Seedlabs

LaunchCast: Few-Shot Demand Forecasts for New SKUs

A tool that predicts sales for brand-new products using a small amount of early data instead of guesses.

A demand-planning tool that forecasts sales for newly launched products from just a few weeks of early data — or none at all at launch decision time — using a frozen language model reprogrammed for forecasting, so initial orders stop relying on guesswork by analogy.

Computer ScienceTime Series Analysis and Forecasting
Cross-Corpus Speech Emotion RecognitionRetail/CPG new-product launch planning: initial order sizing and early replenishment for SKUs in their first weeks of life.

Concept

A demand-planning tool for the first two quarters of a product's life — the window where every statistical forecaster fails for lack of history. The planner supplies the first 4–8 weeks of sell-through for a new SKU (or nothing at all at launch decision time) plus a short structured description: category, price tier, launch promotion, channels, and the closest predecessor item. The engine is a frozen LLM backbone reprogrammed for time series: the numeric signal is patched and aligned to text prototypes, the description is injected as a prompt prefix, and only a lightweight input transformation and output projection are trained — on the retailer's own archive of past launches. Output is a weekly demand forecast with expected ranges for the next 13 weeks, refreshed as early sales arrive. The single wedge: usable forecasts for items with little or no history, replacing the "guess by analogy" spreadsheet that launch teams run today. Delivered as a benchmarked pilot — accuracy on last year's launches is measured against the customer's current analog-based process before it touches a live purchase order.

Why now

Paper [0] shows that a frozen LLM backbone (Llama, GPT-2) reprogrammed for time series outperforms state-of-the-art specialized forecasting models, and specifically demonstrates effectiveness in few-shot and zero-shot settings — exactly the data regime of a product launch. Because only the lightweight input transformation and output projection layers need training, adapting to a customer's launch portfolio is cheap and involves no large-model fine-tuning. The paper is prototype-stage and notes performance depends on the quality of text prototypes and prompt prefixes, so the product is sized as a pilot with planner-reviewed context prompts rather than an autonomous system. The forecasting survey [2] corroborates the underlying pain: specialized time-series models remain confined to single datasets/tasks and struggle with data scarcity and overfitting — which is why the new-item cold start has stayed unsolved by conventional tooling.

Written by GLM · 8 Oct 2026

AI assessment

Backed by 2 papers87

A high-utility application of Time-LLM that targets the specific, high-pain 'cold-start' problem in retail demand planning with a clear, benchmarkable pilot strategy.

Evidence strength
4/5
The idea directly maps to the 'few-shot and zero-shot' capabilities demonstrated in the Time-LLM paper, and the survey corroborates the systemic failure of traditional models in data-scarce environments.
Market pull
5/5
New SKU forecasting is a critical pain point for CPG and retail giants where 'guess by analogy' leads to massive overstock or stockouts, creating high urgency and budget for a solution.
Novelty & moat
4/5
While other catalog ideas handle long-horizon or seasonal forecasting, this specifically addresses the 'cold-start' problem using LLM reprogramming, filling a distinct gap in the existing portfolio.
Feasibility
3/5
While the backbone is frozen, the need to train custom input/output layers on a customer's specific historical launch archive introduces a data-acquisition hurdle for the MVP.
Wedge clarity
5/5
The wedge is exceptionally sharp: replacing the 'analogy spreadsheet' for the first 13 weeks of a new SKU's life, with a clear 'backtest' validation method.
Simplicity / focus
5/5
The product is focused on a single, well-defined task (new SKU forecasting) rather than attempting to build a general-purpose planning platform.

Scored by Gemma 4 · 8 Oct 2026. Rated against a fixed rubric (evidence, market, novelty, feasibility, wedge, simplicity), relative to the closest ideas already in the catalog. A prior estimate to compare ideas before real-world signal arrives.

Persona discussion

AI personas trained on real people's expertise debate this idea as it evolves.

View the discussion →

Act on this idea

Ideas only matter if someone runs with them. Your message goes straight to the founder's inbox — nothing is stored on our servers.

Who benefits

  • Inditexcompany

    Zara's fast-fashion model creates thousands of new SKUs each season with no sales history; a few-shot forecaster that works from the first weeks of sell-through directly attacks their core initial-order problem.

  • Unilevercompany

    Launches new consumer products across many markets; early-life forecasts drive launch quantities, and current statistical tools need roughly a year of history before they become reliable.

  • Targetcompany

    Private-label and new vendor items require initial purchase orders before any sales exist; zero-shot launch forecasts reduce both stockouts and markdowns on new items.

  • Kinaxiscompany

    Its RapidResponse demand-planning suite could embed a few-shot forecaster for the new-item lifecycle stage, a known gap where statistical models fail for lack of history.

Research it builds on

  1. TimeLLM: Time Series Forecasting by Reprogramming Large Language Models
    Ming Jin, Shiyu Wang, Yuan-Fang Li et al. · 2026 · 198 citations
    All ideas from this paper →
  2. A survey of transformer networks for time series forecasting
    Jingyuan Zhao, Fulin Chu, Lili Xie et al. · 2025 · 45 citations
    All ideas from this paper →
Leave feedback
feasibility