You Just Launched a New Product. 6 Weeks Later, You Have a Problem.
Scenario A: Overordered
Forecasted 2,500 units. Sold 1,200. Dead stock: 1,300 units.
Cost: $47K markdowns + $8.4K storage fees.
Scenario B: Underordered
Forecasted 800 units. Sold out in 3 weeks. No stock for continued demand.
Cost: $180K+ missed revenue. Can't get it back.
80% of new product launches fail. Most aren't bad products. They're bad forecasts.
This blog shows how to forecast demand for new launches using historical data, proxy products, seasonality, cross-functional consensus, and AI—so you stop losing $47K-$180K per launch.
The 5 Methods That Actually Work
Method 1: Proxy Data from Similar Products
Example: Launching chocolate protein powder. Use vanilla launch as proxy.
→ Vanilla launch: 1,800 units in 8 weeks (35% of sales)
→ Chocolate historically: 22% of sales
→ Adjustment: 1,800 × (22% / 35%) = 1,131 units expected
Method 2: Seasonality Adjustments
Real example: Wellness brand launched Vitamin D in July using January baseline. Winter demand 3.5× higher.
Ordered 1,200 units. Sold 340. Dead stock 860. Loss $12,900.
Method 3: Delphi Method (Cross-Functional Consensus)
Independent forecasts from product, sales, warehouse, marketing teams
Document reasoning for each forecast
Find consensus number reflecting collective expertise
Method 4: Pre-Orders and Beta Testing
Pre-Orders
Launch pre-order 4-6 weeks before full launch. See actual conversion rate to scale forecast.
Beta Testing
Release limited batch (500-1,000 units). Track sell-through velocity to estimate full demand.
Method 5: AI and Machine Learning Forecasting
Amazon's Cold-Start Forecasting Method
Uses item metadata + similar product patterns to forecast new products. Reduces forecast error by 20-50% vs traditional methods.
Tools: Prediko, Pecan, Amazon Forecast implement this method.
Real Case: $2.7M Apparel Brand (Gut Feel → Data-Driven)
Before: Gut Feel
Marketing: "Order 2,500." Finance: "Order 800." Compromised: 1,600 units.
Sold 980. Dead stock 620. Loss $18,600.
After: Data-Driven
Used proxy + seasonality + Delphi + pre-order + AI. Final forecast: 1,134 units.
Sold 1,112. Error 2.0%. Zero markdowns. $180K+ margin protection.
wMAPE Benchmarks
| Category | wMAPE Benchmark |
|---|---|
| Fashion & Apparel D2C | 25-40% (new: 50%+) |
| Food & Beverage | 20-25% |
| Health & Wellness D2C | 18-30% |
| Durable Goods | 40-50% |
The ROI
Implementation Cost (Year 1)
$25K-$50K
Savings ($3M-$5M brand)
$40K-$80K overstock reduction + $30K-$100K stockout prevention + $50K-$150K markdown avoidance
= $120K-$330K/year
Net ROI: 240-660% Year 1
Stop Losing $100K-$300K Per Year on Bad Forecasts
One $2.5M fashion brand reduced forecast error from 38% to 7% in 3 launches. $180K inventory savings. 3.2-point margin improvement. Went from losing $300K per launch to saving $100K per launch.
Free 15-Minute Demand Forecast Audit
We'll pull your last 5 launches, calculate your actual forecast error, and show you exactly where you're bleeding cash.
FAQ
If AI forecasting is so accurate, why are we still missing demand on 40% of launches?
Because 80% of brands aren't using AI forecasting. They're using historical averages, gut feel, and spreadsheets. AI works, but most companies haven't adopted it yet. The ones that have are reducing forecast error by 20–50%.
How much data do I need to use machine learning?
You can start with 8–12 weeks of proxy product data (a similar SKU from your catalog). You don't need years of history. Amazon's method uses item metadata + similar product patterns.
How do I know if my forecast is good enough?
If your wMAPE is 25% or lower, you're in the top quartile. If it's 15% or lower, you're exceptional. If it's 40%+, you're losing money on every launch.
How often should I update my forecast?
Once a week for the first 4 weeks of launch. Then monthly. If you're seeing 20%+ variance from forecast, increase to twice weekly until you understand why.