Quick commerce demand forecasting for dark stores, explained
How 10-minute operators turn order history, pin-code clusters, weather and festival signals into dark store assortment and replenishment — and what forecast error costs.
The short answer
Quick commerce firms forecast demand by combining historical order data at pin-code level with weather, festival and local event signals, then group SKUs by sales velocity and set replenishment cycles per tier. Forecast error shows up as stockouts on fast movers, wastage on perishables and excess working capital tied up in slow stock.
Key takeaways
- Quick commerce forecasting works at store or pin-code cluster level, not city level, because nearby stores can have very different baskets.
- Weather, festivals, local events and promotions are explicit inputs alongside order history, since they explain the spikes history alone cannot.
- SKUs are sorted into velocity tiers, and each tier gets its own replenishment cycle and safety stock; perishables add a shelf-life constraint.
- Forecast error appears as stockouts on fast movers, wastage on perishables and excess working capital in slow stock.
- Sales history must be adjusted for stockout days, or the model will keep under-ordering products that were simply unavailable.
Quick commerce demand forecasting is the process a 10-minute delivery operator uses to decide how much of each product a dark store should hold, and when to refill it. A dark store is a small, order-only warehouse — typically a few thousand square feet — that serves a two to three kilometre radius. Because it cannot hold everything, the forecast decides what a customer sees as available and what quietly runs out.
The job is harder than normal retail planning. A supermarket sees demand build over days and can reorder weekly. A dark store has to be right within hours, on thousands of SKUs, across hundreds of pin codes that behave differently. The forecast is not a spreadsheet exercise; it is the operating system of the store.
How does quick commerce demand forecasting work in practice?
Most operators run the same loop: collect signals, forecast at a granular level, translate the forecast into an order, and measure how wrong the forecast was.
- Signal collection. Every order, search, add-to-cart and out-of-stock tap is logged with a timestamp, pin code and store ID. Searches that end with no purchase are a demand signal for products the store does not stock.
- Clustering. Stores are grouped by behaviour, not just geography — a cluster of office pin codes orders differently from a cluster of residential towers or a student area. Clusters let a new store borrow a demand pattern before it has its own history.
- Forecasting. Statistical models (moving averages, exponential smoothing) and machine-learning models produce a demand number per SKU per store per day, often broken into time slots.
- Replenishment. The forecast is converted into purchase orders to the mother warehouse or hub, adjusted for lead time, minimum order quantities and shelf life.
- Measurement. Forecast error, stockout rate, wastage and inventory days are reviewed and fed back into the model.
Why does pin-code-level clustering matter more than city averages?
A city-level average hides the variation that actually decides stockouts. Two stores five kilometres apart can have completely different baskets: one sells baby care and breakfast cereal, another sells late-night snacks and cold drinks. Planning both on a city average guarantees one store overstocks and the other runs dry.
Operators therefore forecast at the store or pin-code cluster level. New stores are the tricky case. With no history, the model uses the nearest comparable cluster, then re-weights as real orders arrive. The first four to six weeks are effectively a learning period, and assortment is often deliberately narrow until the store's own pattern emerges.
What signals beyond order history do forecasters use?
Order history explains the base. The spikes come from elsewhere.
- Weather. Rain, heatwaves and cold snaps shift baskets towards hot beverages, ice cream, umbrellas and instant food. A forecast that ignores a rain alert will understock exactly when demand jumps.
- Festivals and calendar events. Diwali, Holi, Eid, regional new years and long weekends change both basket mix and order timing. Gifting, sweets and house-guest staples move; everyday cooking staples may dip.
- Local events. A cricket match, a concert, an exam week or a building society event can lift a single store's demand for hours.
- Promotions and paydays. Discount events and salary-credit dates create predictable demand bumps that the model must learn rather than treat as noise.
Each signal is a feature in the model. The value is in weighting them correctly for each cluster — rain means something different in a hill town than in a coastal city.
How do SKU velocity tiers drive replenishment cycles?
Not every product deserves the same planning effort. Operators sort SKUs into velocity tiers — fast movers that sell many units a day, mid-tier items, and slow movers that sell occasionally but are expected by customers.
Fast movers get frequent replenishment, tight safety stock and daily review, because a stockout there costs real sales. Mid-tier SKUs get scheduled refills a few times a week. Slow movers are stocked in small quantities and reviewed weekly; some are deliberately kept only in larger stores. Perishables — milk, bread, fruit, cut vegetables — add a shelf-life constraint that overrides the pure demand number, because a wrong order spoils rather than waits.
This tiering is why assortment differs between stores even within one brand. It is also why a category manager's job is part forecast, part negotiation with suppliers on pack sizes and lead times.
What does forecast error cost a dark store?
Forecast error is the gap between predicted and actual demand. It shows up in three places.
Stockouts. Under-forecast a fast mover and the shelf is empty. The customer may substitute, or may simply open a rival app. Repeated stockouts train customers to check elsewhere first.
Wastage. Over-forecast perishables and the excess is written off. In fresh and dairy, small errors compound quickly.
Working capital. Over-forecast slow movers and cash sits in inventory that does not move. Every extra day of stock is money not available for new stores, marketing or supplier terms.
What should a founder check before building a forecasting stack?
Before buying or building models, get the basics right.
- Data hygiene. Store IDs, SKU codes and timestamps must be consistent. Bad master data defeats good models.
- Stockout-adjusted history. Sales data records what sold, not what customers wanted. If a product was unavailable, its history understates demand. Flag stockout days or the model will keep under-ordering.
- Lead times and pack sizes. A perfect forecast is useless if the supplier delivers in cases of a couple of dozen and takes three days.
- Shelf life rules. Perishables need expiry-aware ordering, not just demand numbers.
- A simple error metric. Track forecast error, fill rate and wastage weekly before adding complexity.
Common mistakes: planning on city averages, ignoring out-of-stock days in history, treating every SKU with the same review cycle, and expanding assortment faster than the store's data can support.
What it means for Indian retailers
Quick commerce is pulling assortment planning into the open. Brands selling through these platforms now negotiate on pack sizes, replenishment frequency and visibility, and a brand that cannot supply reliably in small, frequent lots loses shelf space. Traditional retailers watching the format can borrow the discipline: cluster your stores by basket behaviour, tier your SKUs by velocity, and measure stockouts honestly. Nykaa's Q2 FY27 results offer one data point on how the format is expanding: Nykaa Now expanded to fifteen cities in the quarter from three in Q1 FY26, with a target of over twenty-five cities by fiscal year-end, according to the company. More cities mean more clusters, more local demand patterns and more forecasting work.
In this story
Logos are trademarks of their owners and are shown to identify them.
Frequently asked questions
What is demand forecasting in quick commerce?
It is the process of predicting how much of each product a dark store will sell over a short period, using order history, pin-code clusters, weather, festivals and promotions, then converting that prediction into replenishment orders and safety stock.
Why do dark stores forecast at pin-code level?
Baskets vary sharply between nearby areas — office districts, residential towers and student zones order differently. A city average would overstock one store and leave another short, so operators cluster stores by behaviour and forecast each cluster separately.
How do weather and festivals change dark store inventory?
Rain, heat and cold shift demand towards umbrellas, beverages, ice cream and instant food. Festivals move gifting, sweets and guest staples. Treating these as model inputs prevents understocking during predictable spikes.
What is SKU velocity tiering?
It is sorting products by how fast they sell. Fast movers get frequent replenishment and tight safety stock, mid-tier items get scheduled refills, and slow movers are stocked in small quantities and reviewed less often.
What does forecast error cost a quick commerce operator?
Under-forecasting causes stockouts and lost sales on fast movers. Over-forecasting perishables causes wastage, and over-forecasting slow movers ties up working capital in inventory that does not move.
