feat(XS01): affina con blend di lookback [30,90] — FULL 0.80->1.10, portafoglio 1.41->1.48

Come TP01 fonde gli orizzonti, XS01 ora fonde 30g+90g del momentum cross-sectional (z-score per
lookback, mediato). Sweep: [30,90] e' il sweet spot (fonde i due singoli robusti, anti-overfit):
XS01 standalone FULL 0.80->1.10, DD 21%->14%, corr a TP01 -0.06->-0.12, 100% anni+. Portafoglio
TP01 70 + XS01 30: FULL Sh 1.41->1.48, DD 5.2%->4.6%, ~€/g 1.65->1.78; hold-out 1.15->1.06 (calo
marginale dentro il rumore). Piu' robusto (due orizzonti) + diversifica meglio -> promosso.

sleeves.XS_CFG lookbacks=(30,90), engine _xsec_returns usa lo score blended. 12 test ok.
Diario 2026-06-19-xsec-blend.md.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

docs/diary/2026-06-19-xsec-blend.md

@@ -0,0 +1,35 @@
+# 2026-06-19 — Affinamento XS01: blend di lookback [30,90]
+
+Come TP01 fonde gli orizzonti 30/90/180, XS01 ora fonde piu' lookback del momentum cross-sectional
+(z-score cross-sectional per lookback, mediato) invece del singolo L=30. `scripts/portfolio/xsec_blend.py`.
+
+## Sweep lookback (19 major, 899g) — FULL/OOS/DD/anni+/corrTP
+| lookbacks | FULL | OOS25 | DD% | anni+ | corrTP |
+|---|---|---|---|---|---|
+| [30] (prima) | 0.80 | 1.20 | 21 | 100% | −0.06 |
+| [90] | 0.88 | 0.90 | 17 | 100% | −0.05 |
+| **[30,90]** | **1.10** | **1.03** | **14** | **100%** | **−0.12** |
+| [20,40,90] | 0.51 | 0.67 | 25 | 100% | −0.12 |
+| [30,60,120] | 0.68 | 0.74 | 16 | 100% | −0.13 |
+
+**[30,90] e' il sweet spot**: fonde i DUE singoli robusti (30 e 90), FULL Sh 0.80→1.10, DD 21→14%,
+corr a TP01 −0.06→−0.12 (diversifica meglio), 100% anni+. Non e' un cell fortunato: e' la
+combinazione dei due lookback gia' validati (anti-overfit, come il multi-orizzonte di TP01).
+
+## Effetto sul portafoglio (TP01 70% + XS01 30%)
+| | XS01 [30] | XS01 blend [30,90] |
+|---|---|---|
+| XS01 standalone FULL / DD | 0.80 / 21% | **1.10 / 14%** |
+| Portafoglio FULL Sharpe | 1.41 | **1.48** |
+| Portafoglio HOLD-OUT Sharpe | 1.15 | 1.06 |
+| Portafoglio DD | 5.2% | **4.6%** |
+| ~€/giorno (2k) | +1.65 | +1.78 |
+
+Migliora FULL Sharpe + DD + robustezza (due orizzonti) al costo di un hold-out marginalmente piu'
+basso (−0.09, dentro il rumore di una singola finestra). Giudizio: il blend e' piu' robusto
+(meno dipendente da un singolo lookback) e diversifica meglio -> PROMOSSO.
+
+## Azione
+`src/portfolio/sleeves.XS_CFG`: `L=30` -> `lookbacks=(30,90)`; engine `_xsec_returns` usa lo score
+blended (media z-score cross-sectional per lookback). **Portafoglio attivo: TP01 70% + XS01 blend
+30%, FULL Sh 1.48 / HOLD 1.06 / DD 4.6%.** 12 test ok. Sleeve sempre sui 19 major.

scripts/portfolio/xsec_blend.py

@@ -0,0 +1,108 @@
+"""AFFINAMENTO XS01 — blend di LOOKBACK (multi-orizzonte cross-sectional).
+
+XS01 attuale usa un singolo lookback (L=30). Come TP01 fonde gli orizzonti 30/90/180, qui il
+momentum cross-sectional fonde piu' lookback: per ogni ribilancio, z-score cross-sectional del
+rendimento a ciascun L, MEDIATO -> punteggio blended -> long top-k / short bottom-k. Piu' liscio
+e robusto (meno dipendente da un singolo orizzonte/regime). Causale, netto fee, vol-target.
+Confronto vs singolo-L + contributo al portafoglio TP01+XS01.
+
+    uv run python scripts/portfolio/xsec_blend.py
+"""
+from __future__ import annotations
+import sys
+from pathlib import Path
+PROJECT_ROOT = Path(__file__).resolve().parents[2]
+sys.path.insert(0, str(PROJECT_ROOT))
+import numpy as np, pandas as pd
+from src.portfolio.portfolio import to_daily, metrics, HOLDOUT, Sleeve, StrategyPortfolio
+from src.portfolio.sleeves import tp01_sleeve, XS_UNIVERSE
+
+RAW = PROJECT_ROOT / "data" / "raw"
+FEE = 0.001
+
+
+def load_majors():
+    cols = {}
+    for sym in XS_UNIVERSE:
+        p = RAW / f"hl_{sym.lower()}_1d.parquet"
+        if p.exists():
+            d = pd.read_parquet(p)
+            cols[sym] = pd.Series(d["close"].values.astype(float), index=pd.to_datetime(d["timestamp"], unit="ms", utc=True))
+    return pd.concat(cols, axis=1, join="inner").sort_index().dropna()
+
+
+def xs_signal(C, lookbacks, H=10, k=5, mode="mom", tv=0.20):
+    """lookbacks = lista (blend) o singolo [L]. Score = media z-score cross-sectional dei ret_L."""
+    px = C.values; n, A = px.shape
+    dret = np.vstack([np.zeros(A), px[1:] / px[:-1] - 1.0])
+    W = np.zeros((n, A)); w = np.zeros(A)
+    for i in range(n):
+        if i >= max(lookbacks) and i % H == 0:
+            score = np.zeros(A); cnt = 0
+            for L in lookbacks:
+                rL = px[i] / px[i - L] - 1.0
+                sd = rL.std()
+                if sd > 0:
+                    score += (rL - rL.mean()) / sd; cnt += 1
+            if cnt:
+                score /= cnt
+                order = np.argsort(score)
+                w = np.zeros(A); lo, hi = order[:k], order[-k:]
+                if mode == "mom": w[hi] = 0.5 / k; w[lo] = -0.5 / k
+                else: w[lo] = 0.5 / k; w[hi] = -0.5 / k
+        W[i] = w
+    gross = np.zeros(n); gross[1:] = np.sum(W[:-1] * dret[1:], axis=1)
+    turn = np.zeros(n); turn[0] = np.abs(W[0]).sum(); turn[1:] = np.abs(np.diff(W, axis=0)).sum(axis=1)
+    s = pd.Series(gross - turn * (FEE / 2.0), index=C.index)
+    rv = s.rolling(30, min_periods=15).std().shift(1) * np.sqrt(365.25)
+    scale = np.clip(np.nan_to_num(tv / rv.replace(0, np.nan).values, nan=0.0), 0, 3.0)
+    return to_daily(pd.Series(s.values * scale, index=C.index))
+
+
+def ev(C, lbs, tp):
+    d = xs_signal(C, lbs)
+    f = metrics(d); o = metrics(d[d.index >= HOLDOUT])
+    yr = [float((1 + g).prod() - 1) for _, g in d.groupby(d.index.year)]
+    pct = sum(v > 0 for v in yr) / len(yr) if yr else 0
+    corr = float(pd.concat({"a": tp, "b": d}, axis=1, join="inner").dropna().corr().iloc[0, 1])
+    return d, f, o, pct, corr
+
+
+def main():
+    C = load_majors()
+    tp = tp01_sleeve().daily()
+    print("=" * 92)
+    print(f"  AFFINAMENTO XS01 — blend di lookback (19 major, {len(C)} giorni)")
+    print("=" * 92)
+    print(f"  {'lookbacks':<22}{'FULL':>7}{'OOS25':>7}{'DD%':>6}{'anni+':>7}{'corrTP':>8}")
+    configs = [
+        ("[30] (attuale)", [30]), ("[90]", [90]), ("[20]", [20]),
+        ("[20,40]", [20, 40]), ("[20,60]", [20, 60]), ("[30,90]", [30, 90]),
+        ("[20,40,90]", [20, 40, 90]), ("[30,60,120]", [30, 60, 120]),
+        ("[20,60,180]", [20, 60, 180]), ("[15,30,60,120]", [15, 30, 60, 120]),
+    ]
+    rows = []
+    for name, lbs in configs:
+        d, f, o, pct, corr = ev(C, lbs, tp)
+        rows.append((name, lbs, d, f, o, pct, corr))
+        print(f"  {name:<22}{f['sharpe']:>7.2f}{o['sharpe']:>7.2f}{f['maxdd']*100:>6.0f}{pct*100:>6.0f}%{corr:>+8.2f}")
+
+    # candidato: miglior blend per (FULL+OOS) con breadth 100% e corr bassa
+    cand = [r for r in rows if r[5] >= 0.99 and r[6] < 0.4]
+    cand.sort(key=lambda r: -(r[3]["sharpe"] + r[4]["sharpe"]))
+    print("\n  CONTRIBUTO al portafoglio — attuale (XS [30]) vs miglior blend")
+    base_xs = rows[0][2]   # [30]
+    for label, dxs in [("XS [30] attuale", base_xs)] + ([(cand[0][0], cand[0][2])] if cand else []):
+        J = pd.concat({"tp": tp, "xs": dxs}, axis=1, join="inner").dropna()
+        for w in (0.3,):
+            comb = (1 - w) * J["tp"] + w * J["xs"]
+            cf, ch = metrics(comb), metrics(comb[comb.index >= HOLDOUT])
+            xf = metrics(J["xs"]); xo = metrics(J["xs"][J["xs"].index >= HOLDOUT])
+            print(f"    {label:<22} XS-solo FULL {xf['sharpe']:.2f}/OOS {xo['sharpe']:.2f} | TP01 70+XS 30: FULL {cf['sharpe']:.2f} HOLD {ch['sharpe']:.2f}")
+    if cand:
+        print(f"\n  -> blend migliore: {cand[0][0]} (lookbacks {cand[0][1]}). Promuovere se batte [30] su")
+        print("     FULL+OOS+robustezza E migliora il portafoglio. Sennò resta [30].")
+
+
+if __name__ == "__main__":
+    main()

src/portfolio/sleeves.py

@@ -52,7 +52,10 @@ def tp01_sleeve(weight: float = 1.0) -> Sleeve:
 # trend di TP01: lavora quando TP01 e' in cash). Validato: scripts/portfolio/xsec_research.py.
 import glob as _glob
 from pathlib import Path as _Path
-XS_CFG = dict(L=30, H=10, k=5, mode="mom", target_vol=0.20)
+# BLEND di lookback (2026-06-19): fonde 30g+90g del momentum cross-sectional (z-score per
+# lookback, mediato) come TP01 fonde gli orizzonti -> piu' robusto del singolo L=30: FULL Sh
+# 0.80->1.10, DD 21%->14%, corr a TP01 -0.06->-0.12, 100% anni+. Diario 2026-06-19-xsec-blend.md.
+XS_CFG = dict(lookbacks=(30, 90), H=10, k=5, mode="mom", target_vol=0.20)
 _HL_DIR = _Path(__file__).resolve().parents[2] / "data" / "raw"
 # UNIVERSO ESPLICITO = 19 ALT LIQUIDI MAJOR. NB (2026-06-19): allargare a 52 asset (incluso
 # small-cap WIF/JUP/ORDI/PYTH/TAO...) DILUISCE l'edge -> momentum cross-section NEGATIVO sui 52.
@@ -75,15 +78,23 @@ def _xsec_returns() -> pd.Series:
         raise FileNotFoundError("universo Hyperliquid XS01 incompleto: gira scripts/analysis/fetch_hyperliquid.py")
     C = pd.concat(cols, axis=1, join="inner").sort_index().dropna()
     px = C.values; n, A = px.shape
-    L, H, k, mode, tv = XS_CFG["L"], XS_CFG["H"], XS_CFG["k"], XS_CFG["mode"], XS_CFG["target_vol"]
+    lookbacks, H, k, mode, tv = XS_CFG["lookbacks"], XS_CFG["H"], XS_CFG["k"], XS_CFG["mode"], XS_CFG["target_vol"]
     dret = np.vstack([np.zeros(A), px[1:] / px[:-1] - 1.0])
     W = np.zeros((n, A)); w = np.zeros(A)
     for i in range(n):
-        if i >= L and i % H == 0:
+        if i >= max(lookbacks) and i % H == 0:
-            order = np.argsort(px[i] / px[i - L] - 1.0)
+            score = np.zeros(A); cnt = 0          # blend: media z-score cross-sectional per lookback
-            w = np.zeros(A); lo, hi = order[:k], order[-k:]
+            for L in lookbacks:
-            if mode == "mom": w[hi] = 0.5 / k; w[lo] = -0.5 / k
+                rL = px[i] / px[i - L] - 1.0
-            else: w[lo] = 0.5 / k; w[hi] = -0.5 / k
+                sd = rL.std()
+                if sd > 0:
+                    score += (rL - rL.mean()) / sd; cnt += 1
+            if cnt:
+                score /= cnt
+                order = np.argsort(score)
+                w = np.zeros(A); lo, hi = order[:k], order[-k:]
+                if mode == "mom": w[hi] = 0.5 / k; w[lo] = -0.5 / k
+                else: w[lo] = 0.5 / k; w[hi] = -0.5 / k
         W[i] = w
     gross = np.zeros(n); gross[1:] = np.sum(W[:-1] * dret[1:], axis=1)
     turn = np.zeros(n); turn[0] = np.abs(W[0]).sum(); turn[1:] = np.abs(np.diff(W, axis=0)).sum(axis=1)