.svg)
Download Brouchure
Click the button to download the PDF file.
Download
%202.svg)
Oops! Something went wrong while submitting the form.
Demand forecasting is the backbone of supply chain management, financial planning, and customer satisfaction. Accurate forecasts reduce the costs and waste tied to excess inventory while preventing the lost sales and customer churn that come with stock-outs. These two objectives are inherently at odds, which is why the starting point must be a measurement framework that can objectively answer: "How accurate are our forecasts, and what impact does the error have on our business?"
There are several metrics for measuring forecast accuracy, each offering a different lens on error. In practice, relying on a single metric is rare — combining metrics to match the business objective is the standard approach.
MAE takes the absolute difference between forecasted and actual values, then averages the result. Because the output is expressed in the same unit as the original data (units, tons, etc.), it's easy to interpret at a glance. The limitation is that it doesn't allow apples-to-apples comparison across product lines with different sales volumes. An MAE of 100 units means something entirely different for a product selling 10,000 units per month versus one selling 200.
MSE squares each error before averaging, which imposes a heavier penalty on large deviations. This makes it useful in scenarios where extreme forecast misses carry disproportionate consequences. On the downside, the output is in squared units (units²), making intuitive interpretation difficult, and a single outlier can disproportionately skew the entire metric.
RMSE applies a square root to MSE, restoring the output to the original scale. It retains MSE's sensitivity to large errors while keeping the result readable in real-world units, which is why it's frequently used for model-to-model performance comparison. Like MSE, however, it remains vulnerable to outliers and does not support direct comparison across datasets with different scales.
Unlike MAPE, MPE does not take the absolute value of errors, so positive errors (over-forecasting) and negative errors (under-forecasting) cancel each other out. This makes it well suited for detecting forecast bias — whether a model consistently skews high or low. Monitoring a bias metric like MPE alongside an accuracy metric like MAPE can reveal situations where overall accuracy looks acceptable but a persistent directional skew is silently creating inventory problems.
Mean Absolute Percentage Error (MAPE) is one of the most widely used accuracy metrics in demand forecasting practice. That said, it requires careful interpretation when actual values are zero or very small, and it's typically used alongside complementary metrics.
Xₜ represents actual demand at time t, Fₜ is the forecasted value, and n is the total number of data points. The formula strips out the direction of error (over or under), converts each error to a percentage of the actual value, and averages across all periods to produce a single percentage figure. The primary reason MAPE became so widely adopted in practice is that it enables performance comparison across product lines with vastly different sales volumes on a common percentage scale.
For items with very low actual values, however, the error rate can be inflated to unrealistic levels. This is why practitioners typically pair MAPE with complementary metrics such as WMAPE.
MAPE's greatest strength is communicative clarity. A report stating "MAE is 500 units" is hard to interpret without knowing total sales volume, but "MAPE is 5%" immediately conveys that forecasts deviate from actual demand by an average of 5%. It requires no technical background to understand, which makes it function as a common language for setting company-wide accuracy targets. And because it's a percentage metric, it enables direct comparison of forecast performance across products with entirely different sales scales.
MAPE's formula is convenient, but it carries two significant weaknesses.
First, it breaks down when actual values are zero or very small. If actual sales are zero, the denominator becomes zero and the calculation is undefined. If actual values are as low as 1 or 2 units, even a minor forecast miss produces error rates in the hundreds of percent. This is the core reason MAPE is unreliable as a standalone metric for slow-movers and intermittent-demand items.
Second, MAPE has a built-in asymmetry that favors under-forecasting. When actual demand is 100, forecasting 0 yields a maximum MAPE of 100%, but forecasting 500 produces a 400% error with no upper bound. This asymmetry means that models optimized to minimize MAPE tend to generate forecasts that systematically lean below actual demand. This is particularly dangerous in industries where stock-outs are more costly than excess inventory — think fresh food or essential pharmaceuticals.
Weighted MAPE (WMAPE) replaces the simple average of individual error rates with the sum of absolute errors divided by the sum of actual values.
Because total sales volume serves as the denominator, the calculation remains valid even when individual actual values are zero. This naturally eliminates the distortion caused by inflated error rates on low-volume items, and it assigns greater weight to errors on high-revenue products — better reflecting real operational impact.
Symmetric MAPE (sMAPE) normalizes the error by the average of the actual and forecasted values, addressing MAPE's directional asymmetry.
The result is bounded between 0% and 200%, applying balanced penalties to both over- and under-forecasting. In environments with high demand volatility or frequent near-zero values, sMAPE offers a more stable alternative to MAPE.
"Good accuracy" in demand forecasting isn't a fixed number — it varies with industry characteristics. Some literature (Lewis, 1982, among others) classifies MAPE below 10% as "highly accurate," 10–20% as "good," 20–50% as "reasonable," and above 50% as "inaccurate," but these are reference thresholds, not absolute standards. Achievable accuracy levels differ significantly based on industry dynamics, product type, forecast granularity, and time horizon.
For example, FMCG (fast-moving consumer goods) products benefit from high repeat-purchase rates and typically report forecast accuracy (1 – MAPE) in the 80–95% range. General CPG (consumer packaged goods) typically lands at 70–85% due to the compounding effects of promotions and seasonality. Consumer electronics, with their rapid innovation cycles, tend to fall in the 65–80% range, while fashion and apparel — highly sensitive to trend shifts — often sit at 60–75%. Promotional and event-driven items, where demand is artificially stimulated, generally achieve 60–70%. These figures can vary across sources, and even within the same industry, results change substantially depending on forecast granularity (daily vs. weekly vs. monthly) and aggregation level (SKU-level vs. category-level).
When setting internal accuracy targets, it's most effective to use external benchmarks as reference points while grounding goals in your own historical forecast performance and the track record of comparable product lines.
For practitioners just getting started with demand forecasting, Excel remains the most accessible tool. Start by structuring a source data table containing date, product code (SKU), actual values, and forecast values, then calculate absolute error using ABS(Actual – Forecast). When computing individual APE, it's standard practice to exclude data points where the actual value is zero. If you set APE to zero when the actual is zero but the forecast is not, it creates a false impression of perfect accuracy for that period.
To address this limitation, it's recommended to calculate MAPE using the AVERAGE function alongside WMAPE using the SUM(Absolute Error) / SUM(Actual) formula as a parallel check.
Aggregate results using pivot tables by product category and owner, and apply conditional formatting to highlight items exceeding threshold values for immediate identification in a dashboard view.
An Excel-based framework is a solid starting point, but once SKU counts grow into the hundreds or thousands and external variables need to be incorporated, an automated solution becomes necessary.
ImpactiveAI's Deepflow is an AI demand forecasting SaaS solution designed for exactly this stage of maturity. It automatically selects the optimal model for each SKU's sales pattern from a library of over 224 machine learning and deep learning models, generating forecasts backed by 72 patented technologies. LLM-powered analytical reports automatically produce historical sales trend analysis, forecast rationale, and department-specific action plans, reducing the time practitioners spend interpreting error causes and formulating responses.
The goal of managing forecast error is not to drive it to zero — it's to understand where error comes from and build an organizational framework where teams align around a shared set of numbers. Starting with MAPE, progressively adding WMAPE and bias analysis, and scaling into AI solutions as needs evolve is the most pragmatic path forward.
.svg)
.svg){kind=icon}
.svg)
