Demand Forecasting AI in India

Indian retailers, D2C brands, and FMCG teams plan inventory around sharp demand swings — the festive season around Diwali, wedding-season spikes, or a monsoon-driven dip in certain categories. Statistical and ML forecasting models remain the right tool for the core numbers; what large language models add is a way to turn messy, unstructured signals — news, promotion calendars, store-manager notes — into structured inputs, and to explain a forecast or anomaly in plain language. unoblox bills every model used in that layer per million tokens, in rupees.

What LLMs are actually good for here

An LLM does not replace a proper time-series or ML forecasting model, and no credible vendor should claim otherwise. What it does well is read unstructured text — a promotions calendar in a PDF, a batch of store-manager notes, a news digest — and turn it into structured fields your existing forecasting pipeline can consume, or take a forecast's output and draft a plain-language explanation for a category manager.

A typical pipeline

1. Collect unstructured signals: promo calendars, regional news, store notes, weather advisories
2. Use an LLM to extract structured fields (event name, date range, affected SKUs or categories)
3. Feed those fields into your existing statistical or ML forecasting model
4. Use an LLM again to draft a short, plain-language summary of the forecast or a flagged anomaly for a human reviewer

Model options for each step

Step	Suggested model	₹ input / ₹ output (per 1M tokens)
Bulk extraction from notes/calendars	deepseek-ai/deepseek-v4.1-flash	₹20.16 / ₹60.48
Structured extraction needing more reasoning	qwen/qwen3-235b-a22b-instruct-2507	₹9.07 / ₹55.44
Complex, longer-report analysis	openai/gpt-5	₹126 / ₹1008
Prototyping the pipeline	qwen/qwen3-1.7b	FREE (₹0)

Sample extraction call

curl https://api.unoblox.ai/v1/chat/completions \
  -H "Authorization: Bearer ub-gw-xxxxxxxxxxxxxxxx" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-ai/deepseek-v4.1-flash",
    "messages": [{"role": "user", "content": "Extract the promotion name, date range, and category from this note as JSON."}]
  }'

Budgeting for scale

Because pricing is per million tokens and quoted in rupees, a forecasting pipeline that processes thousands of short notes a day is straightforward to budget upfront — multiply your expected monthly token volume by the per-model rate above, and check /models for any model not listed here before you plan around a number.

Frequently asked questions

Can an LLM forecast demand directly from historical sales numbers? It's not the right tool for that — a statistical or ML time-series model will outperform an LLM on pure numeric forecasting. Use the LLM for the unstructured-text layer around that model.

What's the actual benefit over just reading the notes manually? Consistency and speed at volume — an LLM can structure hundreds of notes the same way, quickly, at a predictable per-token cost.

Which model should I start with? qwen/qwen3-1.7b is free and good enough to prototype your extraction prompts before you move to a paid model for production volume.

Is pricing genuinely in rupees? Yes — every model is priced and billed in rupees per million tokens, with a monthly GST invoice.

Do foreign models process this data inside India? No — models like GPT-5 and Qwen3 run on their providers' infrastructure outside India; the India benefit is rupee billing, GST invoicing, and one endpoint, not data residency.

Get started in rupees → https://unoblox.ai/sign-in