Ensemble forecasting: combining human and AI predictions for accuracy

Forecast Value Added measures whether each step in the forecasting process, whether models, AI agents or human overrides, improves or worsens forecast accuracy and business outcomes. In fashion merchandising, where demand volatility and short product lifecycles make forecasting inherently difficult, an ensemble approach that weights multiple AI model outputs alongside human planner input offers a structured path to higher accuracy. The method requires discipline: every contributor, whether algorithmic or human, must be evaluated against a baseline to determine whether it adds or destroys forecasting value.

What is ensemble forecasting and why does it matter for fashion?

Ensemble regression methods are widely used to improve prediction accuracy by combining multiple regression models, especially when dealing with continuous numerical targets. The ensemble algorithm trains diverse models on the same datasets and then combines their outcomes for a more accurate final prediction, and the underlying principle is that multiple weak learning models, when strategically combined, can form a stronger, more reliable predictor. For fashion retailers, this means running parallel forecasts using different model types, such as time series models for stable categories, attribute based models for new SKUs without sales history, and social signal models for trend sensitive items, then weighting each model's contribution proportionally to its accuracy.

AI powered forecasting use cases, as noted in The State of Fashion 2025 by McKinsey and Company, have the potential to reduce inventory by 5 to 15 per cent and to achieve a 15 to 25 per cent improvement on stockouts. Boston Consulting Group's retail research found that AI driven demand forecasting has reduced overstock by 20 to 50 per cent at brands with end to end deployment. Fashion faces wider accuracy gaps than most retail verticals because of deep seasonality, frequent new product launches and thin or absent sales history, which makes ensemble methods particularly valuable.

How do you weight multiple AI models in an ensemble?

Weighted average ensembles assume that some models in the ensemble have more skill than others and give them more contribution when making predictions, and weighted average or weighted sum ensemble is an ensemble machine learning approach that combines the predictions from multiple models, where the contribution of each model is weighted proportionally to its capability or skill. The most common weighting approach uses each model's performance on a holdout validation dataset. A model that achieves 10 per cent mean absolute percentage error receives higher weight than one that achieves 15 per cent error.

By using an RRMSE based weighting function, methods give more importance to models that demonstrate higher accuracy, thereby enhancing the overall prediction quality, and by using RRMSE as the basis for weighting, the proposed method prioritises models that exhibit lower relative errors, thereby enhancing the overall prediction quality of the ensemble. In practice, merchandising teams run multiple model architectures in parallel: an LSTM model trained on historical sales data, an XGBoost model incorporating promotional calendars and weather signals, and an attribute similarity model for cold start SKUs. Each model generates a forecast for the same SKU and period, and the final ensemble forecast is the weighted average of these predictions.

Where does human planner input fit in the ensemble?

Based on a field experiment involving AIML driven demand forecasts approximately 1,888 stock keeping units in the retail industry, findings show that human intervention complements AIML driven forecasts most effectively (augmentation) in predictions with long time horizons and low uncertainty. In environments characterised by high uncertainty and short term horizons, automated AI forecasting performs best, and in contrast, human augmentation improves predictive accuracy in more stable environments with long term horizons. This finding matters for fashion merchandising because it defines when to solicit planner overrides and when to automate completely.

The ideal conditions for human AI collaboration are predictions with moderate uncertainty and medium term time horizons, and in these situations, combining human judgment with AI analysis can lead to the most accurate forecasts. For seasonal buys six months ahead, where demand is moderately stable and planners have visibility into marketing calendars and wholesale commitments, human adjustments tend to add value. For volatile, trend driven categories with two week lead times, algorithmic forecasts perform better without manual intervention.

How do you measure Forecast Value Added at each step?

Gilliland suggested that the forecast value added of each step in the forecasting process should be measured when progressing from simple to more complex procedures, and simply put, FVA is the percentage improvement in forecasting performance achieved by each added step. The benchmark that matters is a seasonal naive forecast, which is last season's number repeated, and 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. Measuring FVA requires three inputs: the baseline forecast, the forecast after each model or human contribution, and the actual outcome once the period closes.

Calculate the FVA for each contributor by comparing the accuracy of the statistical forecast before and after that contributor's input, and in turn, the accuracy of the statistical forecast is contrasted with that of the simple benchmark forecast, and contributions that enhance forecast accuracy receive positive FVA, while those that diminish accuracy receive negative FVA. If the baseline seasonal naive achieves 25 per cent MAPE, the first AI model achieves 18 per cent, and a second model brings it to 16 per cent, the first model contributes 7 percentage points of positive FVA and the second adds 2 points. If a planner override then worsens accuracy to 19 per cent, that adjustment contributes negative 3 points of FVA.

What ensemble structure works best for fashion merchandising?

A practical ensemble structure for fashion demand planning begins with a seasonal naive baseline, adds one or more machine learning models weighted by validation accuracy, and conditionally incorporates human planner input based on product and time horizon characteristics. The baseline forecast establishes the threshold that every subsequent step must beat. For fashion categories, the seasonal naive forecast predicts that next spring's demand for a product equals last spring's actual sales.

The first layer of the ensemble typically includes at least two AI models with different strengths. ML demand models use colour, fabric, silhouette and price tier to identify comparable styles from prior seasons and construct a demand curve based on attribute similarity, and social signal integration and consumer sentiment analysis sharpen that early read further. A time series model such as LSTM or Prophet handles SKUs with sufficient history, while an attribute based model addresses cold start scenarios for new styles. Each model's weight reflects its recent accuracy on similar SKUs.

The final layer of the ensemble is the planner adjustment, treated as another weighted input rather than an absolute override. 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, and to identify where such problems are occurring, Forecast Value Added analysis is an increasingly popular approach among practitioners. Planners receive the weighted AI ensemble forecast and can propose adjustments, but those adjustments are tracked, measured and weighted according to past FVA performance, not implemented blindly.

How do you implement FVA tracking in a merchandising workflow?

Implementation requires a structured process for recording forecasts at each step, capturing actuals once the period closes, and calculating accuracy metrics across all SKUs. Merchandising teams establish a data pipeline that logs the baseline forecast, each model's output, the weighted ensemble prediction, any planner adjustments and the final published forecast. When actual sales data becomes available, the system calculates MAPE or mean absolute scaled error for each step and attributes FVA to each contributor.

Any activity that fails to deliver positive FVA (that is, fails to improve forecast quality) is considered process waste, and starting with a naive forecast, FVA analysis seeks to determine whether each subsequent step in the process improves upon the prior steps. A monthly FVA review identifies which models consistently add value, which planners improve accuracy and which adjustment patterns harm performance. Models with sustained negative FVA are removed from the ensemble or retrained. Planners whose overrides consistently worsen accuracy receive feedback and coaching, or their weights in the ensemble are reduced.

The FVA framework also surfaces segmentation opportunities. Some planners may add significant value forecasting core replenishment items but destroy value on trend items, suggesting a workflow where their input is solicited only for stable categories. Some AI models may excel at forecasting full price demand but perform poorly during promotional periods, indicating a need for separate promotional models or adjusted weights during sale windows.

Across 14 enterprise implementations tracked between 2023 and mid 2025, AI augmented demand forecasting reduced mean absolute percentage error by 34 per cent compared to legacy statistical methods (exponential smoothing, ARIMA). For fashion retailers, the ensemble approach delivers the highest returns when it combines diverse model types, weights them by validated accuracy, integrates human judgment selectively and measures every contribution against a disclosed baseline. The discipline is in the measurement: without Forecast Value Added tracking, ensemble forecasting devolves into unweighted model averaging or unchecked planner overrides, both of which forfeit accuracy gains.