How to weight planner judgement, models and AI forecasts in 2026
Ensemble methods combine statistical baselines with machine learning to reduce forecast error, while Forecast Value Added analysis reveals when human overrides destroy accuracy.

KEY TAKEAWAYS Summary by the editors
- Ensemble forecasting combines multiple models to reduce error by 20 to 50 per cent compared to single methods.
- Forecast Value Added measures whether each adjustment improves or worsens accuracy versus a baseline.
- Override AI predictions only when you have context the model lacks, not when output conflicts with intuition.
- Weighted blends of statistical and machine learning models outperform enterprise-wide winner-takes-all strategies.
Ensemble forecasting combines multiple models to overcome uncertainties in data, parameters and models and decrease the risks of relying on a single best model. In fashion merchandising, this means ensemble approaches combine the outputs of multiple models so the outcome is not controlled by any single model's blind spot. The discipline has become standard practice as 75 per cent of fashion executives now prioritise AI for inventory optimisation and demand forecasting.
What is ensemble forecasting and why does it matter for fashion retail?
Evidence from large-scale forecasting competitions and field deployments consistently shows that combining models yields more accurate and robust predictions than selecting a single winner. Using retail demand data from the sports retail industry, ensemble approaches consistently outperformed individual forecasting models and improved forecasting accuracy across different seasonal and demand scenarios.
The practical implementation varies by product category. For a stable, fast-moving SKU with three years of clean history and predictable seasonality, exponential smoothing or ARIMA tuned per series remains hard to beat, costs almost nothing to run, and is fully explainable. If a vendor claims AI will dramatically improve forecasts on your top 500 steady sellers, be sceptical. Mature programmes keep statistical baselines in the ensemble, use machine learning where backtesting shows it wins, and measure per SKU-location rather than declaring one method the enterprise-wide winner.
Weighted blends offer a middle path. For production, the weighted ensemble combining 70 per cent LightGBM and 30 per cent Prophet is highly recommended to combine predictive accuracy with robust trend and seasonality modelling and uncertainty quantification.
How do you measure Forecast Value Added?
Forecast Value Added is a methodology that measures whether each step in the forecasting process, including models, AI agents or human overrides, improves or worsens forecast accuracy and business outcomes. Starting with a naive forecast, FVA analysis seeks to determine whether each subsequent step in the process improves upon the prior steps.
The calculation is straightforward. FVA is the percentage improvement in forecasting performance achieved by each added step. If the departments made the forecast more accurate, they added positive FVA. If they made it less accurate, they added negative FVA.
A representative example shows the pattern. Averaged across all products, naive random walk achieved forecast accuracy of 60 per cent. The company's statistical forecast delivered five percentage points of improvement, but management review and adjustment delivered negative value.
| Process step | Forecast accuracy | FVA vs naive |
|---|---|---|
| Naive forecast | 60% | , |
| Statistical forecast | 65% | +5% |
| Adjusted forecast | 62% | +2% |
Any activity that fails to deliver positive FVA, meaning it fails to improve forecast quality, is considered process waste. Track weighted mean absolute percentage error, bias and forecast value added by SKU segment, and review override quality monthly so accuracy gains stay visible to planning teams.
When should planners override AI recommendations?
In engagements across retail and manufacturing between 2023 and 2025, planner overrides reduced forecast accuracy in roughly 40 per cent of cases, not because planners lack judgement, but because override workflows lack a closed-loop feedback mechanism. Operators override 40 per cent of their AI predictions based on gut feel. When auditing their closed deals three months later, the AI was right 70 per cent of the time.
The discipline required to override responsibly centres on documentation. Override is appropriate when recommendations conflict with domain expertise, violate operational constraints, or are affected by seasonal shifts not yet reflected in the data. Override AI predictions only when you have new information the model does not have access to, such as a champion leaving the company, budget being frozen or procurement adding a new approval layer.
The operators who consistently beat their forecast override less than 15 per cent of AI predictions. Planners who receive regular FVA feedback reduce override frequency by roughly a third within two quarters, and the overrides that remain tend to be the ones that genuinely add value.
Three practices protect accuracy:
- Require every material override to document its evidence source, confidence level and review horizon.
- Compare override cohorts against the untouched model baseline rather than only against final topline revenue.
- Every time you override the AI, log it.
How do blending strategies work in practice?
Leading fashion retailers have moved beyond pilot projects to production deployment. Zara entered a multi-year partnership with IBM Watson to embed real-time demand forecasting across its European supply chain, while H&M teamed up with Google Cloud AI to power visual search on its mobile app. Inditex invested €1.8 billion in tech and logistics for 2025 to 2026, with AI-powered demand forecasting now driving 85 per cent of initial production allocation decisions.
AI-based demand forecasting can decrease the forecasting errors by 20 to 50 per cent, allowing brands to significantly enhance product availability whilst minimising excess inventory. AI-powered forecasting has been so effective that 85 per cent of inventory sells at full price, compared to an industry average of 60 to 70 per cent.
The technical architecture matters less than governance. Teams can view model drivers, adjust parameters or override forecasts when needed. In-season retrending automatically updates projections after strong or weak sales weeks, giving planners agility without losing control. The most accurate forecast in retail history is worthless if planners override it back to last year's buy, and they will, rationally, if the system is a black box that once embarrassed them in a category review.
What are the common blending mistakes to avoid?
The underlying problem is that each human touch point subjects the forecast to the interests of the reviewers, and these interests may not align with creating an accurate, unbiased forecast. Planners override forecasts they cannot interrogate, and blanket overrides usually destroy accuracy.
Disagreeing with the number because it feels wrong is not an override, it is a mood. Require the reason to be written before the action is taken, so the override can be reviewed later as a decision rather than reconstructed from the outcome. Forecasting problems are often blamed on model quality when the larger issue is override behaviour. Forecast override discipline matters because predictive commerce should not only predict demand, it should govern when human instinct gets to overrule evidence.
The organisational structure matters. Reporting into analytics tends to turn the job into forecast maintenance, and the override authority quietly evaporates because nobody in that chain owns the margin. Successful programmes place override authority with the team accountable for sell-through and gross margin outcomes.
How should teams implement ensemble forecasting without disrupting operations?
This ensemble-first strategy is evidence-based, operationally tractable and resilient to regime shifts typical in manufacturing. Retailers still running manual demand forecasting can get structured migration paths that limit disruption as they move towards capable infrastructure. Those already on statistical models get external signal integration, ensemble architecture and agentic AI capabilities built in.
Start with a narrow scope. Analyse by SKU class (A, B or C), life-cycle phase, channel and forecast horizon (for example weeks one to four versus five to 13), because the value of steps varies substantially across segments. Run the ensemble in parallel with existing processes for one season, measure FVA at each stage, then expand to categories where the evidence is strongest.
The best results come from blending AI insights with the expertise of your merchandising team. This balance ensures AI supports, not replaces, the personal touch that makes your brand unique. The future belongs to those who treat analytics as fundamental, blending human taste with AI-powered precision in every assortment and campaign.
Frequently asked questions
What percentage of overrides typically reduce forecast accuracy?
In engagements across retail and manufacturing between 2023 and 2025, planner overrides reduced forecast accuracy in roughly 40 per cent of cases. Operators who consistently beat their forecast override less than 15 per cent of AI predictions.
How much can ensemble forecasting improve accuracy?
AI-based demand forecasting can decrease forecasting errors by 20 to 50 per cent compared to traditional methods. Ensemble approaches consistently outperformed individual forecasting models across different seasonal and demand scenarios.
Should fashion retailers replace statistical models with AI?
Mature programmes keep statistical baselines in the ensemble, use machine learning where backtesting shows it wins, and measure per SKU-location rather than declaring one method the enterprise-wide winner. For stable, fast-moving SKUs with clean history, exponential smoothing or ARIMA tuned per series remains hard to beat.
Sources
- MDPI: Retail Sales Forecasting Using Tree-Based Machine Learning Models
- arXiv: Forecasting: theory and practice
- McKinsey via FashionBI: AI in Fashion Is Moving from Marketing to Merchandising
- Retail News AI: Zara AI Supply Chain
- Netguru: AI demand forecasting implementation
- Supply Chain Management Review: The future of forecast value add
