Warrior_EA/research/deep_test.py

"""
THE OUTSTANDING VALIDATION: does the vol-gated dip-z survive a bear market?

The H4 result (STRATEGY.md) is built on 2021-2026, which contains no 2008 and no
2020. This family's losses are concentrated in bear ONSETS, so a sample without
one proves very little about the drawdown that matters.

Broker intraday history does not reach back far enough, so the bear test runs on
D1, where the indices go back to 2008. The D1 rule is the same shape as the H4
one; it is a proxy, and the point is not to re-measure the edge but to see what
the drawdown does when the market actually falls.

Reported per calendar year, because an average over 18 years hides exactly the
thing being looked for.
"""

from __future__ import annotations

import sys
import numpy as np

sys.path.insert(0, __file__.rsplit("\\", 1)[0] if "\\" in __file__ else ".")

import backtest as bt                    # noqa: E402
from run_screen import zscore_entries    # noqa: E402
from vol_filter_test import vol_pctile   # noqa: E402

STOP_ATR = 3.0
PROP_DD = 0.05
IDX = ["SP500", "NAS100", "US30", "DAX40"]


def trades_for(symbols, period, gated, max_bars=10):
    out = []
    for s in symbols:
        try:
            d = bt.load(s, period)
        except Exception:  # noqa: BLE001
            continue
        e, xma = zscore_entries(d, 20, -1.5, 0)
        if gated:
            e = e & (np.nan_to_num(vol_pctile(d), nan=-1) >= 0.50)
        tr = bt.simulate(d, e, side=1, exit_ma=xma, max_bars=max_bars, stop_atr=STOP_ATR)
        tr = bt.add_r(tr, d, stop_atr=STOP_ATR)
        for t in tr:
            t["symbol"] = s
            t["exit_t"] = d["ts"][t["exit_i"]]
        out += tr
    out.sort(key=lambda t: t["exit_t"])
    return out


def curve_stats(trades, risk):
    eq = [1.0]
    for t in trades:
        eq.append(eq[-1] * (1.0 + risk * t["r"]))
    eq = np.array(eq)
    peak = np.maximum.accumulate(eq)
    dd = (peak - eq) / peak
    return eq, dd


def per_year(trades, risk):
    """Year-by-year return and the worst drawdown inside that year."""
    years = sorted({int(str(t["exit_t"])[:4]) for t in trades})
    rows = []
    for y in years:
        sub = [t for t in trades if int(str(t["exit_t"])[:4]) == y]
        if not sub:
            continue
        eq, dd = curve_stats(sub, risk)
        rows.append((y, len(sub), eq[-1] - 1.0, dd.max(),
                     float(np.mean([t["ret"] for t in sub])) * 1e4))
    return rows


if __name__ == "__main__":
    risk = 0.0025
    print("=== D1, 4 indices, dip-z (z<=-1.5, exit SMA20, 10 bars, stop 3ATR), "
          f"risk {risk:.2%} ===\n")
    for gated in (False, True):
        tag = "VOL-GATED (pct>=0.50)" if gated else "UNGATED"
        tr = trades_for(IDX, "PERIOD_D1", gated)
        if not tr:
            print(f"{tag}: no trades / no data")
            continue
        eq, dd = curve_stats(tr, risk)
        t0 = min(t["t"] for t in tr)
        t1 = max(t["exit_t"] for t in tr)
        yrs = (t1 - t0) / np.timedelta64(365, "D")
        total = eq[-1] - 1.0
        rdd = total / dd.max() if dd.max() > 1e-9 else np.nan
        print(f"--- {tag} --- {str(t0)[:10]} .. {str(t1)[:10]}  "
              f"n={len(tr)}  {len(tr)/max(yrs*12,1e-9):.1f}/mo  "
              f"total {total:.1%}  CAGR {eq[-1]**(1/max(yrs,1e-9))-1:.1%}  "
              f"maxDD {dd.max():.1%}  ret/DD {rdd:.2f}")
        print(f"{'year':>6}{'n':>5}{'return':>9}{'maxDD':>8}{'bp/trade':>10}")
        for y, n, r, d_, bp in per_year(tr, risk):
            mark = "  <-- bear" if y in (2008, 2011, 2015, 2018, 2020, 2022) else ""
            print(f"{y:>6}{n:>5}{r:>9.2%}{d_:>8.2%}{bp:>10.1f}{mark}")
        print()