How to weight planner, statistical and AI forecasts in ensemble models

Ensemble forecasting combines human planner judgment, statistical time series methods and machine learning models into a single system. The question merchandisers face is how to weight each input. The answer lies in measuring baseline accuracy by method, assigning weights through backtesting, and tracking Forecast Value Added to identify which steps genuinely improve performance.

What accuracy should you expect from each forecasting method?

With clean product data, channel aware inventory and wholesale commitments visible as a forward signal, style colour size forecasts in the 70 to 85 per cent accuracy range over an eight week horizon are realistic. The baseline varies by method. Fashion products exhibit four conditions that make classical statistical forecasting structurally unreliable: short life cycles, low historical data availability, high demand volatility, and strong seasonal patterns. Time series methods such as ARIMA or exponential smoothing analyze previous sales data and identify noticeable trends or patterns. Yet a new apparel style may sell for a single season with no comparable historical precedent, leaving insufficient data for ARIMA or similar models to learn meaningful patterns.

Adjustments typically led to improvements in bias and accuracy for only just over half of stock keeping units, though there was variation across datasets. The research, published in the International Journal of Forecasting in 2025, analysed around 147,000 forecasts from six studies. Positive adjustments were confirmed as more likely to worsen performance, whilst negative adjustments typically led to improvements, particularly when they were large.

How do you assign initial weights to each model?

The simplest approach is equal weighting. The final prediction is derived by averaging individual model predictions for stability. Equal weight ensembles serve as a useful baseline, but they treat all inputs as equally reliable regardless of performance.

Dynamic inverse RMSE based weighting strategies adaptively assign higher weights to models exhibiting superior validation performance. This approach, published in Nature in December 2025, calculates root mean square error for each model on a validation set, then assigns weights inversely proportional to error. Models with lower error receive higher weight in the final ensemble.

Weighted averaging schemes integrate individual models, with particle swarm optimisation also determining the optimal model weights, ensuring the sum equals one. Optimisation methods search the weight space to minimise ensemble error on historical data.

What does a practical backtesting workflow look like?

Backtesting evaluates forecast performance using data the model has not seen. Backtest each saved forecast version against actual demand known after the period. The workflow requires four steps.

• Set the data cutoff before comparing a forecast with actual demand to prevent judging a forecast against data that was not available when the decision was made.
• Generate forecasts from each method using only data available at the cutoff date.
• Combine forecasts using your chosen weighting scheme.
• Forecast error measures how far the estimate was from actual demand, whilst forecast bias shows whether the process repeatedly forecasts too high or too low.

Archive five forecast versions: system statistical, demand sensing overlay for weeks one to four, planner overrides, marketing promo overlays, and consensus final. This allows you to trace where error enters or exits the process.

How do you measure Forecast Value Added?

Forecast value added analysis is commonly used to measure the improved accuracy and bias achieved by judgmentally modifying system forecasts. Forecast Value Add is a diagnostic framework that measures the incremental impact each step, participant or data input has on forecast accuracy versus a simple baseline.

Compute the incremental change versus the baseline or prior stage and label as value added if error reduced, value neutral if no material change, or value destroyed if error increased. On A SKUs outside promo periods, planner overrides increased WMAPE by 2.8 points on average versus the statistical forecast, whilst during promo periods marketing overlays improved weeks one to four WMAPE by 4.1 points but worsened weeks five to thirteen due to lingering uplift.

Analyse by SKU class, life cycle phase, channel and forecast horizon, because the value of steps varies substantially across segments. A planner adjustment that adds value for core items may destroy value for new launches.

Should weights change over time or stay fixed?

Dynamic weights that update as conditions change outperform static schemes. Most 2026 era planning stacks combine statistical baselines, AI driven pattern recognition for products with thin history, and a planner review layer for exceptions. The optimal weight for each component shifts depending on product lifecycle, seasonality and data availability.

Demand sensing showed positive FVA in weeks one to four for 70 per cent of SKUs, especially in e commerce. This suggests near term forecasts should weight real time signals more heavily, whilst longer horizons may rely more on statistical baselines. Automating FVA ensures that evaluation becomes a continuous discipline rather than a periodic exercise.

What mistakes prevent accurate ensemble forecasting?

The evidence that forecasters made effective use of relevant information not available to the algorithm was weak, as they appeared to respond to irrelevant cues or those of less diagnostic value. Planner adjustments often reflect bias rather than genuine insight.

The same model will underperform a planner's spreadsheet when the inputs are dirty, because the planner is doing implicit data cleaning in her head and the model is not. Clean, consistent SKU codes, accurate timestamps, clear channel mapping and properly tagged returns are prerequisites for reliable ensemble performance.

Set improvement goals from the brand's own backtests rather than a universal accuracy target. Comparing your ensemble to industry benchmarks ignores the specific mix of products, channels and volatility in your business. Measure progress against your own naive baseline, then optimise weights to beat it consistently.