Kyle Moloney
Procurement & Operations | 10+ Years
Kyle has spent over a decade managing procurement and operations for manufacturing companies ranging from small food producers to mid-size contract manufacturers. He now writes about practical inventory management, supply chain, and production operations.
Nearly every article on demand forecasting assumes you have years of clean sales data and a stable product line — the kind of dataset a big retailer or CPG brand has. Small-batch manufacturers rarely have that. You've got 12-18 months of history, a handful of SKUs that keep changing as you tweak the line, and sales that swing hard based on one wholesale order or a single social post going somewhere unexpected. Statistical forecasting models built for large, stable datasets don't fail gracefully on data like that — they just produce confident-looking numbers that are wrong.
That doesn't mean forecasting is pointless at small scale. It means the method has to match the data you actually have.
Simple Moving Averages, With Worked Numbers
For most small-batch manufacturers, a simple moving average is a better starting point than anything more sophisticated — it's transparent, it's easy to sanity-check by eye, and it doesn't pretend to more precision than lumpy data supports.
Worked example: a candle SKU
Say you sell a single candle SKU and want to forecast next month's demand using a 3-month moving average. Your last three months of units sold:
- Month 1: 340 units
- Month 2: 410 units
- Month 3: 375 units
3-month moving average forecast = (340 + 410 + 375) ÷ 3 = 1,125 ÷ 3 = 375 units for Month 4.
That's the base number. On its own, a moving average smooths out month-to-month noise, but it also lags — it won't catch a trend or a seasonal shift until it's already partway into the data. That's exactly where judgment needs to step in, which is the next section.
Seasonality by Eyeball, Plus a Judgment Overlay
With only 12-18 months of history, you don't have enough data to statistically model seasonality — you can't separate "this month is genuinely slower every year" from "this month happened to be slow once." What you can do is apply what you actually know about your business on top of the moving average.
If Month 4 in the candle example above is November and you know from last year (even a single data point) or from general knowledge of your category that scented candles pick up heading into the holidays, that's a reason to adjust the moving-average number upward — say, to 450-500 units — rather than treat 375 as the forecast. The moving average gives you a defensible baseline; the seasonal adjustment is a deliberate, documented judgment call on top of it, not a guess pulled from nowhere.
The discipline that matters here: write down *why* you adjusted the number, and check the adjustment against what actually happened once the month closes. That feedback loop is what turns "eyeballing it" into a judgment that gets more accurate over time, instead of a habit that never improves.
Why Per-SKU Forecasting Fails at Low Volume
If you're selling 15-20 units a week of any individual SKU, a per-SKU moving average is mostly noise — a single large wholesale order or a slow week from one retail account can swing the number by 50% or more, and the forecast will chase that noise instead of tracking real demand.
The fix is to aggregate to the product family level before forecasting, then allocate back down to individual SKUs using their historical share of the family's volume. If you sell a candle in four scents and each scent moves 15-20 units a week individually, but the family as a whole moves 70-80 units a week, forecast at the family level (which has enough volume for a moving average to mean something) and split that forecast across scents using each scent's typical percentage of family sales. You'll still be wrong on any individual scent in any individual week — but the family-level forecast, and the purchasing decisions for shared raw materials (wax, wicks, jars), will be far more stable than four separate noisy SKU forecasts.
MOQ and Lead-Time Interaction With Forecasts
A forecast tells you how much you'll likely need. It doesn't tell you how much you can actually order — that's where minimum order quantities (MOQs) and lead times complicate small-batch purchasing in ways big-retailer forecasting guides never address.
If your forecast says you need 375 units of wax for next month but your wax supplier's MOQ is 500, the forecast doesn't change what you order — you order the MOQ and carry the extra 125 units of wax forward into the following month's starting inventory. That's not a forecasting error; it's a structural constraint the forecast has to account for on the next cycle, not fight against on this one. Nstock's EOQ calculator can help you think through whether your standing order quantity should track closer to the MOQ or your actual usage, depending on holding cost and how much the supplier's MOQ exceeds your typical order.
Lead time works the other way — it determines how far ahead of the forecasted need you have to act. A supplier with a 6-week lead time means you're committing to a purchase based on a forecast that's 6 weeks stale by the time the material arrives, which is exactly why safety stock matters more, not less, for manufacturers with long-lead-time inputs and thin sales history. Nstock's safety stock calculator turns your usage variability and lead time into a specific buffer number rather than a guess.
Where AI Actually Helps at This Scale
It's fair to be skeptical of "AI forecasting" claims when you've only got a year of data — a lot of AI forecasting is built and marketed for exactly the large, stable datasets small manufacturers don't have. But at small scale, AI forecasting isn't valuable because it replaces your judgment with a black-box number. It's valuable because it can weigh multiple signals simultaneously (recent usage trend, open purchase orders, actual observed supplier lead time rather than the quoted one) faster and more consistently than manually recalculating a moving average and a safety stock buffer by hand every time something changes. Nstock's AI forecasting is built to layer onto whatever history you have — it starts producing usable projections well before you'd have enough data for a classic statistical model, and it gets more accurate as your production and sales history accumulates, rather than requiring years of clean data before it's useful at all.
The practical takeaway: don't wait until you have "enough" data to start forecasting formally. Start with a moving average and a documented judgment overlay now, keep a record of what you predicted versus what happened, and layer in more automated forecasting as your history builds.
Frequently Asked Questions
How much sales history do I actually need before forecasting is worth doing?
You can start a simple moving average with as little as 2-3 months of history, though it will be noisy and heavily influenced by one-off orders. Meaningful judgment overlays for seasonality generally need at least one full year of observed patterns, even if that's just a single data point per season rather than a statistically robust sample.
Should I forecast every SKU individually or by product family?
At low weekly volume — roughly under 30-50 units a week per SKU — individual SKU forecasts are usually too noisy to be useful. Forecast at the product family level, where volume is higher and the moving average is more stable, then allocate the family forecast down to SKUs using their historical share of family sales.
What do I do when my supplier's MOQ is bigger than what I've forecasted?
Order the MOQ and carry the excess forward as starting inventory for the next period rather than trying to force the order down to match the forecast exactly. Track whether the MOQ consistently exceeds a full period's need — if so, it's worth revisiting whether a different supplier or packaging size brings the MOQ closer to your actual usage.
Is a 3-month moving average always the right window?
Not always — a shorter window (like 2 months) reacts faster to genuine trend changes but is noisier; a longer window (like 6 months) smooths out more noise but lags further behind real shifts. For most small-batch manufacturers with lumpy monthly sales, 3 months is a reasonable default, but it's worth testing a longer window if your sales are especially erratic month to month.
Read the complete guide to AI inventory forecasting →
— Kyle Moloney, Procurement & Operations