From ffd8b191ef15fdea73dfa66d2162c0234baba725 Mon Sep 17 00:00:00 2001
From: Sven Geboers <geboers.sven@gmail.com>
Date: Sun, 29 Mar 2026 20:12:24 +0200
Subject: [PATCH] fix: use annual-only windows for SVD to restore EVR (~20%
 PC1)

Quarterly windows (29 of 41 total) diluted PC1 explained variance ratio
from ~20% down to ~14.6%. The fix splits the vector collection loop into:
- pca_vecs: annual windows only (re.match r'^\d{4}$') -> M_pca used for SVD
- all_vecs: every window -> M used for projections onto derived axes

Centering for SVD and global_mean for projection both now use M_pca.mean(axis=0)
so axes are consistent. Falls back to all windows if no annual windows exist.
---
 analysis/political_axis.py | 31 ++++++++++++++++++++++++-------
 1 file changed, 24 insertions(+), 7 deletions(-)

diff --git a/analysis/political_axis.py b/analysis/political_axis.py
index 3adfd14..b296be9 100644
--- a/analysis/political_axis.py
+++ b/analysis/political_axis.py
@@ -14,6 +14,7 @@ Both modes return a dict mapping mp_name → scalar score for the given window.
 
 import json
 import logging
+import re
 from typing import Dict, List, Optional, Tuple
 
 import numpy as np
@@ -190,17 +191,23 @@ def compute_2d_axes(
 
     aligned_window_vecs = _trajectory._procrustes_align_windows(raw_window_vecs)
 
-    # Stack all vectors across windows into a single matrix for PCA if needed
+    # Stack all vectors across windows into a single matrix for PCA if needed.
+    # pca_vecs / pca_index: annual windows only (e.g. "2024") — used for SVD axis derivation.
+    # all_vecs / entity_index: every window — used for projection onto the derived axes.
+    pca_vecs = []
     all_vecs = []
     entity_index = []  # parallel list of (window_id, entity)
     for wid, d in aligned_window_vecs.items():
         for ent, v in d.items():
             if normalize_vectors:
                 n = np.linalg.norm(v)
-                all_vecs.append(v / n if n > 1e-10 else v)
+                vec = v / n if n > 1e-10 else v
             else:
-                all_vecs.append(v)
+                vec = v
+            all_vecs.append(vec)
             entity_index.append((wid, ent))
+            if re.match(r"^\d{4}$", wid):
+                pca_vecs.append(vec)
 
     if len(all_vecs) == 0:
         _logger.info("No vectors loaded for windows %s", window_ids)
@@ -208,9 +215,19 @@ def compute_2d_axes(
 
     M = np.vstack(all_vecs)
 
+    # If no annual windows found, fall back to all windows for SVD.
+    if len(pca_vecs) == 0:
+        _logger.warning(
+            "No annual windows found; falling back to all %d windows for SVD axis derivation",
+            len(aligned_window_vecs),
+        )
+        M_pca = M
+    else:
+        M_pca = np.vstack(pca_vecs)
+
     if method == "pca":
-        # centre globally
-        Mc = M - M.mean(axis=0)
+        # centre using annual-only mean so SVD axes are not diluted by quarterly windows
+        Mc = M_pca - M_pca.mean(axis=0)
         try:
             U, s, Vt = np.linalg.svd(Mc, full_matrices=False)
         except np.linalg.LinAlgError:
@@ -358,8 +375,8 @@ def compute_2d_axes(
                 evr1 * 100,
             )
 
-        # project per-window vectors (centre by global mean)
-        global_mean = M.mean(axis=0)
+        # project per-window vectors (centre by annual-window global mean, consistent with SVD axes)
+        global_mean = M_pca.mean(axis=0)
         axes["global_mean"] = global_mean
         positions_by_window: Dict[str, Dict[str, Tuple[float, float]]] = {
             wid: {} for wid in window_ids