Trent TurnerIntegrated Adaptive Demand Forecast
Combining seasonality and recent demand signals into an out-of-sample weekly forecasting system
- Question
- Can an interpretable adaptive seasonal model outperform simple benchmarks on genuinely unseen demand?
- Method
- Frozen adaptive model developed on 2023–2024; evaluated on untouched 2025.
- Evidence
- 2025 holdout MAE 1.97 vs. 2.43; RMSE 3.23 vs. 3.70, both vs. recent 4-week benchmark.
- Decision
- Use as a provisional weekly acquisition-demand benchmark; check exact 2026 alignment before stronger claims.
Business Question
Can an interpretable adaptive seasonal model outperform simple benchmarks on genuinely unseen demand?
Why It Matters
Fixed seasonality misses annual level changes; flat recent averages miss seasonal direction. Integrating shape, relative level, and local momentum addresses both weaknesses.
Data / Method
2023–2024 development used chronological rolling-origin selection. The selected architecture and score scale were frozen before evaluation on the untouched 2025 holdout. The final model uses no weather or economic inputs.
Analysis
- Recurring seasonal curve
- Normalize completed demand: (New Customers + 1) ÷ S(w)
- Estimate six-week current level and slope
- Reapply seasonal curve
- Expected weekly New Customers
Method detail
S(w) is the retained two-harmonic seasonal factor. L is the documented six-week endpoint trend of completed normalized observations. Each forecast excludes the target week.
- 2023–2024 development
- Rolling-origin model selection
- Model and score scale frozen
- Untouched 2025 holdout
- 2026 calendar-alignment sensitivity check
Results
Untouched 2025 holdout: absolute error
MAE · weekly New Customers · lower is better
Untouched 2025 holdout: large-error sensitivity
RMSE · weekly New Customers · lower is better
Method detail
2025 holdout supporting diagnostics: predictive R² = 0.778; actual–predicted correlation = 0.891. Forecast error relative to the recent four-week benchmark is the primary evaluation.
Improvements are the source-reported values calculated before rounding; displayed two-decimal errors do not reproduce those percentages exactly.
The supplied sources report holdout summaries but do not contain all 52 actual/predicted 2025 rows. A weekly holdout line chart cannot be reconstructed accurately from these summaries.
Interpretation
Provisionally validated: the model generalized usefully to one untouched annual holdout. The later 2026 check is a separate, approximate alignment-sensitivity analysis.
Same week number. Different days.
- 4 days from pivot W01
- 3 days from pivot W02
- Approximate frozen-model W01 = 4/7 × count A + 3/7 × count B
Assumes uniform within-week timing. Exact validation needs daily Customer Since dates or daily counts.
Calendar definitions are part of the model specification, not a formatting detail.
2026 calendar-alignment sensitivity check · W01–W36
MAE · weekly New Customers · approximate calendar alignment
The model’s advantage persisted under this additional alignment check, but the margin was smaller: approximately 1.56 versus 1.62 New Customers per week in MAE. This is not a repeat of the untouched 2025 holdout. The next checkpoint is exact daily re-bucketing of 2026 New Customers, followed by a frozen-model rerun without retraining.
Limitations
One annual holdout supports provisional rather than full validation. The 2026 approximation assumes uniform timing within each pivot week; exact daily counts are required for a definitive comparison.
How this fits into the forecasting system
Flagship integration: seasonality, adaptive demand level, and local momentum form the final architecture. Weather and economic candidates were rejected. The next layer translates forecasts into a Demand Score.
Next: Demand Score & Decision Support