Every inventory decision your team makes — how much to buy, when to reorder, which SKUs to prioritize — rests on some version of demand forecasting. The question isn't whether you're forecasting. It's whether your forecast is explicit, defensible, and accurate enough to act on — or whether it's buried in a gut call made by an ops manager staring at last month's spreadsheet.
For brands at the $5M–$50M GMV range, the gap between those two approaches isn't academic. It's the difference between entering Q4 with the right inventory position and scrambling to expedite freight at a 40% premium three weeks before Black Friday.
What demand forecasting actually means
Demand forecasting is the process of estimating future customer demand for a product over a specified time period. For an e-commerce brand, that means: how many units of SKU WMN-HOODIE-BLK-M are you going to sell in the next 30, 60, or 90 days?
The output of that estimate drives every downstream decision in your replenishment workflow: your reorder point, your purchase order quantities, your safety stock buffer, and your days of cover target. Get the forecast wrong — in either direction — and you pay for it. Overforecast and you're sitting on capital in a warehouse. Underforecast and you stockout mid-peak.
What distinguishes a good forecast from a guess isn't complexity. It's the structured use of historical demand signals, adjusted for known future conditions: upcoming promotions, seasonal patterns, channel mix changes, and lead time windows.
Forecasting methods: from simple to sophisticated
Most mid-market brands don't need to implement academic forecasting algorithms manually. But understanding the methods matters because they determine what your forecasting tool can and cannot model.
Moving average and exponential smoothing
The simplest approaches use weighted averages of recent sales. A basic 4-week moving average takes the last four weeks of unit sales and divides by four. Exponential smoothing — specifically the Holt-Winters triple exponential smoothing model — extends this by applying decay weights so that recent data counts more than older data, and adds components for trend and seasonality.
For SKUs with relatively stable, frequent demand, exponential smoothing is reliable and computationally cheap. The weakness shows up when demand is intermittent — a SKU that sells 3 units one week, 0 the next, 12 the week after. Holt-Winters wasn't built for that pattern.
Croston's method for intermittent demand
Intermittent demand is common in mid-market catalogs. Slow-moving accessories, seasonal colorways, B-tier SKUs in a large assortment — these don't sell every day or every week. Croston's method handles this by separately forecasting the demand size and the demand interval, then combining them into a per-period estimate. It's the standard approach for intermittent demand in inventory management.
A catalog of 1,500 SKUs might have 600 that sell daily, 700 that sell weekly, and 200 that sell monthly or less. The right forecasting method differs across those three segments.
ARIMA and machine learning approaches
ARIMA (Autoregressive Integrated Moving Average) models capture autocorrelation in a time series — the idea that this week's sales are correlated with last week's and the week before. ARIMA handles complex temporal patterns and can incorporate external variables (promotions, ad spend) through ARIMAX extensions.
Machine learning approaches — gradient boosted trees, neural networks, Bayesian forecasting — can incorporate a much wider feature set: web traffic, marketing calendar, competitor pricing signals, weather patterns for seasonal goods. These methods improve accuracy at scale but require data volumes and engineering depth that most mid-market teams don't have in-house. Purpose-built forecasting platforms handle this under the hood, which is the primary reason brands at 500+ SKUs look beyond spreadsheets.
Why category-level forecasting fails growing brands
Many brands start by forecasting at the category level: "we expect denim to be up 15% next quarter." The problem is that category-level signals don't translate to SKU-level purchase orders. You buy a specific style, in a specific colorway, in specific sizes. A 15% category lift that's concentrated in one colorway and ignores size run distribution will leave you stocked out in M/L and sitting on XS/XXL you can't move.
SKU-level forecasting, sometimes called bottom-up forecasting, models demand at the level at which you actually place purchase orders: the individual SKU. This is more computationally intensive — 1,500 SKUs means 1,500 individual demand models — but it's the only approach that produces actionable replenishment quantities.
Hierarchical forecasting frameworks reconcile the two: you maintain SKU-level models, but constrain them so that the sum of SKU-level forecasts is consistent with category and total revenue projections. This prevents the "I'm right about the total but wrong about the mix" error that trips up brands heading into a promotional period.
The forecast accuracy problem: what MAPE actually tells you
Most forecasting tools surface a headline accuracy metric. MAPE — Mean Absolute Percentage Error — is the most commonly cited. A MAPE of 20% means your forecast is off by 20% on average.
MAPE has a known failure mode: it's undefined when actual demand is zero (you can't divide by zero), and it's systematically biased against low-volume SKUs. A SKU that sells 2 units and you forecast 3 shows 50% MAPE. A SKU that sells 200 and you forecast 240 shows 20% MAPE. MAPE treats these the same, but they have very different cash flow implications.
MASE (Mean Absolute Scaled Error) addresses some of these issues by scaling error relative to a naive benchmark. RMSE (Root Mean Squared Error) penalizes large misses more heavily than small ones, which aligns better with the actual inventory cost structure — a 100-unit underforecast on a hero SKU costs far more than ten 10-unit misses across slow movers.
We're not saying MAPE is useless — it's fast to compute and easy to communicate to a CFO. We're saying it shouldn't be your only accuracy signal, and it should never be applied uniformly across SKUs with very different demand patterns and velocity tiers.
Where spreadsheet forecasting breaks
Consider a mid-market apparel brand based in Portland managing roughly 800 active SKUs across Shopify (their DTC channel) and Amazon Seller Central. In early 2025, their ops team was running demand planning in a shared Google Sheet — pulling weekly sales reports from both channels, manually combining them, and applying a rough multiplier based on last year's comparable period.
The spreadsheet worked at 200 SKUs. At 800, it became a bottleneck: two and a half days of manual data consolidation every month, persistent formula errors from channel-merge inconsistencies, and no systematic handling of lead time variability (their overseas supplier ranged from 45 to 90 days depending on season). When demand for a new colorway surprised to the upside during a mid-season sale, they didn't catch the signal in time. They stocked out three weeks into the run, missed the sell-through window, and carried unsold inventory from the following purchase order well into the next season.
This isn't an unusual story. According to Digital Commerce 360's annual operations survey data, inventory management consistently ranks among the top three operational challenges for mid-market e-commerce brands. The constraint isn't lack of data — Shopify and Amazon generate plenty. The constraint is the tooling to turn raw channel data into SKU-level demand signals at the cadence decisions actually need to be made.
What a good forecast enables downstream
Demand forecasting isn't an end in itself. Its value shows up in the decisions it enables:
- Reorder point calculation: Your reorder point is a function of your average daily demand and your supplier lead time. Without an accurate demand forecast, your reorder point is a guess — and it's typically wrong in the direction that hurts most (too low), because ops teams are more afraid of overstock than stockouts when they set the initial number.
- Safety stock sizing: Safety stock is designed to cover demand variability during the lead time window. Sizing it correctly requires understanding the statistical distribution of your demand — not just the mean, but the variance. Forecast-based safety stock models (using forecast error standard deviation) are materially more precise than the rule-of-thumb buffers most brands use.
- Purchase order quantities and economic order quantity: EOQ optimization balances order cost against holding cost. A more accurate demand forecast reduces the uncertainty that forces you to over-order defensively.
- Cash flow planning: Knowing your expected sell-through by SKU 60–90 days out lets your finance team plan working capital more precisely. Inventory is typically the largest current asset on a mid-market brand's balance sheet.
If you're evaluating whether your current forecasting approach is working, the most direct test isn't accuracy metrics — it's this: how often does your team get surprised by a stockout or overstock situation that a forecast should have predicted? If the answer is "often," the issue isn't your team's competence. It's the tooling.
Stockorlo connects to your existing e-commerce stack — Shopify, Amazon Seller Central, WooCommerce, NetSuite — and builds demand models at the SKU level using your actual sales history. See how the forecasting engine works, or explore the platform integrations to check if your current stack is covered.