Demand Forecasting

The Cost of Guessing Wrong

Every retailer in India knows this pain. You ordered 200 kg of sugar before Diwali because "last year it sold well." But this year, customers bought more dry fruits and less sugar. Now you have 80 kg of sugar sitting on the shelf, tying up capital. Meanwhile, the dry fruits you stocked conservatively sold out by day two, and you turned away customers who went to the shop across the street.

This is the forecasting problem. Overstocking costs you money (spoilage, storage, tied-up capital). Understocking costs you sales and customer trust. The average Indian retailer loses 5-10% of revenue to poor demand estimation — and for perishable goods like dairy, fruits, and vegetables, the number can reach 20%.

AI does not eliminate uncertainty. But it reduces it dramatically by analyzing patterns in data that no human can process manually. This chapter teaches you how demand forecasting works, why it matters, and how to start using it — even if your "data system" is a notebook and a calculator.

Gut Feel vs. Data-Driven Forecasting

Most small and mid-size retailers in India forecast demand using one of three methods:

The difference between 65% accuracy and 85% accuracy might not sound dramatic, but for a business doing Rs 10 lakh per month in sales, that gap represents Rs 1-2 lakh in either lost sales or wasted inventory. Over a year, that is Rs 12-24 lakh — enough to fund a second store.

How AI Demand Forecasting Works

AI forecasting is not magic. It follows a logical process that you can understand without any technical background:

Step 1: Collect Historical Data

The AI needs your past sales data — ideally 12-24 months. More data means better predictions. This includes: date of sale, product/SKU, quantity sold, price at the time, and any promotions running.

Step 2: Identify Patterns

The AI looks for three types of patterns:

Step 3: Generate Forecasts

Based on these patterns, the AI predicts demand for the next day, week, or month — with confidence intervals. Instead of saying "you will sell 100 units," it says "you will likely sell 85-115 units, with 90% confidence." This range is far more useful for ordering decisions.

Open data/sales-history.csv in the code panel. This file contains 12 months of daily sales data for a mid-size grocery store in Pune — 15 product categories, daily quantities, prices, and whether a promotion was active. This is the raw material AI works with.

Festival-Driven Demand: India's Unique Challenge

India's festival calendar creates demand spikes that would confuse any forecasting model not trained on Indian data. This is why generic global AI tools often fail for Indian retail.

The Festival Effect

The challenge is not just the spike — it is the timing. Diwali falls on a different date every year (lunar calendar). Ramadan shifts by about 10 days annually. A good AI forecasting system accounts for these moving dates automatically.

Regional Variation

A forecasting model trained on Delhi data will fail in Chennai. Pongal matters in Tamil Nadu but not in Punjab. Bihu matters in Assam but not in Maharashtra. AI must be trained on local data to capture these regional patterns.

Open data/festival-calendar.json to see a structured dataset of Indian festivals by region, typical demand multipliers, and key product categories. This is the kind of reference data that makes AI forecasting India-aware.

Building Your First Forecast

You do not need expensive software to start. Here is a practical approach using AI chat tools:

The Simple Moving Average (Baseline)

Before using AI, understand the baseline. Take your last 4 weeks of sales for any product and calculate the average. That is your simple forecast.

Week 1: 45 units
Week 2: 52 units
Week 3: 38 units
Week 4: 55 units

Simple forecast for Week 5: (45 + 52 + 38 + 55) / 4 = 47.5 units

This works reasonably well for stable products. But it completely misses trends, festivals, and external factors.

The AI-Powered Forecast

Now paste those same 4 weeks (or ideally 12+ weeks) into Claude or ChatGPT with this prompt:

Here is my weekly sales data for [product] at my store in [city]:
[paste your data]

Predict next week's sales considering:
- Any upcoming festivals or events
- The trend in the data
- Day-of-week patterns if daily data is available
- Any anomalies in the historical data

Give me a point estimate and a range.

The AI will identify patterns you might miss — like the fact that Week 3 was low because it had a public holiday, or that the upward trend suggests Week 5 should be closer to 58 than 47.

Working with Your Sales Data

What Data to Collect

If you are starting from scratch, begin tracking these fields daily:

Common Data Problems

Practical Exercise: Your 5-Step Forecast

Key Takeaways