SniperGold_ML/ml/p3/regime_diagnostic.py

# -*- coding: utf-8 -*-
"""P3.1e REGIME DIAGNOSTIC — SniperGold_ML.

Diagnostic ONLY. Answers the fifth research question:
  "Apakah relationship feature -> label berubah berdasarkan market state?"

Regime proxies use EXISTING semantics only (no new regime model):
  * ATR percentile   : rolling percentile rank of ATR over trailing 500 bars
  * Realized vol     : std of 20-bar log returns (trailing)
  * Trend proxy      : |f16 mom20_atr| magnitude (already a contract feature)
  * Range proxy      : f17 range_atr (already a contract feature)

Analysis:
  * P(Y=1) / P(Y=-1) per regime tercile (low/med/high)
  * feature x regime: univariate AUC of key features within each tercile
    -> does the feature->label relationship change across regimes?
  * regime x SMC state: P(Y=1) for selected combos

No MS-GARCH / HMM / regime gate is built.
"""
import os
import sys
import json

import numpy as np

HERE = os.path.dirname(os.path.abspath(__file__))
if HERE not in sys.path:
    sys.path.insert(0, HERE)
import p3_common as P3
import train_model as TM

KEY_FEATS = [0, 5, 7, 9, 16, 18]


def terciles(x):
    q33, q66 = np.quantile(x, [1 / 3, 2 / 3])
    lo = x <= q33
    hi = x >= q66
    md = ~lo & ~hi
    return lo, md, hi


def main():
    t, o, h, l, c, v, htf = P3.load_data()
    F = P3.load_F()
    A = P3.atr_series(h, l, c)
    n = len(c)
    lab = P3.make_label(c, A)
    labeled = lab != 0
    midx = np.where(labeled)[0]
    y = lab[midx]
    ypos = (y == 1).astype(int)

    # ---- regime proxies ----
    atr_pct = np.zeros(n)
    for i in range(500, n):
        atr_pct[i] = (A[i - 500:i] <= A[i]).mean()
    atr_pct[:500] = 0.5

    lr = np.zeros(n)
    lr[1:] = np.log(c[1:] / np.maximum(c[:-1], 1e-12))
    rv = np.zeros(n)
    for i in range(20, n):
        rv[i] = lr[i - 19:i + 1].std(ddof=1)
    rv[:20] = np.nan
    rv = np.nan_to_num(rv, nan=np.nanmedian(rv[20:]))

    proxies = {
        "ATR_percentile": atr_pct,
        "RealizedVol20": rv,
        "TrendAbs_mom20(f16)": np.abs(F[:, 16]),
        "RangeATR(f17)": F[:, 17],
    }

    report = {"provenance": P3.provenance(),
              "dataset_hash": P3.dataset_hash(F, lab, A, t),
              "regimes": {}}

    print("=== REGIME DIAGNOSTIC (tercile low/med/high) ===")
    for pname, px in proxies.items():
        lo, md, hi = terciles(px)
        r = {}
        print(f"\n[{pname}]")
        for rname, m in (("low", lo), ("med", md), ("high", hi)):
            ml = m[midx]
            if ml.sum() < 100:
                continue
            p1 = float(ypos[ml].mean())
            pm1 = float((y[ml] == -1).mean())
            r[rname] = {"n": int(ml.sum()), "P1": p1, "Pm1": pm1,
                        "lift_P1": float(p1 / ypos.mean())}
            print(f"  {rname:>4}: n={int(ml.sum()):>6} P(+1)={p1:.4f} "
                  f"P(-1)={pm1:.4f} lift={p1 / ypos.mean():.3f}")
            for j in KEY_FEATS:
                xs = F[midx, j][ml]
                if len(np.unique(xs)) < 3:
                    continue
                a = TM.auc(ypos[ml].astype(int), xs)
                r[rname]["auc_" + P3.FEAT_NAMES[j]] = float(a)
        print("  univariate AUC by regime (long label):")
        for fname in [P3.FEAT_NAMES[j] for j in KEY_FEATS]:
            vals = []
            for rname in ("low", "med", "high"):
                vv = r.get(rname, {}).get("auc_" + fname)
                vals.append(vv if vv is not None else float("nan"))
            print("    %-16s low=%.4f med=%.4f high=%.4f" % (fname, vals[0], vals[1], vals[2]))
        report["regimes"][pname] = r

    # ---- regime x SMC state combos ----
    print("\n=== REGIME x SMC STATE: P(Y=1) selected combos ===")
    atr_lo, atr_md, atr_hi = terciles(atr_pct)
    f9 = F[:, 9]
    f7 = F[:, 7]
    f18 = F[:, 18]
    combos = {
        "BUY_cand & ATR_low": (f9 == 1) & (f7 > 0) & atr_lo,
        "BUY_cand & ATR_high": (f9 == 1) & (f7 > 0) & atr_hi,
        "SELL_cand & ATR_low": (f9 == 1) & (f7 < 0) & atr_lo,
        "SELL_cand & ATR_high": (f9 == 1) & (f7 < 0) & atr_hi,
        "conf>=60 & ATR_low": (f18 >= 60) & atr_lo,
        "conf>=60 & ATR_high": (f18 >= 60) & atr_hi,
        "MTF_bull & ATR_low": (F[:, 0] > 0) & (F[:, 1] > 0) & (F[:, 2] > 0) & atr_lo,
        "MTF_bull & ATR_high": (F[:, 0] > 0) & (F[:, 1] > 0) & (F[:, 2] > 0) & atr_hi,
    }
    comb = {}
    for cname, mask in combos.items():
        ml = mask[midx]
        if ml.sum() < 100:
            comb[cname] = {"n": int(ml.sum()), "note": "too few"}
            print(f"  {cname:<24} n={int(ml.sum())} (too few)")
            continue
        p1 = float(ypos[ml].mean())
        pm1 = float((y[ml] == -1).mean())
        comb[cname] = {"n": int(ml.sum()), "P1": p1, "Pm1": pm1,
                       "lift_P1": float(p1 / ypos.mean())}
        print(f"  {cname:<24} n={int(ml.sum()):>6} P(+1)={p1:.4f} "
              f"P(-1)={pm1:.4f} lift={p1 / ypos.mean():.3f}")
    report["regime_x_smc"] = comb

    P3.save_json("regime_diagnostic.json", report)
    print("\nRegime diagnostic selesai. Output: ml/p3/output/regime_diagnostic.json")


if __name__ == "__main__":
    main()