radar love

Author: michielkallenberg

cybench/runs/analysis/collect_walk_forward_results.py

@@ -397,8 +397,8 @@ def summary_rows_to_dashboard_records(
         ("spatial", "r_agg", "r_spatial_agg"),
         ("temporal", "r", "r_temporal"),
         ("temporal", "r_agg", "r_temporal_agg"),
+        ("anomaly", "r", "r_res"),
         ("anomaly", "r2", "r2_res"),
-        ("anomaly", "r_pooled", "r_res"),
     ]
     records: list[dict[str, Any]] = []
     for row in summary_rows:

cybench/runs/analysis/global_insights_lib.py

@@ -44,11 +44,10 @@
     {
         "id": "anomaly",
         "label": "Anomaly",
-        "note": "Pooled residual R² on location-demeaned yields (variance explained in anomalies). "
-        "Spatial and temporal axes use Pearson r because they assess pattern agreement; "
-        "anomaly uses R² because the objective is explained anomaly variance, not ranking alone.",
+        "note": "Pooled Pearson r on location-demeaned yields (all region×year residuals). "
+        "Distinct from temporal median per-region r; measures whether anomalous years line up globally.",
         "metrics": (
-            {"id": "r2", "column": "r2_res", "label": "R²", "higher_better": True},
+            {"id": "r", "column": "r_res", "label": "r", "higher_better": True},
         ),
     },
 )

cybench/runs/analysis/model_family_radar_lib.py

@@ -39,9 +39,9 @@
     },
     {
         "label": "Anomaly",
-        "metric": "r2_res",
-        "display": "R²",
-        "question": "How much interannual anomaly variance is explained after removing regional means (pooled residual R²)?",
+        "metric": "r_res",
+        "display": "r",
+        "question": "Across all regions and years, do anomalous years line up (pooled Pearson r on location-demeaned yields)?",
     },
 )
 
@@ -86,13 +86,13 @@
     "nrmse": {"lo": 0.1, "hi": 0.30, "higher_better": False, "display": "NRMSE"},
     "r_spatial": {"lo": 0.0, "hi": 1.0, "higher_better": True, "display": "r"},
     "r_temporal": {"lo": 0.0, "hi": 1.0, "higher_better": True, "display": "r"},
-    "r2_res": {"lo": 0.0, "hi": 1.0, "higher_better": True, "display": "R²"},
+    "r_res": {"lo": 0.0, "hi": 1.0, "higher_better": True, "display": "r"},
 }
 
 RADAR_ABSOLUTE_NOTE = (
     "Absolute: fixed scales per axis — NRMSE 0.10 (outer, best) to 0.30 (center, worst); "
-    "Pearson r axes and Anomaly R²: 0 (center) to 1 (outer). Values outside the range are clamped; "
-    "negative values are shown at 0 on the radar."
+    "Pearson r axes 0 (center) to 1 (outer). Values outside the range are clamped; "
+    "negative r is shown at 0."
 )
 
 MODEL_DISPLAY_NAMES: dict[str, str] = {

cybench/runs/viz/global_insights_template.html

@@ -220,7 +220,7 @@ <h3 style="font-size:1rem;margin:1rem 0 0.5rem;">Model × country</h3>
       </div>
       <p class="muted" style="margin-top:0;" id="matrix-metric-note">
         Median column = median across countries per model (click column headers to sort). Green = better.
-        Pattern axes (spatial, temporal) show Pearson r; anomaly shows pooled residual R²; overall shows R²/NRMSE.
+        Pattern axes (spatial, temporal, anomaly) show Pearson r; overall shows R²/NRMSE.
         Country cells link to the matching country dashboard with that model pre-selected.
       </p>
       <div class="table-scroll" id="model-country-wrap"></div>
@@ -378,7 +378,7 @@ <h3 style="font-size:1rem;margin:1rem 0 0.5rem;">Detail per crop × country × m
       },
       { id: "spatial", label: "Spatial", note: "Median per-year r across regions (typical-year slice).", metrics: [{ id: "r", label: "r", higher_better: true }] },
       { id: "temporal", label: "Temporal", note: "Median per-region r across years (typical-region slice).", metrics: [{ id: "r", label: "r", higher_better: true }] },
-      { id: "anomaly", label: "Anomaly", note: "Pooled residual R² on location-demeaned yields.", metrics: [{ id: "r2", label: "R²", higher_better: true }] },
+      { id: "anomaly", label: "Anomaly", note: "Pooled Pearson r on location-demeaned yields (r_res).", metrics: [{ id: "r", label: "r", higher_better: true }] },
     ];
 
     function syncCropSelects(value) {

cybench/runs/viz/model_family_radar_template.html

@@ -195,7 +195,7 @@
     <p class="back"><a href="index.html">← All dashboards</a></p>
     <header>
       <h1>Model family comparison</h1>
-      <p class="lead">NRMSE (Overall); Pearson r pattern slices (Spatial, Temporal); pooled residual R² (Anomaly) — one best-NRMSE representative per modeling paradigm (five families).</p>
+      <p class="lead">NRMSE (Overall); Pearson r (Spatial, Temporal, and pooled anomaly r<sub>res</sub>) — one best-NRMSE representative per modeling paradigm (five families).</p>
       <p class="muted">Paper-level summary. For every model and crop×country drill-down, see <a href="insights.html">Global insights</a> (leaderboard and model×country heatmap).</p>
     </header>
 

tests/runs/test_collect_walk_forward_results.py

@@ -241,7 +241,7 @@ def test_summary_rows_to_dashboard_records(tmp_path: Path):
     assert sp_agg["value"] == 0.12
     tm_agg = next(r for r in records if r["view"] == "temporal" and r["metric"] == "r_agg")
     assert tm_agg["value"] == 0.48
-    an_pool = next(r for r in records if r["view"] == "anomaly" and r["metric"] == "r_pooled")
+    an_pool = next(r for r in records if r["view"] == "anomaly" and r["metric"] == "r")
     assert an_pool["value"] == 0.05
     scatter_recs = [r for r in records if r.get("images", {}).get("scatter")]
     assert scatter_recs

tests/runs/test_model_family_radar.py

@@ -116,7 +116,7 @@ def test_relative_scores_span_unit_interval():
             "nrmse": [0.2, 0.8],
             "r_spatial": [0.1, 0.9],
             "r_temporal": [0.3, 0.7],
-            "r2_res": [0.0, 1.0],
+            "r_res": [0.0, 1.0],
         }
     )
     rel = relative_scores(raw)
@@ -132,7 +132,7 @@ def test_absolute_scores_use_fixed_scales():
             "nrmse": [0.1, 0.2, 0.30],
             "r_spatial": [-0.2, 0.5, 1.0],
             "r_temporal": [0.0, 0.5, 1.0],
-            "r2_res": [0.25, 0.5, 0.75],
+            "r_res": [0.25, 0.5, 0.75],
         }
     )
     abs_scores = absolute_scores(raw)
@@ -148,7 +148,7 @@ def test_absolute_scores_clamps_out_of_range_nrmse():
             "nrmse": [0.05, 0.40],
             "r_spatial": [0.5, 0.5],
             "r_temporal": [0.5, 0.5],
-            "r2_res": [0.5, 0.5],
+            "r_res": [0.5, 0.5],
         }
     )
     abs_scores = absolute_scores(raw)
@@ -189,7 +189,7 @@ def test_build_family_dataset_rows_includes_all_view_metrics():
     rows = build_family_dataset_rows(df, reps)
     assert len(rows) == 1
     assert rows[0]["dataset"] == "maize_DE"
-    assert set(rows[0]["metrics"]) == {"nrmse", "r_spatial", "r_temporal", "r2_res"}
+    assert set(rows[0]["metrics"]) == {"nrmse", "r_spatial", "r_temporal", "r_res"}
 
 
 def test_build_radar_payload_structure(tmp_path: Path):
@@ -205,32 +205,32 @@ def test_build_radar_payload_structure(tmp_path: Path):
                     nrmse=0.22,
                     r_spatial=0.62,
                     r_temporal=0.33,
-                    r2_res=0.12,
                     r_res=0.2,
+                    r2_res=0.12,
                 ),
                 _summary_row(
                     "lightgbm",
                     nrmse=0.15,
                     r_spatial=0.74,
                     r_temporal=0.1,
-                    r2_res=0.28,
                     r_res=0.5,
+                    r2_res=0.28,
                 ),
                 _summary_row(
                     "transformer_lf",
                     nrmse=0.17,
                     r_spatial=0.73,
                     r_temporal=0.21,
-                    r2_res=0.18,
                     r_res=0.4,
+                    r2_res=0.18,
                 ),
                 _summary_row(
                     "tabpfn",
                     nrmse=0.16,
                     r_spatial=0.77,
                     r_temporal=0.31,
-                    r2_res=0.22,
                     r_res=0.45,
+                    r2_res=0.22,
                 ),
             ]
         ).to_csv(d / "walk_forward_summary.csv", index=False)