research(premium): ondata premi cross-venue auto-calcolati — CBPREM morto, KIMCHI d30 ucciso dallo scettico (spike+lag)

CBPREM (Coinbase vs feed cert, 2015->): HEDGE, hold -0.46. KIMCHI (Upbit/ECB, 2017->,
ancora 00:00 UTC verificata): EARNS_SLOT=True al marginal scorer (ADDS, robust_oos,
uplift + ogni anno) ma DSR 0.891 -> scettico obbligatorio: niente plateau
(d15/30/45/60 = 0.44/1.07/0.51/0.24) e lag +1g azzera (hold 0.74->0.14) = parameter
luck. Lezione codificata: EARNS_SLOT con DSR<0.95 -> sempre plateau+lag skeptic.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

scripts/research/r0724_premium_wave.py

@@ -0,0 +1,141 @@
+"""r0724_premium_wave — premi cross-venue AUTO-CALCOLATI: Coinbase premium + Kimchi (2026-07-24).
+
+Seconda parte dell'ondata "trova altre strategie" (dopo r0724_onchain_wave): la ricerca
+web (agente on-chain/dati) indica i premi regionali come UNICA famiglia flow con dati
+100% auto-calcolabili da candele raw -> zero rischio-revisione/vintage del vendor
+(coerente con la dottrina dati del progetto). Evidenza accademica: kimchi = anomalia
+documentata (violazione persistente della legge del prezzo unico, capital controls);
+lead-lag ASIMMETRICO e TEMPO-VARIANTE (MDPI 2026) -> nessuna regola pubblicata onesta,
+qui si meccanizza da zero.
+
+SEGNALI (mai il prezzo): per asset a e giorno d
+  CBPREM_a(d)  = close Coinbase USD (00:00 UTC) / close feed certificato - 1
+  KIMCHI_a(d)  = close Upbit KRW (00:00 UTC, ancora 09:00 KST) / (USDKRW_ECB x close cert) - 1
+Le candele Upbit daily sono ancorate a mezzanotte UTC (=09:00 KST) -> stesso istante di
+chiusura del feed certificato, nessun premio finto da mismatch orario. FX = fixing ECB
+del giorno (ffill weekend; il KRW si muove ~nulla vs la vol crypto — caveat dichiarato).
+NB: il feed certificato Deribit e' esso stesso un indice multi-exchange che include
+Coinbase -> il CBPREM misurato e' SMORZATO (caveat strutturale).
+
+Griglie (piccole, tutte contate nel deflated-Sharpe): z-score rolling 180g del livello
+(follow / contrarian) e segno della variazione 30g (follow), LF e LS -> 6 celle/famiglia.
+Gate: study_family_honest (cella in-sample, DSR, marginal scorer vs TP01).
+
+Dati: data/external/premium/ (fetch: scratchpad/fetch_premium.py — Coinbase Exchange
+public candles, Upbit public candles, frankfurter.app ECB; tutti tokenless).
+
+Uso: `uv run python scripts/research/r0724_premium_wave.py`
+"""
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+import numpy as np
+import pandas as pd
+
+ROOT = Path(__file__).resolve().parents[2]
+sys.path.insert(0, str(ROOT))
+sys.path.insert(0, str(ROOT / "scripts" / "research" / "alt"))
+
+import altlib  # noqa: E402
+from altlib import study_family_honest, fmt_marginal, get  # noqa: E402
+
+EXT = ROOT / "data" / "external" / "premium"
+
+
+def _cert_close_by_day(asset: str) -> pd.Series:
+    df = get(asset, "1d")
+    days = pd.DatetimeIndex(pd.to_datetime(df["datetime"], utc=True)).floor("D")
+    return pd.Series(df["close"].values.astype(float), index=days)
+
+
+def _cb_close(asset: str) -> pd.Series:
+    d = pd.read_csv(EXT / f"cb_{asset.lower()}.csv")
+    idx = pd.DatetimeIndex(pd.to_datetime(d["ts"], unit="s", utc=True)).floor("D")
+    return pd.Series(d["close"].values.astype(float), index=idx)
+
+
+def _upbit_close(asset: str) -> pd.Series:
+    d = pd.read_csv(EXT / f"upbit_{asset.lower()}.csv")
+    idx = pd.DatetimeIndex(pd.to_datetime(d["utc"], utc=True)).floor("D")
+    return pd.Series(d["close_krw"].values.astype(float), index=idx)
+
+
+def _fx() -> pd.Series:
+    d = pd.read_csv(EXT / "usdkrw.csv")
+    idx = pd.DatetimeIndex(pd.to_datetime(d["date"], utc=True))
+    s = pd.Series(d["usdkrw"].values.astype(float), index=idx)
+    full = pd.date_range(s.index[0], s.index[-1] + pd.Timedelta(days=3), freq="D", tz="UTC")
+    return s.reindex(full).ffill()
+
+
+def _premia() -> dict:
+    fx = _fx()
+    out = {}
+    for a in ("BTC", "ETH"):
+        cert = _cert_close_by_day(a)
+        cb = _cb_close(a).reindex(cert.index)
+        up = _upbit_close(a).reindex(cert.index)
+        out[a] = pd.DataFrame({
+            "cbprem": cb / cert - 1.0,
+            "kimchi": up / (fx.reindex(cert.index) * cert) - 1.0,
+        }, index=cert.index)
+    return out
+
+
+_PREM = _premia()
+
+
+def prem_factory_maker(col: str):
+    def factory(tf: str, sig: str = "z180", mode: str = "folLF"):
+        def fn(df, asset):
+            p = _PREM[asset][col]
+            if sig == "z180":
+                mu = p.rolling(180, min_periods=90).mean()
+                sd = p.rolling(180, min_periods=90).std()
+                z = (p - mu) / sd
+            else:                          # d30: variazione 30g del premio
+                z = p - p.shift(30)
+            s = np.sign(z) if mode.startswith("fol") else -np.sign(z)
+            pos = s if mode.endswith("LS") else (s > 0).astype(float)
+            days = pd.DatetimeIndex(pd.to_datetime(df["datetime"], utc=True)).floor("D")
+            return np.nan_to_num(pos.reindex(days).values.astype(float))
+        return fn
+    return factory
+
+
+GRID = [
+    dict(sig="z180", mode="folLF"), dict(sig="z180", mode="conLF"),
+    dict(sig="z180", mode="folLS"), dict(sig="z180", mode="conLS"),
+    dict(sig="d30", mode="folLF"), dict(sig="d30", mode="folLS"),
+]
+
+
+def main() -> None:
+    print("=" * 100)
+    print("  ONDATA PREMI CROSS-VENUE — CBPREM + KIMCHI via study_family_honest")
+    for a in ("BTC", "ETH"):
+        P = _PREM[a].dropna()
+        print(f"  {a}: {len(P)} giorni | cbprem medio {P['cbprem'].mean()*1e4:+.1f}bps "
+              f"(p1/p99 {P['cbprem'].quantile(0.01)*1e4:+.0f}/{P['cbprem'].quantile(0.99)*1e4:+.0f}) | "
+              f"kimchi medio {P['kimchi'].mean()*1e4:+.1f}bps "
+              f"(p1/p99 {P['kimchi'].quantile(0.01)*1e4:+.0f}/{P['kimchi'].quantile(0.99)*1e4:+.0f})")
+    print("=" * 100)
+    for fam, col in (("CBPREM-coinbase", "cbprem"), ("KIMCHI-korea", "kimchi")):
+        print("-" * 100)
+        rep = study_family_honest(fam, prem_factory_maker(col), GRID, tfs=("1d",))
+        if rep.get("chosen") is None:
+            print(f"=== {fam}: nessuna cella valida in-sample")
+            continue
+        ch = rep["chosen"]
+        print(f"=== {fam}: cella IS {ch['params']} (IS Sh {ch['insample_sharpe']}, "
+              f"full {ch['full_sharpe']}) su {rep['n_cells']} celle")
+        print(f"    deflated-Sharpe {rep['deflated_sharpe']} (null-max {rep['expected_null_max']})"
+              f"  dsr_pass={rep['dsr_pass']}")
+        print(fmt_marginal(rep["marginal"]))
+        print(f"    >>> EARNS_SLOT_HONEST = {rep['earns_slot_honest']}")
+
+
+if __name__ == "__main__":
+    main()