1 — Overview
Best model
casa_neighbours_only
Mean RMSE (mm)
80.34
± Std
7.62
Models × Folds
8 × 5
Full summary table
| model | rmse_mean | rmse_std | mae_mean | mae_std | n_folds |
|---|---|---|---|---|---|
| casa_neighbours_only | 80.34 | 7.62 | 60.63 | 4.11 | 5 |
| lstm_static | 80.36 | 7.64 | 60.64 | 4.13 | 5 |
| casa_static_alpha | 80.50 | 7.58 | 60.73 | 4.04 | 5 |
| casa_no_climate | 80.50 | 7.60 | 60.74 | 4.07 | 5 |
| casa_full | 80.51 | 7.57 | 60.74 | 4.04 | 5 |
| casa_no_geo | 80.87 | 7.68 | 60.97 | 3.92 | 5 |
| gstarx | 86.82 | 5.13 | 66.18 | 2.41 | 5 |
| gstar | 87.09 | 4.50 | 66.36 | 1.74 | 5 |
2 — Model Comparison
3 — Diebold-Mariano Test (HLN-corrected)
|−log₁₀(p)| > 1.30 ≈ p < 0.05; > 2.0 ≈ p < 0.01; > 3.0 ≈ p < 0.001. Negative-signed (blue) cells: row outperforms column.
Significant pairs (p < 0.05)
| model_A | model_B | dm_hln | p_value | A_better | sig_001 |
|---|---|---|---|---|---|
| casa_neighbours_only | gstar | -4.216 | 5.50e-05 | True | True |
| casa_no_geo | gstar | -4.201 | 5.83e-05 | True | True |
| gstar | lstm_static | 4.193 | 5.99e-05 | False | True |
| casa_static_alpha | gstar | -4.177 | 6.36e-05 | True | True |
| casa_full | gstar | -4.171 | 6.51e-05 | True | True |
| casa_no_climate | gstar | -4.161 | 6.77e-05 | True | True |
| casa_neighbours_only | gstarx | -3.753 | 2.95e-04 | True | True |
| casa_no_geo | gstarx | -3.733 | 3.17e-04 | True | True |
| gstarx | lstm_static | 3.731 | 3.19e-04 | False | True |
| casa_static_alpha | gstarx | -3.710 | 3.42e-04 | True | True |
| casa_full | gstarx | -3.702 | 3.52e-04 | True | True |
| casa_no_climate | gstarx | -3.693 | 3.63e-04 | True | True |
4 — Regime-Stratified RMSE
Hypothesis (Spec Proposition 4): CASA outperforms static during El Niño / La Niña while geographic prior dominates in neutral periods.
5 — α_t Timeline (CASA blending gate)
Pilih ablation via tombol di atas chart untuk switch CASA variant.
6 — Geographic Adjacency Matrix (A_geo)
Density 12.82% = k/(N−1). Rows sum to 1.0. W_t at runtime = α_t · A_geo + (1−α_t) · A_learned^t.
7 — Geographic Network & Per-province RMSE
KNN edges (k=5) overlaid. Marker colour = per-province RMSE from best CASA ablation; orange squares = boundary nodes (PNG, Borneo non-IDN), not forecast targets.