The expensive part of an online sale is rarely the cardboard box. It is the box you bought six months too early, filled with 800 units of a colour nobody wanted, now ageing gracefully beside the bubble wrap. To forecast ecommerce demand properly is to make fewer heroic guesses before the warehouse starts looking like a museum of optimism.
For founders, marketers and operations teams, demand forecasting is not a crystal ball exercise. It is a practical way to decide what to buy, where to place it, when to replenish it and when to stop pretending a one-week TikTok spike is a permanent lifestyle choice.
Forecast ecommerce demand from signals, not vibes
Last year’s sales are useful. They are also dangerous when treated as scripture. A skincare brand that sold brilliantly during a December gifting rush should not automatically order the same quantity for February, when consumers are spending more cautiously and their gift-buying aunties have retreated until the next birthday.
A useful forecast combines historical demand with what has changed. Start with the product’s sales history by week, not just month. Weekly data shows patterns that a monthly total politely hides: payday lifts, weekend spikes, campaign effects, out-of-stock periods and the strange Tuesday when a creator mentioned your product to 400,000 people.
Then put those figures in context. Ask whether the product was discounted, bundled, featured in an email, backed by paid media or temporarily unavailable. A sales number without context is like a restaurant review that omits the fact the reviewer arrived after three cocktails.
Your forecast should also separate demand from availability. If you sold 50 units because only 50 were left, demand may have been 50 or it may have been 200. Check page views, add-to-basket rates, waitlist requests, customer service messages and abandoned baskets around the stockout. They are imperfect clues, but far better than assuming the shelf told the whole story.
Build a base case before chasing the upside
For most SMEs, a simple base forecast beats an elaborate spreadsheet nobody updates. Take comparable weeks from recent months, adjust for seasonality, and calculate a reasonable expected run rate. Then create an upside case for promotional activity and a downside case for softer trading.
The point is not to predict a precise number. Precision can be a very convincing costume for uncertainty. The point is to give purchasing and fulfilment teams a range they can act on.
For example, a reusable bottle brand may expect 300 orders a week based on recent sales. An upcoming school holiday, a marketplace campaign and a creator partnership could push that to 500. Rather than buying as though 500 is guaranteed, it may order enough stock for the base case, secure a fast replenishment option, and reserve packing capacity for the higher scenario. That is not timid planning. It is paying attention.
The demand data that actually earns its keep
Not every metric deserves a seat at the forecasting table. Revenue can flatter a forecast when prices change. Total orders can mislead when bundles become popular. Start with units sold by SKU, then layer in the signals that explain why those units moved.
The most useful data usually includes historical unit sales, current stock on hand, stock already in transit, lead times from suppliers, returns, cancellations, marketplace promotion calendars and marketing activity. If you sell across Shopify, Shopee, Lazada, TikTok Shop and a physical counter, channel-level data matters too. A marketplace flash sale can empty inventory intended for your own site by lunchtime, where it will be rediscovered only when someone says, “Why are we overselling?”
Returns deserve particular respect. Apparel sellers know this well. A dress may appear to be a bestseller until the return rate reveals that half of buyers are ordering three sizes and sending two back. Forecast gross orders alone and you may overstate demand, understate handling work and make your warehouse team unusually familiar with disappointment.
Use a rolling view of three periods where possible: the latest four weeks for momentum, the same period last year for seasonality, and a longer trailing average for stability. Each has flaws. Recent sales can be noisy; last year may be irrelevant after a price change or new channel launch; long averages react slowly. Together, they produce a more honest picture.
Treat channels as different shops with different weather
A product does not behave identically everywhere. Shoppers on a brand site may buy thoughtfully, read ingredients and return for refills. Marketplace shoppers may respond to vouchers, free shipping thresholds and the general thrill of seeing a countdown timer. Live-commerce customers may buy because a host demonstrated the product with the urgency of a person escaping a sinking ship.
Forecast each channel separately before combining the total. That lets you see whether growth is genuine demand or simply demand moving from one channel to another. It also helps with inventory allocation. The last thing a brand needs is to win a marketplace campaign and lose its higher-margin direct customers because every unit was sent to the wrong virtual shelf.
This is where operational design starts to matter as much as the spreadsheet. uParcel in Singapore and Malaysia works as a multi-channel fulfilment partner on a cloud-based, commerce-enabled operation, with live studios beside fulfilment and marketplace management in the same orbit. Its engineering, fleet network, warehouse and commerce teams are directly controlled, which matters when stock must move between channels without a game of email ping-pong.
You do not need a large operation to borrow the principle: keep one dependable view of inventory, and ensure marketing knows what operations can actually fulfil. The promotional calendar should never be a surprise sprung on the people packing the orders.
Put lead time at the centre of the forecast
Demand forecasting without replenishment lead time is just interesting trivia. If your supplier needs 45 days, sea freight takes 14, quality checks take three, and a public holiday appears in the middle like a plot twist, then a forecast for next week is mostly too late to help.
Map the full lead time from purchase order to sellable stock. Include manufacturing, freight, customs, receiving, put-away and any labelling or kitting. Then add a buffer based on variability, not nerves. A supplier that reliably delivers in 20 days needs a different buffer from one that delivers in 20 days, except when it takes 38.
Set a reorder point for each meaningful SKU: expected daily sales multiplied by lead time, plus safety stock. Review it regularly. A reorder point written during a quiet month can become dangerously quaint after a campaign, a new listing or a sudden rise in repeat purchase.
For fast movers, check stock and inbound inventory daily. For slower, stable products, a weekly review may be enough. The right rhythm depends on sales velocity, margin, storage cost and how painful a stockout would be. Running out of a hero product can cost more than the missed sale. It can send a customer to a competitor and teach them a new habit.
Make the forecast a conversation, not a spreadsheet burial ground
The best forecasts are revised. That is not failure; it is the job. Hold a short weekly review involving whoever owns buying, marketing, marketplace activity and operations. Look at forecast versus actuals, explain material gaps and update the next few weeks.
Keep the meeting blunt. Did sales rise because demand improved, or because a voucher made the price temporarily irresistible? Did they fall because customers lost interest, or because the product page went out of stock at 2pm? Was a campaign successful, or did it merely pull next week’s orders into this week?
Write down the assumptions. If the forecast expects 1,000 units from a TikTok Shop live, note the expected audience, conversion rate, offer and available inventory. Afterwards, compare reality to the assumption. Over time, your team becomes better at forecasting because it learns which beliefs were useful and which were merely enthusiastic.
Watch for the traps that make clever teams look careless
Forecasting fails most often when people mistake a special event for a trend. A viral post, a one-off corporate order, a clearance sale or a competitor’s stockout can distort demand. Flag these events rather than letting them quietly inflate the average.
New products require a different approach because they have no history. Use an analogue: a similar price point, category, customer and channel. Start with a conservative buy, especially when replenishment is possible. Ordering enormous quantities because the packaging looked excellent in a presentation is a rite of passage, but an expensive one.
Finally, do not forecast only what you intend to sell. Forecast the work that sale creates: picks, packs, inserts, gift wrapping, split shipments, returns and customer enquiries. A campaign that doubles orders may more than double complexity if every order contains a bundle assembled from three locations.
A good forecast will not remove uncertainty from ecommerce. Nothing can, short of banning people from changing their minds. What it can do is turn uncertainty into choices: buy now or later, hold more stock or protect cash, push a campaign or wait. That is a far better position than discovering, at 11.47pm, that your bestseller sold out during the livestream and the next container is still somewhere at sea.
Leave a Reply