AI demand forecasting predicts how many units of each product you will sell, broken down by location and date, by learning patterns from your own sales history plus the signals that move demand: price, promotions, festivals, day of week, season, and stock on hand. The output is a concrete number, units per product per location per week, that drives a single decision: how much to reorder and when.
That last part is what separates a forecast that earns money from a chart nobody acts on. A good forecast flows straight into a reorder calculation and ends as a purchase order, not a slide. Applied to supply chains, AI-driven forecasting can cut forecast errors by 20 to 50 percent and reduce lost sales from product being unavailable by up to 65 percent (McKinsey, 2022). The stakes are large because the failure is expensive on both sides: globally, the mix of overstocks and stockouts (inventory distortion) cost retailers an estimated 1.77 trillion US dollars, of which lost sales from running out alone were about 1.2 trillion (RELEX and IHL Group, 2023).
India’s e-commerce leaders run exactly this loop at extreme scale. With the market at roughly 211.6 billion US dollars in 2025 (IBEF, 2025), and festive weeks compressing a huge share of the year into days, Flipkart, Amazon India, and the quick-commerce platforms live or die on getting the next reorder right. The mechanics they use scale down to a 500-product retailer. This is the real how.
What demand forecasting actually is
A demand forecast is an estimate of future sales at a specific level of detail: which product, which location, which time window. The level matters more than people expect. A single company-wide number is easy and useless. The useful forecast is granular, “this SKU at this warehouse next week,” because that is the level at which you actually place orders.
Two ideas underpin everything below. First, demand is not the same as sales. If you stocked out, your sales were capped below true demand, so raw sales history understates what people wanted. Good forecasting corrects for this. Second, demand is not random; it is driven. Price, promotions, the festive calendar, day of week, and weather all push it, and the job of a model is to learn how much each one pushes.
The data inputs that drive a forecast
A forecast is only as good as the signals feeding it. The raw material is scattered across tools, which is why this is a data problem before it is a model problem. The inputs that matter:
- Historical sales: daily or weekly units per product per location, ideally one to two years so the model sees a full festive cycle.
- Seasonality and calendar: day of week, month, and the Indian festive calendar (Diwali, Dussehra, Eid, Raksha Bandhan, regional festivals), tagged as known events.
- Festival and sale events: Big Billion Days, Great Indian Festival, and your own flash sales, marked with their start and end dates.
- Price and promotions: your selling price over time and any discount or offer, since a price cut on Tuesday explains the spike on Tuesday.
- Stock on hand and stockouts: so the model knows when low sales meant low demand versus an empty shelf.
- External signals: weather for weather-sensitive goods, competitor pricing, and search or traffic trends where you can get them.
Getting these into one clean, time-aligned table is the unglamorous half of the project. Sales sit in one system, prices in another, the festive calendar in someone’s head. Pulling them together and lining them up by date is data engineering work, and it is exactly where most forecasting efforts quietly fail. At Galific we audit the data first: we check whether the history is long enough, consistent, and complete enough to support a forecast, before anyone builds a model. Honest forecasting starts with honest data.
The model families, and when each one fits
There is no single best algorithm. The skill is matching the model to the demand pattern, and trying a few rather than betting on one. Three families cover most SME needs.
Time-series models (ARIMA, Prophet). These learn from the shape of the sales line itself: its trend, its weekly and yearly seasonality, and its response to known events. ARIMA (AutoRegressive Integrated Moving Average) is a classical statistical workhorse. Prophet, an open-source library from Meta, is popular because it handles seasonality and holiday effects with little tuning, which suits steady, seasonal sellers like staple groceries or fast-moving consumer goods.
Machine-learning models (gradient boosting, LightGBM). When demand depends heavily on price, promotions, and many interacting factors, tree-based gradient boosting shines. LightGBM and XGBoost are the standard tools: you give them columns for price, discount, festival flags, day of week, stock, and they learn the combined effect. This is the family most large e-commerce players lean on for promotion-heavy catalogs, because it captures “a 20 percent discount during Big Billion Days lifts this category by X” far better than a plain trend line.
Intermittent-demand methods (Croston, TSB). Slow movers break normal models. A product that sells zero units most weeks and three units occasionally has no smooth trend to fit, so ARIMA and even gradient boosting produce nonsense. Croston’s method, and its successor TSB (Teunter, Syntetos, Babai), were built exactly for this: they forecast the size of a sale and the gap between sales separately, then combine them. For any catalog with a long tail of rarely sold items (spare parts, niche SKUs), this is the family that prevents you from either drowning in dead stock or stocking out on the odd order.
How you measure whether a forecast is any good
A forecast you cannot score is a guess with confidence. Two metrics do the heavy lifting.
MAPE (Mean Absolute Percentage Error) is the average percentage you were off, across all products and weeks. A MAPE of 20 means your forecasts were typically 20 percent wrong, in either direction. It is intuitive and widely used.
WMAPE (Weighted MAPE) fixes MAPE’s main flaw. Plain MAPE treats a small miss on a slow product the same as a small miss on a bestseller, and it blows up to huge percentages on low-volume items. WMAPE weights each error by sales volume, so a big miss on a high-runner hurts your score more than a miss on something you sell twice a month. For a real catalog with a long tail, WMAPE is the fairer number, which is why most serious operators track it.
The discipline that matters more than the metric: always compare against a naive baseline. The simplest baseline is “next week equals last week” or “this week equals the same week last year.” If your fancy model cannot beat that, it is not earning its cost. A model is worth deploying only when it beats the baseline on the products that matter.
Lessons from Indian e-commerce
The giants did not invent new math. They got disciplined about four things that are specific to selling in India, and each one transfers directly to a smaller business.
Festivals are events, not noise. Diwali week and Big Billion Days are not random spikes to be smoothed away; they are known, dated events with their own demand uplift. Festive periods in India account for roughly one in four new online shoppers and drive close to triple the normal daily traffic during big sale events (Bain and Company, 2024). The platforms tag these dates as inputs so the model learns each category’s festive lift. A plain monthly average does the opposite: it spreads the Diwali peak across surrounding weeks and you stock out exactly when demand is highest.
India is not one market. Demand for the same product varies sharply by region, language, and city tier. A category that booms in Tier 2 towns may be flat in metros. The lesson, which carries straight into any retail and e-commerce operation, is to forecast at a regional or warehouse level where the data allows, rather than forcing one national number, because that national number hides the very variation you need to stock against.
Cash on delivery distorts demand. A large share of Indian orders are cash on delivery, and a meaningful portion are refused at the door and returned. Industry logistics data puts the return-to-origin rate on COD orders at around 26 percent, against under 2 percent for prepaid orders (Shipway, 2024). That means gross orders overstate the demand that actually sticks. The platforms net out expected returns so they plan inventory against kept demand, not placed demand. Plan on gross COD orders and you will reliably over-order.
Re-forecast constantly. Quick commerce pushed this to an extreme: forecasts refreshed continuously, down to the pincode and the hour, because a dark store has minutes of buffer, not days. You do not need per-minute forecasting, but the principle holds: a forecast refreshed weekly beats one set once a month and forgotten, because it absorbs last week’s surprises.
How the forecast becomes a reorder
This is the step that turns analysis into money saved. The forecast is an input to a reorder calculation, not the end product. The logic is straightforward:
Order quantity = (forecast demand over the supplier lead time) plus (safety stock) minus (stock on hand) minus (stock already on order)
Each piece earns its place. The lead time is how long your supplier takes to deliver, so you must cover demand for that whole window, not just next week. Safety stock is a deliberate buffer sized to the forecast’s uncertainty and your service-level target; a 95 percent service level (rarely stocking out) needs more buffer than 90 percent. Subtracting stock on hand and on order stops you double-ordering what is already coming. When the result crosses your reorder point, the system raises a purchase suggestion, and that suggestion is the whole point. McKinsey’s supply-chain work finds AI-driven planning can reduce inventory levels by 20 to 30 percent while improving service, precisely because better forecasts let you hold less safety stock for the same protection (McKinsey).
For an SME, this is the wiring that pays for the project: forecasts flowing into reorder points inside the tools you already use, so a number on a screen becomes a goods-inward at your warehouse. It is the same loop behind demand forecasting and the broader data intelligence work, where clean data becomes a forecast becomes a decision. And once the forecast is sitting in a database, you can simply ask your data questions like “which SKUs are at risk of stockout this festive week” in plain language, instead of waiting for a report.
What a small business can realistically expect
You will not match a giant’s data team, and you do not need to. With one to two years of clean sales history, the right model per product group, an honest WMAPE you track every week, and forecasts wired to your reorder points, the gains are real: fewer stockouts on your bestsellers, less capital frozen in slow stock, and far less firefighting around festive weeks. The published benchmarks (20 to 50 percent lower forecast error, up to 65 percent fewer lost sales) come from exactly this discipline, applied honestly.
The cheapest insurance against a wasted project is the order of operations. Audit the data first to confirm it can support a forecast, prove the model beats a naive baseline on the products that matter, then wire it into reordering. That is how Galific builds demand forecasting for Indian SMEs: data first, honest accuracy, and a forecast that ends as a decision rather than a chart.
