New Product Forecasting

Forecast SKUs with no sales history

HorizonAI · Walkthrough · ~3 min
The Problem

SKU cold-start

You're launching a new product. Traditional forecasting models need historical data, but your new SKU has zero or near-zero sales records. So what do you forecast?

What doesn't work

ARIMA, ETS, Prophet, LightGBM — all need history.

"Use last year's number" — there's no last year.

"Guess from gut" — works at small scale, falls apart across hundreds of launches.

What does work

Analog blend — borrow from similar SKUs that DO have history.

Attribute regression — learn how attributes drive volume across the catalog.

Seasonality libraries — apply attribute-matched 52-week shapes.

Three Approaches

Three ways to forecast a new SKU

All three are built into the workbench. Pick by what data you have — and you can combine them.

🔍

Analog Blend history match

Find similar SKUs (by attribute or by AI judgment), then blend their historical forecasts weighted by similarity.

Output: 13-week forecast curve.

🧠

Attribute Launch recommended

Train Ridge + Random Forest on every historical SKU's attributes → first-period volume. Predict launch volume from your new SKU's attributes directly.

Output: first-period units + 80% CI + implied life total.

📅

Seasonal Component library match

Match attributes to a 52-week profile in the seasonality library, multiply by an analog-derived BSR. Returns a fully-shaped weekly forecast.

Output: weekly forecast × seasonal shape.

Step 1

Enable New Product Mode & set attributes

In the SKU Forecaster workbench, scroll to the purple "🆕 New Product Mode" panel and click Enable. Attribute selectors populate from your active dataset's columns.

🆕 NEW PRODUCT MODE 3 attrs

Set product characteristics:

category: Apparel subcategory: Polos color: Navy

More attributes = more accurate analog matches. You can leave some blank — the system widens the search until it finds candidates.

Step 2a · Analog Blend

Find similar SKUs — your choice

A toggle above the "Find Similar SKUs" button lets you pick how analogs get chosen.

🤖 AI-Powered
📐 Rule-Based

Results: 5 analogs found

AP-POLO-001 Perfect match — same category, subcategory, and color 6.0
AP-TEE-001 Same category, similar casual shirt styling 5.4
AP-SWIM-001 Same category and color, both casual wear items 4.8

🤖 AI-Powered

Claude reads the catalog and picks analogs based on semantic similarity. Slower, but catches "T-shirt as an analog for Polo" cases. One-line reason per pick.

📐 Rule-Based

Deterministic exact-attribute scoring: +2 per exact match, +1.5 for "similar" numerics. Repeatable, no LLM round-trip. Each match is explicit in the results.

Step 2b · Attribute Launch

Predict launch volume from attributes

If you have 20+ historical SKUs to train on, the Attribute-Based Launch Model fits a Ridge + Random Forest ensemble on (attributes → first-period sales) and predicts directly for your new SKU.

🧠 ATTRIBUTE-BASED LAUNCH MODEL

Trained on all 42 historical SKUs · Ridge + RF ensemble

115
First period
54 – 177
80% CI
330
Implied life
0.92
R² (fit)

Feature importance: subcategory 50% · price 29% · category 11% · color 10%

Click "Auto-Fill Sell-Through" in the modal to push implied life into the Sell-Through panel and turn this point estimate into a weekly schedule.

Recommended Workflow

Use them together

Most teams use Attribute Launch to set the initial buy quantity, and Analog Blend or Seasonal Component to schedule weekly receipts.

1

Set attributes for the new SKU

Category, subcategory, color, price — whatever the catalog tracks. More attributes = better matches.

2

Run Attribute Launch Model

Get a defensible first-period number + 80% CI. Use this as your initial buy or shelf-set quantity.

3

Run Analog Blend (AI or Rule-Based)

Get a 13-week forecast curve. Compare against the Attribute Launch life total — if they agree within ~30%, you have a defensible plan. If they disagree, that's a signal to investigate.

4

Auto-fill Sell-Through (optional)

Click Auto-Fill Sell-Through in the Attribute Launch modal to distribute the implied life total across weeks using a configurable decay curve.

Honest Framing

What to trust, what to second-guess

Cold-start forecasting is inherently uncertain. Use the right method for the right situation.

Trust more when…

• You have 20+ historical SKUs in similar categories
• Attributes are well-populated (≥3 set)
• Analog Blend & Attribute Launch agree within ~30%
• The Attribute Launch model R² > 0.5

⚠️

Second-guess when…

• Methods disagree by >50% (orchestrator flags these)
• Fewer than 5 analogs found
• Confidence reported as "low" or "very_low"
• The new SKU represents a category your catalog hasn't seen

Try It

Load a dataset → Enable New Product Mode → Set attributes

Open Workbench → Help docs Methodology §8