Forecasting Methodology
📌
Naïve Baseline
Ŷₜ₊₁ = Yₜ
(Last observed value)
Simple benchmark. No trend or pattern recognition. Used as the baseline to measure improvement from statistical models. Performs poorly with seasonal or trending demand.
Avg MAPE: ~14.1%
📉
Simple Moving Average (3-month)
Ŷₜ₊₁ = (Yₜ + Yₜ₋₁ + Yₜ₋₂) / 3
Smooths out short-term fluctuations by averaging recent periods. Better than Naïve but slow to respond to trend changes. Window = 3 months chosen via cross-validation.
Avg MAPE: ~11.6%
🏆
Holt's Double Exponential Smoothing
Lₜ = α·Yₜ + (1−α)(Lₜ₋₁ + Bₜ₋₁)
Bₜ = β·(Lₜ − Lₜ₋₁) + (1−β)·Bₜ₋₁
Ŷₜ₊ₕ = Lₜ + h·Bₜ
α=0.50 · β=0.15 (optimised)
Captures both level and trend. Parameters tuned via rolling-window cross-validation on 18 months of training data. Outperforms Naïve by ~30% MAPE reduction across all categories.
Avg MAPE: ~9.9% ✓ Best
Actual vs. Forecast — Select Product Category
Electronics Accessories
Training period (Jan–Jun 2024) shown with solid lines. Forecast period (Jul–Nov 2024) shown with dashed lines beyond the training boundary.
Model Performance Comparison — Rolling Window MAPE (%)
Mean Absolute Percentage Error by Method & Category
Evaluated using 6-period rolling window cross-validation on training data. Lower MAPE = better accuracy.
| Product Category |
Naïve MAPE |
SMA MAPE (3-mo) |
Holt's MAPE |
Improvement |
Holt's Accuracy |
MAPE Comparison Across All Models
Grouped bar chart — lower bar = better forecasting accuracy
Key Methodology Notes
Cross-Validation Approach
Rather than a single train/test split, a rolling-window cross-validation was applied: for each month from month 7 onwards, the model was trained on all prior months and evaluated on the next month. This prevents look-ahead bias and provides a robust estimate of real-world performance.
Parameter Optimisation
Holt's model parameters α (level smoothing) and β (trend smoothing) were set to α=0.50 and β=0.15. Higher α gives more weight to recent observations (appropriate for MENA consumer goods with responsive demand). Lower β avoids over-reacting to short-term trend fluctuations.
Business Context — ANDS Dubai Application
This methodology was applied to ANDS Dubai's China-MENA-GCC supply chain. By replacing Naïve (last-month) forecasting with Holt's DES, a ~30% reduction in MAPE was achieved across product categories, translating directly into reduced safety stock requirements, fewer stockouts, and improved S&OP plan accuracy. The 30% forecast accuracy gain reported in the portfolio corresponds to this model migration.
Next Steps — Advanced Models
Future enhancements include: (1) Holt-Winters Seasonal model for products with clear annual seasonality; (2) Facebook Prophet for handling irregular events (Ramadan, DSF promotions); (3) XGBoost with external regressors (price, promotional calendar, market data); (4) Ensemble methods combining statistical and ML forecasts for maximum accuracy.