How to measure Forecast Value Added and prove your AI beats baseline

What is Forecast Value Added and why does it matter for fashion merchandising?

Forecast Value Added is a tool for evaluating the performance of each step and contributor in the demand forecasting process, measured against the step before it or a simple benchmark. In fashion merchandising, where typical weighted MAPE runs 25 to 40 percent and style rotation, trend sensitivity and size or colour complexity all work against accuracy, FVA provides the evidence to justify investment in AI models or identify which manual adjustments actually destroy value.

FVA aims to eliminate waste and increase demand forecasting accuracy by encouraging and evaluating inputs from multiple departments, providing companies actionable data on the overrides that make the forecast worse. More than 35 percent of fashion executives already use generative AI in daily operations and rank AI as the industry's biggest opportunity, according to the State of Fashion 2026 from McKinsey and Business of Fashion. Yet proving the model improves on simpler methods remains the barrier to wider adoption.

How do you establish a naive baseline forecast?

The naive baseline is the no-skill threshold every forecast must beat. A seasonal naive forecast defines the no-skill threshold at a MASE of 1.00, and any vendor claim should be quoted against that baseline rather than against no forecast at all. For fashion SKUs, two naive methods serve as standard benchmarks.

The naive forecast uses the most recent observed value as the next forecast. If a jacket sold 120 units last week, the naive forecast for next week is 120 units. The naive forecast can attain a MAPE below 1 percent in highly persistent aggregate retail sales where a random-walk forecast is hard to beat. The seasonal naive forecast uses the value from the same period in the prior cycle. For a spring dress launching in week 12, the seasonal naive forecast uses sales from week 12 of the previous year's comparable SKU.

The naive baseline performs quite competitively on the one-week forecast, and a more detailed analysis reveals that especially for high-selling articles using the naive forecast poses a strong baseline. This is why FVA requires comparison: if your AI model achieves 15 percent MAPE but the naive forecast achieves 12 percent, the AI has added negative value.

Which accuracy metrics should you use to calculate FVA?

FVA can be calculated using any forecast accuracy metric, but fashion merchandisers typically track three.

• MAPE (Mean Absolute Percentage Error): MAPE measures the average percentage error between forecast and actual demand, calculated as the sum of absolute value of actual minus forecast divided by actual, times 100. It is intuitive and widely used, but it is distorted by low-volume items and can mask large absolute errors in high-volume SKUs.
• WAPE (Weighted Absolute Percentage Error): WAPE is the total absolute error divided by total actual demand across all items; unlike MAPE, it weights each SKU by its volume, so high-selling products have more influence on the score, making it more representative of real business impact.
• Forecast Bias: Bias is the sum of forecast minus actual divided by sum of actual, also tracked as percent bias and signed mean percentage error. Positive bias indicates over-forecasting, negative bias indicates under-forecasting.

Despite its deficiencies, MAPE remains one of the most used accuracy metrics; in calculating the values of FVA, individual SKUs are equally weighted. Choose WAPE if your assortment has a long tail of low-volume SKUs, or MAPE if you need a metric non-technical stakeholders understand immediately.

How do you calculate FVA step by step?

An FVA analysis involves defining the contributors and the order in which they will contribute, generating a statistical forecast and then a naive one, then calculating the FVA for each contributor at each step of the process. The calculation compares accuracy at each layer against the previous layer and against the baseline.

For each forecasting step, calculate the chosen accuracy metric (typically MAPE or WAPE) using the same set of actual sales. FVA is the percentage improvement in forecasting performance achieved by each added step. A step adds value when the forecast after it is more accurate than the forecast before it, and destroys value when it is less accurate; FVA is reported as a stairstep where each layer's value is the difference between its error and the error of the layer before it.

The formula for FVA in percentage points is: FVA = Accuracy of previous step minus Accuracy of current step, where accuracy is expressed as 100 minus MAPE. Alternatively, express FVA directly as the change in error: FVA = MAPE of previous step minus MAPE of current step. A positive FVA means the new step reduced error.

What does a worked FVA example look like for fashion SKUs?

Consider a women's knitwear SKU across 50 stores over 12 weeks. Actual sales totalled 4,800 units. Four forecasting approaches were tested, each generating a forecast before the period began. The table below shows the resulting MAPE and FVA for each step.

The statistical model improved MAPE by 7 percentage points compared to the naive baseline, delivering positive FVA. The AI model added a further 6 points of improvement. The planner override, however, increased MAPE by 3 points, destroying value relative to the AI model. The final forecast of record still beats the naive baseline by 10 points, but the manual step should be examined: any activity that fails to deliver positive FVA is considered process waste.

In absolute terms, a 13-point MAPE reduction from 35 percent to 22 percent on 4,800 units means approximately 624 fewer units mis-forecast per period. Fashion retailers that have implemented AI demand forecasting report reductions in forecast error of 30 to 50 percent compared to traditional methods; for a retailer with 20 million euros in annual inventory purchases, a 35 percent reduction in forecast error translates directly into millions in reduced excess inventory.

How do you turn FVA analysis into action?

Each layer's forecast is scored against the same actuals over the same period, and the recommendations that follow from that work are that each adjustment carry a recorded reason and that adjustments be measured against the forecast they replaced. Track FVA at each step for multiple periods to separate signal from noise.

An FVA analysis conducted for many periods enables identifying stages and persons who generate added value; for them, FVA is greater than zero. If your AI model consistently delivers positive FVA of 5 to 10 percentage points against the naive baseline, you have quantified its value. If manual overrides consistently add negative FVA, either retrain the planners, restrict override authority or eliminate the step entirely.

Granularity matters more than most teams realise; a MAPE calculated at the product-category-month level will always look better than the same metric calculated at SKU-store-week level, but supply chain decisions happen at SKU-store-week; if you are reporting accuracy at a level that is too aggregated, you are masking the errors that actually cause stockouts and overstock. Run FVA at the decision grain: SKU-location-week for allocation, style-colour-week for initial buys.

Grocery and CPG retailers typically target a MAPE of 10 to 20 percent at the product-week level; fashion and apparel retailers often treat 30 to 40 percent as acceptable at the same granularity. Use these benchmarks to set realistic expectations, but always measure improvement relative to your own naive baseline, not an industry average. Track MAPE for magnitude, bias for direction and Forecast Value Added to prove your process beats a naive baseline.