Demand Forecasting for Inventory Optimization

Demand Forecasting for Inventory Optimization#

Most companies carry 30-50% more inventory than needed because forecasts are inaccurate. Improving forecast accuracy from 60% to 80% (measured by MAPE) reduces safety stock by 25% and stockouts by 40%. Methods: moving averages for stable demand, exponential smoothing for trending products, and machine learning for seasonal/promotional items.

Safety Stock Calculation Methods#

Safety stock = Z × σ × √L, where Z = service level factor (1.65 for 95%), σ = demand standard deviation, L = lead time in periods. For a product with 100 units/week average demand, 20 units standard deviation, and 4-week lead time: safety stock = 1.65 × 20 × √4 = 66 units. This covers 95% of demand scenarios during lead time.

Collaborative Forecasting with Key Customers#

Your top 20% of customers drive 80% of demand. Share forecast data with them through vendor-managed inventory (VMI) or collaborative planning (CPFR). When your largest customer shares their promotion calendar, your forecast accuracy for that account jumps from 65% to 85%. The effort to collaborate pays for itself in lower inventory costs.

Walking Through a Moving Average Forecast Step by Step#

A simple moving average forecast takes the average of the most recent N periods of actual demand and uses that as the forecast for the next period, updating as new data arrives. Consider a specialty tea importer forecasting demand for a bestselling blend using a 4-week moving average. If the last four weeks of sell-through were 210, 195, 230, and 245 units, the moving average forecast for the coming week is (210+195+230+245)/4 = 220 units. As week five's actual demand comes in — say 260 units, reflecting a slow upward trend — the forecast for week six recalculates using weeks two through five: (195+230+245+260)/4 = 233 units. The moving average is simple to compute and explain to a small team, which is its main strength, but it lags behind trends because it weights all four periods equally rather than giving more importance to the most recent data. For a product with a steady upward trend, a moving average will consistently under-forecast; for a product trending down, it will consistently over-forecast and lead to excess stock. This is exactly why exponential smoothing, which weights recent periods more heavily, tends to outperform simple moving averages for any product with a persistent trend rather than flat, stable demand.

Measuring Forecast Accuracy the Right Way#

Mean Absolute Percentage Error (MAPE) is the most common way small businesses track forecast accuracy, calculated as the average of the absolute difference between forecast and actual, divided by actual, across a set of periods. A pet supplies distributor forecasting a mid-tier SKU predicted 500 units for a month and actually sold 430 — an absolute percentage error of |500-430|/430 = 16.3% for that period. Tracked across 12 months and averaged, a MAPE in the 15-25% range is fairly typical for an SMB without sophisticated forecasting tools, while a MAPE under 10% is considered strong performance achievable mainly for stable, non-seasonal products. What matters more than the raw number is tracking MAPE by SKU category over time: a rising MAPE trend for a specific product line is an early signal that something has changed in the demand pattern — a new competitor, a shifting customer base, a seasonality shift — before it shows up as a stockout or overstock event. Businesses that only look at forecast accuracy in hindsight, after a stockout has already happened, are using the metric too late to be useful.

Common Forecasting Mistakes That Undermine Inventory Planning#

The most common mistake is applying the same forecasting method to every SKU regardless of its demand pattern — using a simple moving average on a highly seasonal product, or an unadjusted exponential smoothing model on a product with a hard promotional spike, guarantees a bad forecast no matter how carefully the math is done. A second mistake is forecasting at too aggregated a level, such as forecasting total category demand and then splitting it evenly across SKUs, which masks the fact that individual products within a category often move in opposite directions. A third mistake is failing to strip out one-off demand events — a single large wholesale order, a stockout-driven demand spike as customers overbuy in fear of another shortage — from the historical data used to build the forecast, which distorts every future period's baseline. A fourth mistake, especially common among growing SMBs, is never revisiting the forecast method itself; a business that has grown from 5 to 50 SKUs often still uses the same basic spreadsheet approach that worked for a much smaller, simpler catalog, and the forecasting error compounds as complexity grows without a matching upgrade in method.

Connecting Forecasts to Purchasing Decisions in Practice#

A forecast only creates value once it changes a purchasing decision, and the businesses that get the most from demand forecasting build a direct, largely automated link between the forecast number and the reorder trigger, rather than treating forecasting as a separate analytical exercise reviewed occasionally by a manager. AskBiz's trade intelligence tracking supports this by keeping historical sell-through, current stock position, and supplier lead time together in one place, so a business can see not just what the forecast says but what it implies for the next purchase order given current stock coverage. For an importer juggling dozens of SKUs across multiple suppliers with different lead times, having this connection visible in one dashboard — rather than reconstructed manually in a spreadsheet each planning cycle — is often the difference between a forecast that actually gets used and one that gets built once and then ignored until the next stockout forces a scramble.

Handling Seasonal and Promotional Demand Spikes#

Products with a strong seasonal pattern need a fundamentally different forecasting approach than stable, steady-demand items — applying a moving average or simple exponential smoothing to a seasonal SKU will chronically under-forecast the peak and over-forecast the trough. A garden furniture importer selling into a spring-summer season should build a seasonal index by comparing each month's historical sales to the annual average, then apply that index to the current year's baseline trend. If April historically runs at 180% of the monthly average and the underlying trend suggests average monthly demand is growing 10% year over year, next April's forecast should be the new baseline average multiplied by 1.8, not a simple continuation of March's numbers. Promotional spikes require a separate adjustment layered on top of the seasonal index — if a planned promotion historically lifts demand by 40% during its run, that lift should be added explicitly to the forecast for those specific weeks rather than assumed to be captured by the general trend, and removed again from the historical baseline once the promotion ends so it does not distort future forecasts.

Starting Small: A Practical First Step for SMBs#

Businesses without an existing forecasting practice do not need to build a sophisticated statistical model on day one. The highest-leverage starting point is picking your top 10-20 SKUs by revenue, pulling 12-18 months of weekly or monthly sell-through history for each, and simply calculating a 4-period moving average alongside actual results to see how far off a naive forecast already is. This baseline exercise alone usually surfaces which products are stable enough for simple methods and which need seasonal or trend adjustments, giving a clear roadmap for where more sophisticated forecasting effort will actually pay off rather than spreading equal attention across every SKU regardless of its value or predictability.