Đánh giá Mô hình Dự báo

Đã tạo lúc: 2026-11-19 23:56:13 · MAPE gate: 25.0%

Series
9
Models
4
Pass
36
Fail
0

Backtest Results (Rolling Origin)

Báo cáođánh giá dự báo (rolling backtest + MAPE gate)
Dùng đểĐánh giá chất lượng dự báo out-of-sample trước khi công bố báo cáo — cổng chất lượng bắt buộc.
Tạo lúc2026-11-19 23:56
Dữ liệu đến ngày2026-11-19
Phương pháp tính (kỹ thuật)
Rolling-origin backtest: mỗi fold train trên quá khứ, predict 3 tháng tới, tính metrics out-of-sample. MASE < 1 = tốt hơn seasonal naive. Coverage 80%/95% phải gần nominal. MAPE gate 25%.
Series Model MAPE sMAPE MASE RMSE MAE Cov 80% Cov 95% Gate
Latex_Ecom AutoARIMA 14.2% 11.8% 0.78 66.2 46.7 78% 89% ✅
Latex_Ecom AutoETS 14.2% 11.8% 0.78 66.2 46.7 78% 89% ✅
Latex_Ecom CrostonSBA 13.6% 11.8% 0.79 57.5 46.7 44% 44% ✅
Latex_Ecom SeasonalNaive 14.8% 12.5% 0.86 59.1 50.1 89% 89% ✅
Latex_Offline AutoARIMA 13.9% 12.0% 1.00 98.7 82.9 89% 89% ✅
Latex_Offline AutoETS 13.9% 12.0% 1.00 98.7 82.9 89% 89% ✅
Latex_Offline CrostonSBA 11.4% 10.1% 0.82 82.7 68.1 44% 44% ✅
Latex_Offline SeasonalNaive 9.6% 8.4% 0.67 75.1 55.9 89% 100% ✅
Latex_Online AutoARIMA 18.3% 15.3% 0.97 34.1 28.2 78% 89% ✅
Latex_Online AutoETS 18.3% 15.3% 0.97 34.1 28.2 78% 89% ✅
Latex_Online CrostonSBA 15.0% 12.9% 0.80 28.3 23.1 56% 56% ✅
Latex_Online SeasonalNaive 17.3% 14.9% 0.94 31.5 27.6 89% 100% ✅
Pillow_Ecom AutoARIMA 14.9% 12.9% 0.83 77.2 64.4 89% 89% ✅
Pillow_Ecom AutoETS 14.9% 12.9% 0.84 77.2 64.5 89% 100% ✅
Pillow_Ecom CrostonSBA 11.9% 10.5% 0.66 64.5 51.1 56% 56% ✅
Pillow_Ecom SeasonalNaive 11.6% 10.1% 0.63 66.6 48.3 89% 100% ✅
Pillow_Offline AutoARIMA 14.7% 12.5% 0.85 138.1 106.2 78% 89% ✅
Pillow_Offline AutoETS 14.7% 12.6% 0.85 138.2 106.2 89% 100% ✅
Pillow_Offline CrostonSBA 13.0% 11.5% 0.76 118.7 95.6 56% 56% ✅
Pillow_Offline SeasonalNaive 10.0% 8.7% 0.59 100.3 72.8 89% 100% ✅
Pillow_Online AutoARIMA 15.3% 13.1% 0.87 38.1 30.4 89% 89% ✅
Pillow_Online AutoETS 15.3% 13.1% 0.87 38.1 30.4 89% 100% ✅
Pillow_Online CrostonSBA 13.8% 12.3% 0.80 32.9 28.2 56% 56% ✅
Pillow_Online SeasonalNaive 10.3% 9.2% 0.59 24.6 20.3 89% 100% ✅
Spring_Ecom AutoARIMA 9.3% 8.3% 0.55 27.2 20.3 89% 100% ✅
Spring_Ecom AutoETS 9.3% 8.3% 0.55 27.2 20.3 89% 100% ✅
Spring_Ecom CrostonSBA 8.5% 7.8% 0.52 23.1 19.1 67% 67% ✅
Spring_Ecom SeasonalNaive 12.3% 11.5% 0.77 33.5 28.2 89% 100% ✅
Spring_Offline AutoARIMA 16.1% 13.6% 0.91 72.6 56.1 78% 89% ✅
Spring_Offline AutoETS 16.1% 13.6% 0.91 72.7 56.2 78% 89% ✅
Spring_Offline CrostonSBA 13.2% 11.4% 0.73 61.9 45.6 67% 67% ✅
Spring_Offline SeasonalNaive 12.8% 10.9% 0.74 59.3 45.2 89% 100% ✅
Spring_Online AutoARIMA 13.0% 13.7% 0.87 20.0 15.5 100% 100% ✅
Spring_Online AutoETS 17.6% 15.5% 1.13 23.4 20.4 89% 100% ✅
Spring_Online CrostonSBA 9.3% 8.9% 0.60 12.5 10.6 78% 78% ✅
Spring_Online SeasonalNaive 11.3% 11.5% 0.74 17.6 13.2 100% 100% ✅

Methodology

Rolling-origin backtest: For each fold, models are trained on all data before the test window and predict 3 steps ahead. Metrics are computed out-of-sample.

Challengers: SeasonalNaive and Croston serve as baselines. AutoARIMA/AutoETS must beat these to be considered useful.

MASE: Mean Absolute Scaled Error vs seasonal naive. MASE < 1 means better than naive.

Coverage: Percentage of actuals falling within 80%/95% prediction intervals. Should be close to nominal.

MAPE gate: Series with MAPE > 25.0% are marked as failing and should not be used for decisions.