Understanding forecast bias in FBA inventory planning
Forecast bias measures whether your demand forecasts are consistently too high or too low over time. A forecast that is off by 10% randomly each month is workable because the errors cancel out, and your safety stock buffer covers the swings. A forecast that is off by 10% in the same direction every month is a problem: it causes either chronic excess inventory (over-forecasting) or recurring stockouts (under-forecasting).
Forecast bias is different from forecast accuracy (MAPE). Accuracy measures how far off the forecast is regardless of direction. Bias measures the direction. You can have excellent accuracy (low MAPE) and still have significant forecast bias if positive and negative errors average out when you look at individual periods but accumulate when you look at the trend.
For FBA sellers, forecast bias has direct financial consequences. Positive bias (over-forecasting) means you consistently order too much, which inflates storage costs, increases your risk of aged inventory surcharges, and drags down your IPI score. Negative bias (under-forecasting) means you consistently order too little, leading to stockouts that cost you revenue and organic ranking.
Forecast bias formula
Negative bias = under-forecasting (stockout risk)
| Variable | Meaning |
|---|---|
Forecast | Predicted demand for each period |
Actual | Real sales in each period |
n | Number of periods measured. Use at least 6 months to detect meaningful bias patterns. |
Example: detecting over-forecasting bias
You sell a natural cleaning spray at $38 ASP with a 60-day lead time. Your forecasting system has been running for 6 months. Let’s check for bias:
| Month | Forecast | Actual | Error (F−A) | Cumulative |
|---|---|---|---|---|
| 1 | 350 | 320 | +30 | +30 |
| 2 | 345 | 310 | +35 | +65 |
| 3 | 340 | 325 | +15 | +80 |
| 4 | 338 | 295 | +43 | +123 |
| 5 | 335 | 305 | +30 | +153 |
| 6 | 330 | 300 | +30 | +183 |
Bias = 183 / 6 = +30.5 units per month. Average actual demand = 309. Bias % = 30.5 / 309 = +9.9%. Your forecast consistently over-predicts by about 10%.
The financial impact: you are ordering roughly 30 extra units per month at $38 ASP = $1,140/month in excess inventory. Over 6 months, that is $6,840 sitting in FBA warehouses, accumulating storage fees and dragging down your IPI score.
FBA-specific considerations
Forecast bias has outsized financial consequences for FBA sellers because Amazon charges you on multiple dimensions for excess or insufficient inventory. Positive bias (over-forecasting) means consistent excess inventory, which increases monthly storage fees, raises your exposure to aged inventory surcharges after 181 days, and lowers your IPI score. A lower IPI cuts your storage allocation, which forces you to either pay per-unit storage fees or send less inventory next time.
Negative bias (under-forecasting) is just as costly but in different ways. Stockouts kill organic ranking, which compounds the revenue loss for weeks after you restock. Stockout periods also distort future statistical forecasts because the suppressed sales data feeds back into the next forecast cycle, making the bias worse.
Most FBA sellers do not segment their bias analysis. They check accuracy in aggregate and miss that some product categories run +15% bias while others run -10% bias, which cancels out at the portfolio level but represents real category-by-category problems. Always check forecast bias by product type, lead time bucket, and seasonality profile.
Common mistakes
- Confusing forecast bias with forecast accuracy. MAPE measures how far off the forecast is on average; bias measures the direction. A forecast can have low MAPE but persistent bias if errors are consistently in the same direction even when small. Always check both metrics.
- Only checking accuracy monthly instead of tracking cumulative bias. A forecast that is off by +3% each month for 6 months looks fine when you check individual months but represents 18% cumulative over-forecasting. Watch the running sum of forecast errors (RSFE) over a 6-month rolling window to catch drift early.
- Not segmenting bias by category. A portfolio-level bias of 0% can hide significant category-level bias. Apparel might run +12% (over-forecasting) while supplements run -8% (under-forecasting), netting near zero. Slice bias by product category, lead time bucket, and season to find the real problems.