forecasting revenue instead of units. A price rise looks like demand growth and you overbuy.
- Treating stockout months as low demand. The forecast learns the shortage and repeats it.
- Using a straight-line forecast on a seasonal product. Google Sheets’ FORECAST and a 28-day average both flatten the peak.
- Buying an app without a backtest. Run your own hidden-months test during the trial and compare it with the app’s numbers.
- Ignoring lead time. A correct quantity that arrives late still means empty shelves in December.
- Forecasting every SKU the same way. Spend your time on your A items. Shopify’s ABC analysis grades variants by their share of the last 28 days of revenue, which gives you a starting list.
What you should check next
- Run Shopify’s ABC analysis report and pick the 10 to 20 variants that drive most of your revenue.
- Export units by SKU and month for those variants and count how many months of history each has.
- Flag stockout months and one-off spikes.
- Backtest at least two methods on the last 12 months and keep the one with the lowest error and acceptable bias.
- Forecast your next order cycle, add a buffer from the backtest error, and subtract available and incoming stock.
- Check that current stock lasts until the order arrives. If it does not, expedite or order sooner.
Frequently asked questions
Do I need an AI forecasting app at all?
Not to start. For a catalog of a few dozen core SKUs, a spreadsheet with the methods above and a monthly backtest goes a long way. Apps start to earn their fee when you have many SKUs, several locations, purchase orders to manage and too little time to update the sheet. Our guide on how to choose inventory software covers what else to compare.
What should I use now that Stocky is gone?
Shopify’s FAQ points merchants to its own inventory management features. Those include the inventory reports above, but the reports describe past sales rather than forecasting. Pair them with a spreadsheet forecast or a forecasting app.
How do I forecast a brand new product?
You cannot backtest it, so borrow. Use the launch curve and seasonal shape of your most similar existing product, scale it to your expected launch volume, and order a small first batch. Our guide to testing demand before buying stock covers cheap ways to set that launch volume.
How often should I update the forecast?
At least monthly, and before every purchase order. Rerun the backtest every quarter so you notice when a method stops working, for example after a price change or a new sales channel.
Can an AI agent place purchase orders for me?
Some apps list automated purchase order features. Keep a person approving the order, at least until you have several months of backtests showing the forecast is reliable. Our guide to AI agents for store operations covers which tasks hand off well.
The takeaway
You do not need a data team to forecast inventory. You need units by month, a method that matches how much history you have, a backtest that tells you how wrong it is, and a timing check. Let AI do the reshaping and the formula writing. Keep the decision, and the arithmetic, where you can see it.
Sources
- Shopify Help Center, “Inventory reports”. help.shopify.com
- Shopify Help Center, “Stocky frequently asked questions”. help.shopify.com
- Microsoft Support, “FORECAST.ETS function”. support.microsoft.com
- Google Docs Editors Help, “FORECAST”. support.google.com
- Prediko, “Pricing”. prediko.io
- Shopify App Store, “Prediko Inventory Management”. apps.shopify.com
- Shopify App Store, “Inventory Planner by Sage”. apps.shopify.com
- Shopify App Store, “Inventory Planner Essentials”. apps.shopify.com
Last reviewed: November 16, 2026. Shopify report definitions, the Stocky end date, spreadsheet function behavior and app prices match the pages above as checked on that date.
How this article was made: researched and drafted with AI assistance, with every factual claim checked against the sources above. The product, its sales history, stock level, lead time, buffer rule and all forecast results are a hypothetical model, and we have not tested the apps named here.
Related guides
- Reorder points and safety stock: how to calculate both
- How to read Shopify Analytics: the reports that matter weekly
- How to choose inventory software
- Can you sell on Amazon and Shopify at the same time?
