How to choose a demand forecasting tool: a checklist for fashion brands

KEY TAKEAWAYS Summary by the editors

1. A demand forecasting tool for fashion must handle new products without history, short life cycles, size curves and sales capped by stockouts, not only continuous replenishment items.
2. The most reliable way to compare tools is a back-test on your own past seasons, measured at the level where buying, allocation and re-order decisions are taken.
3. Accuracy should be reported with metrics that cope with zero sales, such as WAPE or MASE, plus forecast bias, rather than MAPE alone.
4. Data readiness, including consistent product attributes and stockout-corrected sales history, often limits results more than the choice of algorithm.
5. A forecast only creates value when it changes a decision, so integration with planning, allocation and ordering processes belongs in the selection criteria.

Choose a demand forecasting tool by testing it on your own past seasons, at the level where you actually make decisions, and by checking how it handles fashion-specific problems: new products, short life cycles, sizes and stockouts. Data readiness, transparency, integration and total cost matter as much as headline accuracy. This checklist helps structure the evaluation without favouring any vendor.

Why is choosing a forecasting tool different in fashion?

Many forecasting tools were built for grocery or consumer goods, where products sell continuously for years. Fashion is different: a large share of the range is new each season, products live for weeks or months, demand is split across sizes and colours, and sales are often capped by the stock that was bought. A tool that excels at replenishing basics may struggle with a seasonal dress. Clarifying which of these problems matters most to your business is the first step.

Which forecasting use cases should the tool cover?

Read also

How do you measure forecast accuracy in fashion? MAPE, WAPE, bias and MASE

How should you test forecast accuracy before buying?

Ask each shortlisted vendor to forecast one or two past seasons that their model has not seen, using only the data that would have been available at the time. Compare the results with your current method and with a simple naive benchmark. Hyndman and Koehler, in their widely cited paper on forecast accuracy measures, recommend scaled errors such as MASE, where a value below one means a forecast beats a naive method, partly because percentage errors such as MAPE become undefined when actual sales are zero.

Make sure every vendor is measured the same way: same products, same horizon, same level of detail, same treatment of stockouts and promotions. Accuracy at brand-month level says little about a size run in a single store.

Ask also for bias, the average tendency to forecast too high or too low, broken down by category and by new versus continuing products. A tool with good average error that consistently over-forecasts will still build excess stock. And check whether the vendor's forecasts were produced automatically or tuned by their data scientists for the test, since only the former reflects what your planners will receive in daily use.

What questions should be on the checklist?

• New products: How does the tool forecast items with no history? Which attributes, images or analogues does it use?
• Censored demand: Does it correct historical sales for stockouts and limited distribution?
• Granularity: Can it forecast at style, colour, size, store and week, and reconcile those levels?
• External signals: Can it use weather, events, search or marketing calendars, and does it show their measured effect?
• Explainability: Can planners see why a forecast changed, and which drivers mattered?
• Overrides: Can planners adjust forecasts, and does the tool measure whether overrides helped?
• Uncertainty: Does it provide ranges or probabilities, not only a single number?
• Integration: How does the forecast reach buying, allocation, ERP and order systems?
• Data ownership: Who owns trained models and outputs, and can data be exported if you leave?
• Cost: What are licence, implementation, data preparation and ongoing support costs over three years?

How much does data readiness matter?

More than most selection processes assume. A model that matches new products to past ones needs attributes that were tagged consistently across seasons. A model that learns demand needs sales corrected for stockouts: researchers working with the online retailer Rue La La first used machine learning to estimate lost sales on sold-out items before forecasting new products, as Harvard Business School's Working Knowledge describes. Before a pilot, audit attribute completeness, stock history, price and promotion records, and store master data. Gaps found here will limit any tool.

How do you judge the return on investment?

Forecast accuracy is a means, not the goal. The business case comes from decisions that change: fewer units left at season end, fewer stockouts on bestsellers, better allocation. Published evidence suggests gains are real but incremental. In Zara's field experiment, published in Operations Research in 2015, a system combining updated forecasts with stock allocation increased average season sales by about 2% and reduced end-of-season unsold units by about 4%. Research on new fashion products using the VISUELLE dataset found that adding Google Trends signals improved weighted error by 1.5%. Be sceptical of vendor claims that are an order of magnitude larger without a comparable test behind them.

Translate measured error reduction into inventory and margin terms with your finance team, using your own cost of excess stock and of lost sales.

Read also

AI demand forecasting in fashion: how it works and where it fails

How should the selection process run?

1. Define the priority use cases and decision levels before contacting vendors.
2. Audit and prepare data, and agree a common test dataset.
3. Shortlist tools that can demonstrate fashion-specific capabilities, especially for new products.
4. Run back-tests on the same seasons and metrics, including WAPE or MASE and bias.
5. Pilot the best candidate live on one category for a full season, alongside the current method.
6. Decide based on measured decision outcomes, integration effort and three-year cost.

Finally, avoid the mistakes that most often undermine selection projects:

• Comparing vendor demos on their sample data instead of your own history.
• Accepting a single headline accuracy figure without knowing level, horizon or metric.
• Underestimating the effort of data preparation and process change.
• Selecting a tool before deciding who will own and act on the forecast.

Frequently asked questions

What should I look for in demand forecasting software for fashion?

Look for strong handling of new products without history, forecasts at style, colour, size and store level, correction for stockouts, uncertainty ranges and transparent drivers. Above all, insist on a back-test on your own past seasons compared with your current method.

How do you compare forecasting tools fairly?

Give every vendor the same historical dataset, ask them to forecast seasons their models have not seen, and measure results with the same metrics at the same decision level. Include your current method and a naive benchmark in the comparison.

How long does it take to implement a forecasting tool?

Timelines vary with data quality, integration needs and scope. Data preparation and process change often take longer than technical installation, so a single-category pilot over a full season is a realistic way to validate results before scaling.

What ROI can a fashion brand expect from AI forecasting?

Published field experiments show incremental gains rather than dramatic ones, for example about 2% more season sales and about 4% fewer unsold units in Zara's 2012 allocation experiment. Expected ROI should be estimated from your own back-test results and costs.

GuideThe complete guide to AI in fashion merchandising and buyingRead the complete guide