Stockouts and Dead Stock Are Both Killing Your Shopify Margin: The AI Forecasting Workflow to Fix Both

Shopify demand forecasting advice usually stops at "analyze your sales data and predict future demand," which is not a workflow. Here is the actual five-step process, with exact prompts and thresholds at each stage.

Why Generic Forecasting Advice Does Not Work

Forecasting fails in practice because building a usable forecast requires specific decisions, how far back to look, how to handle seasonality, what counts as a meaningful shift, that generic advice skips entirely.

Step 1: Export the Right Sales History Window

Export weekly Sales by Product data covering the trailing 12 months, or full history if shorter. Weekly granularity matters since monthly data smooths over spikes that determine reorder timing.

Step 2: Establish Your Baseline and Flag Seasonality

Prompt: "Here is my weekly unit sales data for [product name] over the last 12 months: [paste]. Calculate the average weekly sales rate, then identify any weeks where sales were more than 40 percent above or below that average. For each flagged week, tell me the date range and whether it aligns with an obvious calendar event or appears to be an anomaly."

Step 3: Build a Forward Forecast With Explicit Assumptions

Prompt: "Based on the baseline and seasonal pattern you identified, project weekly demand for the next 12 weeks. State your assumptions explicitly: are you assuming the recent trend continues, reverts to the 12-month average, or follows the same seasonal pattern from last year. If there is a seasonal event coming up, project a specific higher estimate and explain the multiplier used."

Step 4: Convert the Forecast Into a Reorder Point

Prompt: "My supplier lead time is [X] weeks from order to inventory being sellable. Using the 12-week demand forecast, calculate the reorder point, my current inventory level at which I need to place a new order today, accounting for a safety buffer of one additional week."

Step 5: Set a Recheck Cadence Based on Product Velocity

Products selling over 50 units weekly should be rechecked every two weeks. Lower-velocity products can run monthly. Build this into a recurring calendar reminder.

Explore our Amazon Fee Calculator if you sell the same product on Amazon too, since your reorder point needs combined channel demand.

A Worked Example From Steps 2 Through 4

A seller running this on a ceramic mug line found their weekly baseline sat around 85 units, with step two flagging a 6-week November window running 60 percent above baseline, aligning with holiday gifting season. Step three, projecting from a September starting point, explicitly stated its assumption: apply the same November multiplier observed the prior year, projecting roughly 135 to 140 units weekly during that window. Step four took a 5-week supplier lead time and calculated a reorder point of 675 units, accounting for the safety buffer. Without the explicit multiplier from step three, a naive forecast using only the trailing average would have suggested reordering roughly 40 percent less inventory than the season actually required.

What This Forecast Cannot Account For

This is built entirely from historical data, with no visibility into a competitor stockout, a viral moment, or a planned marketing push. Treat it as a baseline to adjust with your own context.

Check our newsletter for ongoing demand planning coverage.

Frequently Asked Questions

What if I do not have 12 months of sales history for a new product?

Use whatever history you have, even 8 to 10 weeks establishes a rough baseline, though seasonal flagging will be less reliable.

How do I account for a planned promotion in the forecast?

Add it explicitly to your step three prompt, describing the discount depth and asking for a demand multiplier based on past promotions if available.

Does this workflow work for products I sell on multiple channels?

Core forecasting works per-channel, but the reorder point in step four needs total demand across every channel drawing from the same inventory.

How accurate is a Claude-generated forecast compared to dedicated forecasting software?

Dedicated software often incorporates more sophisticated models, but this workflow is a strong, transparent starting point with explicitly stated assumptions.

Takeaways

• This is a five-step process: export weekly data, establish baseline and seasonality, forecast with explicit assumptions, convert to reorder point, set recheck cadence.
• Weekly, not monthly, granularity is necessary to catch demand spikes that matter for reorder timing.
• Explicit stated assumptions let you catch projections that do not match your own read of the business.
• The reorder point, not the raw forecast, is the actual operational output that matters.
• Recheck high-velocity products every two weeks, lower-velocity monthly, on a scheduled cadence.