okx-codex-trader/scripts/search_eth_adaptive_state_exit.py

from __future__ import annotations

import argparse
import sys
from collections import Counter
from dataclasses import dataclass
from pathlib import Path

import pandas as pd

sys.path.insert(0, str(Path(__file__).resolve().parents[1]))

from okx_codex_trader.models import Candle
from okx_codex_trader.sampled_report import SegmentResult, mark_to_market, trade_equity


ETH_SYMBOL = "ETH-USDT-SWAP"
BTC_SYMBOL = "BTC-USDT-SWAP"
BAR = "15m"
YEARS = 10.0
LEVERAGE = 3
INITIAL_EQUITY = 10_000.0
DATA_DIR = Path("data/okx-candles")
OUTPUT_DIR = Path("reports/eth-exploration")
PRIMARY_COST = "maker_taker"
COSTS = (
    ("maker_maker", 0.0012),
    ("maker_taker", 0.0021),
    ("taker_taker", 0.0030),
)
HORIZONS = (
    ("full", None),
    ("3y", pd.DateOffset(years=3)),
    ("1y", pd.DateOffset(years=1)),
    ("6m", pd.DateOffset(months=6)),
    ("3m", pd.DateOffset(months=3)),
    ("30d", pd.DateOffset(days=30)),
)


@dataclass(frozen=True)
class Variant:
    band_length: int
    bandwidth_lookback: int
    bandwidth_quantile: float
    side_mode: str
    btc_filter: str
    eth_vol_cap: float | None
    cooldown_bars: int
    stop_loss_pct: float
    trend_momentum_bars: int
    trend_middle_buffer_pct: float
    trend_middle_confirm_bars: int
    neutral_middle_buffer_pct: float
    neutral_middle_confirm_bars: int
    protect_trigger_pct: float
    protect_lock_pct: float
    protect_trail_giveback_pct: float
    protect_middle_confirm_bars: int
    fast_vol_pct: float
    fast_bandwidth_ratio: float
    fast_middle_buffer_pct: float
    fast_middle_confirm_bars: int
    max_hold_bars: int

    @property
    def name(self) -> str:
        vol = "none" if self.eth_vol_cap is None else f"{self.eth_vol_cap:g}"
        return (
            f"eth-adaptive-state-exit-l{self.band_length}-bw{self.bandwidth_lookback}"
            f"-q{self.bandwidth_quantile:g}-{self.side_mode}-{self.btc_filter}-vc{vol}"
            f"-sl{self.stop_loss_pct:g}-tr{self.trend_momentum_bars}"
            f"-tb{self.trend_middle_buffer_pct:g}-tc{self.trend_middle_confirm_bars}"
            f"-nb{self.neutral_middle_buffer_pct:g}-nc{self.neutral_middle_confirm_bars}"
            f"-pt{self.protect_trigger_pct:g}-pl{self.protect_lock_pct:g}"
            f"-pg{self.protect_trail_giveback_pct:g}-pc{self.protect_middle_confirm_bars}"
            f"-fv{self.fast_vol_pct:g}-fr{self.fast_bandwidth_ratio:g}"
            f"-fb{self.fast_middle_buffer_pct:g}-fc{self.fast_middle_confirm_bars}"
            f"-mh{self.max_hold_bars}"
        )


def _format_ts(ts: int) -> str:
    return pd.to_datetime(ts, unit="ms", utc=True).strftime("%Y-%m-%d %H:%M")


def _load_candles(symbol: str, bar: str) -> list[Candle]:
    frame = pd.read_csv(DATA_DIR / symbol / f"{bar}.csv")
    return [
        Candle(
            symbol=symbol,
            ts=int(row.ts),
            open=float(row.open),
            high=float(row.high),
            low=float(row.low),
            close=float(row.close),
            volume=float(row.volume),
        )
        for row in frame.itertuples(index=False)
    ]


def _align_pair(left: list[Candle], right: list[Candle]) -> tuple[list[Candle], list[Candle]]:
    right_by_ts = {candle.ts: candle for candle in right}
    left_out: list[Candle] = []
    right_out: list[Candle] = []
    for candle in left:
        other = right_by_ts.get(candle.ts)
        if other is not None:
            left_out.append(candle)
            right_out.append(other)
    return left_out, right_out


def close_position(
    *,
    trades: list[dict[str, object]],
    exits: list[dict[str, object]],
    position: dict[str, object],
    candle: Candle,
    exit_price: float,
    reason: str,
) -> tuple[float, bool]:
    margin_used = float(position["margin_used"])
    exit_equity = trade_equity(
        side=str(position["side"]),
        margin_used=margin_used,
        entry_price=float(position["entry_price"]),
        exit_price=exit_price,
        leverage=LEVERAGE,
    )
    pnl = exit_equity - margin_used
    trades.append(
        {
            "side": "Long" if position["side"] == "long" else "Short",
            "entry_ts": int(position["entry_time"]),
            "exit_ts": candle.ts,
            "entry_time": _format_ts(int(position["entry_time"])),
            "exit_time": _format_ts(candle.ts),
            "entry_price": round(float(position["entry_price"]), 4),
            "exit_price": round(exit_price, 4),
            "pnl": round(pnl, 4),
            "return_pct": round(pnl / margin_used * 100.0, 4),
            "cost_weight": 1.0,
            "exit_reason": reason,
            "entry_state": str(position["state"]),
            "mfe_pct": round(float(position["mfe_pct"]) * 100.0, 4),
            "hold_bars": int(position["hold_bars"]),
        }
    )
    exits.append({"ts": candle.ts, "price": exit_price, "side": position["side"]})
    return exit_equity, pnl > 0.0


def favorable_move(side: str, entry_price: float, candle: Candle) -> float:
    if side == "long":
        return candle.high / entry_price - 1.0
    return entry_price / candle.low - 1.0


def close_profit(side: str, entry_price: float, close: float) -> float:
    if side == "long":
        return close / entry_price - 1.0
    return entry_price / close - 1.0


def hard_stop_exit(position: dict[str, object], candle: Candle) -> tuple[float, str] | None:
    side = str(position["side"])
    stop = float(position["stop_price"])
    if side == "long":
        if candle.open <= stop:
            return candle.open, "stop_gap"
        if candle.low <= stop:
            return stop, "stop"
    else:
        if candle.open >= stop:
            return candle.open, "stop_gap"
        if candle.high >= stop:
            return stop, "stop"
    return None


def protection_stop_exit(position: dict[str, object], candle: Candle, variant: Variant) -> tuple[float, str] | None:
    if str(position["state"]) != "protect":
        return None
    side = str(position["side"])
    entry_price = float(position["entry_price"])
    mfe = float(position["mfe_pct"])
    if side == "long":
        lock_stop = entry_price * (1.0 + variant.protect_lock_pct)
        trail_stop = entry_price * (1.0 + mfe - variant.protect_trail_giveback_pct)
        stop = max(float(position["stop_price"]), lock_stop, trail_stop)
        if candle.open <= stop:
            return candle.open, "protect_gap"
        if candle.low <= stop:
            return stop, "protect_trail"
    else:
        lock_stop = entry_price * (1.0 - variant.protect_lock_pct)
        trail_stop = entry_price * (1.0 - mfe + variant.protect_trail_giveback_pct)
        stop = min(float(position["stop_price"]), lock_stop, trail_stop)
        if candle.open >= stop:
            return candle.open, "protect_gap"
        if candle.high >= stop:
            return stop, "protect_trail"
    return None


def adverse_middle(position: dict[str, object], candle: Candle, middle: float, buffer_pct: float) -> bool:
    if position["side"] == "long":
        return candle.close < middle * (1.0 - buffer_pct)
    return candle.close > middle * (1.0 + buffer_pct)


def trend_continues(position: dict[str, object], candle: Candle, middle: float, momentum: float, required_bars: int) -> bool:
    if position["side"] == "long":
        above_middle = candle.close > middle
        same_direction = momentum > 0.0
    else:
        above_middle = candle.close < middle
        same_direction = momentum < 0.0
    if above_middle and same_direction:
        position["trend_bars"] = int(position["trend_bars"]) + 1
    else:
        position["trend_bars"] = 0
    return int(position["trend_bars"]) >= required_bars


def update_state(
    *,
    position: dict[str, object],
    candle: Candle,
    middle: float,
    momentum: float,
    realized_vol: float,
    bandwidth: float,
    threshold: float,
    variant: Variant,
) -> None:
    position["hold_bars"] = int(position["hold_bars"]) + 1
    position["mfe_pct"] = max(float(position["mfe_pct"]), favorable_move(str(position["side"]), float(position["entry_price"]), candle))
    if float(position["mfe_pct"]) >= variant.protect_trigger_pct:
        position["state"] = "protect"
        return
    if realized_vol >= variant.fast_vol_pct or bandwidth >= threshold * variant.fast_bandwidth_ratio:
        position["state"] = "fast"
        return
    if trend_continues(position, candle, middle, momentum, variant.trend_momentum_bars):
        position["state"] = "trend"
    else:
        position["state"] = "neutral"


def state_signal_exit(position: dict[str, object], candle: Candle, middle: float, variant: Variant) -> str | None:
    state = str(position["state"])
    if state == "trend":
        buffer_pct = variant.trend_middle_buffer_pct
        confirm_bars = variant.trend_middle_confirm_bars
    elif state == "protect":
        buffer_pct = 0.0
        confirm_bars = variant.protect_middle_confirm_bars
    elif state == "fast":
        buffer_pct = variant.fast_middle_buffer_pct
        confirm_bars = variant.fast_middle_confirm_bars
    else:
        buffer_pct = variant.neutral_middle_buffer_pct
        confirm_bars = variant.neutral_middle_confirm_bars

    if adverse_middle(position, candle, middle, buffer_pct):
        position["middle_exit_streak"] = int(position["middle_exit_streak"]) + 1
    else:
        position["middle_exit_streak"] = 0
    if int(position["middle_exit_streak"]) >= confirm_bars:
        return f"{state}_middle"
    if int(position["hold_bars"]) >= variant.max_hold_bars:
        return "time_exit"
    return None


def run_variant(eth: list[Candle], btc: list[Candle], variant: Variant) -> tuple[SegmentResult, dict[str, int]]:
    eth_close = pd.Series([candle.close for candle in eth], dtype=float)
    btc_close = pd.Series([candle.close for candle in btc], dtype=float)
    middle_series = eth_close.rolling(variant.band_length).mean()
    stdev_series = eth_close.rolling(variant.band_length).std(ddof=0)
    upper_values = middle_series + 2.0 * stdev_series
    lower_values = middle_series - 2.0 * stdev_series
    middle = middle_series.tolist()
    upper = upper_values.tolist()
    lower = lower_values.tolist()
    bandwidth = ((upper_values - lower_values) / middle_series).tolist()
    threshold = pd.Series(bandwidth, dtype=float).rolling(variant.bandwidth_lookback).quantile(variant.bandwidth_quantile).tolist()
    btc_sma = btc_close.rolling(480).mean().tolist()
    btc_momentum = (btc_close / btc_close.shift(96) - 1.0).tolist()
    eth_realized_vol = eth_close.pct_change().rolling(96).std(ddof=0).tolist()
    eth_momentum = (eth_close / eth_close.shift(variant.band_length // 2) - 1.0).tolist()
    warmup_bars = max(variant.band_length, variant.bandwidth_lookback, 480, 96)

    equity = INITIAL_EQUITY
    ending_equity = equity
    peak_equity = equity
    max_drawdown = 0.0
    wins = 0
    trades: list[dict[str, object]] = []
    entries: list[dict[str, object]] = []
    exits: list[dict[str, object]] = []
    equity_curve: list[dict[str, float | int]] = []
    position: dict[str, object] | None = None
    pending_entry_side: str | None = None
    pending_exit_reason: str | None = None
    cooldown_until = -1
    exit_counts: Counter[str] = Counter()

    for index in range(warmup_bars, len(eth)):
        candle = eth[index]
        if pending_exit_reason is not None and position is not None:
            equity, won = close_position(
                trades=trades,
                exits=exits,
                position=position,
                candle=candle,
                exit_price=candle.open,
                reason=pending_exit_reason,
            )
            wins += int(won)
            exit_counts[pending_exit_reason] += 1
            position = None
            pending_exit_reason = None
            cooldown_until = index + variant.cooldown_bars
        if pending_entry_side is not None and position is None and equity > 0.0:
            entry_price = candle.open
            position = {
                "side": pending_entry_side,
                "entry_time": candle.ts,
                "entry_price": entry_price,
                "margin_used": equity,
                "stop_price": entry_price * (1.0 - variant.stop_loss_pct if pending_entry_side == "long" else 1.0 + variant.stop_loss_pct),
                "mfe_pct": 0.0,
                "state": "neutral",
                "trend_bars": 0,
                "middle_exit_streak": 0,
                "hold_bars": 0,
            }
            entries.append({"ts": candle.ts, "price": entry_price, "side": pending_entry_side})
            pending_entry_side = None

        current_equity = equity
        if position is not None:
            risk_exit = hard_stop_exit(position, candle)
            if risk_exit is None:
                values = (middle[index], bandwidth[index], threshold[index], eth_realized_vol[index], eth_momentum[index])
                if not any(value != value for value in values):
                    update_state(
                        position=position,
                        candle=candle,
                        middle=float(middle[index]),
                        momentum=float(eth_momentum[index]),
                        realized_vol=float(eth_realized_vol[index]),
                        bandwidth=float(bandwidth[index]),
                        threshold=float(threshold[index]),
                        variant=variant,
                    )
                    risk_exit = protection_stop_exit(position, candle, variant)
                    if risk_exit is None:
                        signal_reason = state_signal_exit(position, candle, float(middle[index]), variant)
                        if signal_reason is not None:
                            pending_exit_reason = signal_reason
            if risk_exit is not None:
                exit_price, reason = risk_exit
                equity, won = close_position(
                    trades=trades,
                    exits=exits,
                    position=position,
                    candle=candle,
                    exit_price=exit_price,
                    reason=reason,
                )
                wins += int(won)
                exit_counts[reason] += 1
                current_equity = equity
                position = None
                pending_exit_reason = None
                cooldown_until = index + variant.cooldown_bars
        if position is not None:
            current_equity = mark_to_market(
                side=str(position["side"]),
                margin_used=float(position["margin_used"]),
                entry_price=float(position["entry_price"]),
                mark_price=candle.close,
                leverage=LEVERAGE,
            )
        peak_equity = max(peak_equity, current_equity)
        max_drawdown = max(max_drawdown, (peak_equity - current_equity) / peak_equity)
        equity_curve.append({"ts": candle.ts, "equity": current_equity, "close": candle.close})
        ending_equity = current_equity
        if index == len(eth) - 1 or equity <= 0.0:
            continue

        values = (middle[index], upper[index], lower[index], bandwidth[index], threshold[index], btc_sma[index], btc_momentum[index], eth_realized_vol[index])
        if any(value != value for value in values):
            continue
        if position is not None or index < cooldown_until:
            continue
        if variant.eth_vol_cap is not None and float(eth_realized_vol[index]) > variant.eth_vol_cap:
            continue
        if variant.btc_filter == "btc-up" and not (btc_close.iloc[index] > float(btc_sma[index])):
            continue
        if variant.btc_filter == "btc-up-momo" and not (
            btc_close.iloc[index] > float(btc_sma[index]) and float(btc_momentum[index]) > 0.0
        ):
            continue
        if bandwidth[index] <= threshold[index]:
            if candle.close > float(upper[index]):
                pending_entry_side = "long"
            elif variant.side_mode == "both" and candle.close < float(lower[index]):
                pending_entry_side = "short"

    trade_count = len(trades)
    result = SegmentResult(
        trade_count=trade_count,
        total_return=(ending_equity - INITIAL_EQUITY) / INITIAL_EQUITY,
        win_rate=wins / trade_count if trade_count else 0.0,
        max_drawdown=max_drawdown,
        trades=trades,
        open_position=position,
        candles=eth[warmup_bars:],
        equity_curve=equity_curve,
        entries=entries,
        exits=exits,
    )
    return result, dict(exit_counts)


def cost_equity_frame(result: SegmentResult, cost: float) -> pd.DataFrame:
    rows = [{"ts": pd.to_datetime(result.equity_curve[0]["ts"], unit="ms", utc=True), "equity": INITIAL_EQUITY}]
    equity = INITIAL_EQUITY
    for trade in result.trades:
        equity *= 1.0 + float(trade["return_pct"]) / 100.0 - cost * float(trade.get("cost_weight", 1.0))
        rows.append({"ts": pd.to_datetime(int(trade["exit_ts"]), unit="ms", utc=True), "equity": equity})
    return pd.DataFrame(rows)


def max_drawdown(values: list[float]) -> float:
    peak = values[0]
    dd = 0.0
    for value in values:
        peak = max(peak, value)
        dd = max(dd, (peak - value) / peak if peak else 0.0)
    return dd


def equity_metrics(frame: pd.DataFrame, first_ts: int, last_ts: int) -> dict[str, float]:
    years = (last_ts - first_ts) / 86_400_000 / 365
    total_return = float(frame["equity"].iloc[-1] / frame["equity"].iloc[0] - 1.0)
    annualized = (1.0 + total_return) ** (1.0 / years) - 1.0 if total_return > -1.0 and years > 0.0 else 0.0
    dd = max_drawdown([float(value) for value in frame["equity"]])
    return {
        "net_total_return": total_return,
        "net_annualized_return": annualized,
        "net_max_drawdown": dd,
        "net_calmar": annualized / dd if dd else 0.0,
    }


def trade_stats(trades: list[dict[str, object]]) -> dict[str, float | int]:
    if not trades:
        return {
            "trades": 0,
            "win_rate": 0.0,
            "profit_factor": 0.0,
            "payoff_ratio": 0.0,
            "avg_return_pct": 0.0,
            "avg_mfe_pct": 0.0,
            "avg_hold_bars": 0.0,
        }
    returns = [float(trade["return_pct"]) for trade in trades]
    wins = [value for value in returns if value > 0.0]
    losses = [-value for value in returns if value < 0.0]
    return {
        "trades": len(trades),
        "win_rate": len(wins) / len(trades),
        "profit_factor": sum(wins) / sum(losses) if losses else 0.0,
        "payoff_ratio": (sum(wins) / len(wins)) / (sum(losses) / len(losses)) if wins and losses else 0.0,
        "avg_return_pct": sum(returns) / len(returns),
        "avg_mfe_pct": sum(float(trade["mfe_pct"]) for trade in trades) / len(trades),
        "avg_hold_bars": sum(int(trade["hold_bars"]) for trade in trades) / len(trades),
    }


def exit_reason_counts(trades: list[dict[str, object]]) -> dict[str, int]:
    counts = Counter(str(trade["exit_reason"]) for trade in trades)
    return {
        "stop_exits": counts["stop"] + counts["stop_gap"],
        "protect_exits": counts["protect_trail"] + counts["protect_gap"],
        "trend_middle_exits": counts["trend_middle"],
        "protect_middle_exits": counts["protect_middle"],
        "fast_middle_exits": counts["fast_middle"],
        "neutral_middle_exits": counts["neutral_middle"],
        "time_exits": counts["time_exit"],
    }


def horizon_frame(frame: pd.DataFrame, first_ts: int, last_ts: int, offset: pd.DateOffset | None) -> tuple[pd.DataFrame, int, str, str]:
    end_time = pd.to_datetime(last_ts, unit="ms", utc=True)
    if offset is None:
        start_time = pd.to_datetime(first_ts, unit="ms", utc=True)
        return frame[["ts", "equity"]].copy(), first_ts, start_time.strftime("%Y-%m-%d %H:%M"), end_time.strftime("%Y-%m-%d %H:%M")
    cutoff = end_time - offset
    before = frame[frame["ts"] <= cutoff]
    if len(before):
        start_equity = float(before["equity"].iloc[-1])
        after = frame[frame["ts"] > cutoff]
        out = pd.concat([pd.DataFrame([{"ts": cutoff, "equity": start_equity}]), after[["ts", "equity"]]], ignore_index=True)
    else:
        out = frame[["ts", "equity"]].copy()
        cutoff = pd.Timestamp(out["ts"].iloc[0])
    return out, int(cutoff.timestamp() * 1000), cutoff.strftime("%Y-%m-%d %H:%M"), end_time.strftime("%Y-%m-%d %H:%M")


def horizon_rows(result: SegmentResult, frame: pd.DataFrame, first_ts: int, last_ts: int) -> list[dict[str, object]]:
    rows: list[dict[str, object]] = []
    for label, offset in HORIZONS:
        sliced_frame, start_ts, start_text, end_text = horizon_frame(frame, first_ts, last_ts, offset)
        trades = [trade for trade in result.trades if int(trade["exit_ts"]) >= start_ts]
        rows.append(
            {
                "horizon": label,
                "horizon_start": start_text,
                "horizon_end": end_text,
                **equity_metrics(sliced_frame, start_ts, last_ts),
                **trade_stats(trades),
                **exit_reason_counts(trades),
            }
        )
    return rows


def worst_month(frame: pd.DataFrame) -> tuple[str, float]:
    monthly = frame.set_index("ts")["equity"].resample("ME").last().ffill().pct_change().dropna()
    if not len(monthly):
        return "", 0.0
    idx = monthly.idxmin()
    return idx.strftime("%Y-%m"), float(monthly.loc[idx])


def build_variants() -> list[Variant]:
    bases = (
        (48, 960, 0.25, "both", "none", 0.006, 0.010),
        (48, 960, 0.25, "both", "none", 0.006, 0.012),
        (96, 960, 0.25, "both", "btc-up", 0.006, 0.012),
        (96, 960, 0.25, "both", "btc-up-momo", 0.006, 0.012),
        (96, 480, 0.15, "both", "none", 0.006, 0.010),
        (48, 960, 0.25, "long", "btc-up", 0.006, 0.010),
    )
    exits = (
        (2, 0.0015, 2, 0.0005, 1, 0.006, 0.0000, 0.006, 1, 0.0070, 1.40, 0.0, 1, 192),
        (2, 0.0025, 3, 0.0010, 1, 0.008, 0.0005, 0.007, 1, 0.0065, 1.30, 0.0, 1, 288),
        (3, 0.0035, 3, 0.0010, 2, 0.010, 0.0010, 0.008, 2, 0.0060, 1.25, 0.0, 1, 384),
        (3, 0.0045, 4, 0.0015, 2, 0.012, 0.0010, 0.010, 2, 0.0060, 1.20, 0.0, 1, 480),
        (2, 0.0020, 2, 0.0000, 1, 0.006, 0.0005, 0.005, 1, 0.0055, 1.15, 0.0, 1, 192),
        (4, 0.0050, 4, 0.0015, 2, 0.015, 0.0015, 0.012, 2, 0.0065, 1.35, 0.0, 1, 576),
    )
    variants: list[Variant] = []
    for band_length, lookback, quantile, side_mode, btc_filter, vol_cap, stop_loss in bases:
        for exit_spec in exits:
            variants.append(
                Variant(
                    band_length=band_length,
                    bandwidth_lookback=lookback,
                    bandwidth_quantile=quantile,
                    side_mode=side_mode,
                    btc_filter=btc_filter,
                    eth_vol_cap=vol_cap,
                    cooldown_bars=24,
                    stop_loss_pct=stop_loss,
                    trend_momentum_bars=exit_spec[0],
                    trend_middle_buffer_pct=exit_spec[1],
                    trend_middle_confirm_bars=exit_spec[2],
                    neutral_middle_buffer_pct=exit_spec[3],
                    neutral_middle_confirm_bars=exit_spec[4],
                    protect_trigger_pct=exit_spec[5],
                    protect_lock_pct=exit_spec[6],
                    protect_trail_giveback_pct=exit_spec[7],
                    protect_middle_confirm_bars=exit_spec[8],
                    fast_vol_pct=exit_spec[9],
                    fast_bandwidth_ratio=exit_spec[10],
                    fast_middle_buffer_pct=exit_spec[11],
                    fast_middle_confirm_bars=exit_spec[12],
                    max_hold_bars=exit_spec[13],
                )
            )
    return variants


def format_cell(value: object) -> str:
    if isinstance(value, float):
        return f"{value:.6g}"
    return str(value).replace("|", "\\|")


def markdown_table(frame: pd.DataFrame) -> str:
    columns = list(frame.columns)
    rows = [columns, ["---" for _ in columns]]
    for record in frame.to_dict("records"):
        rows.append([record[column] for column in columns])
    return "\n".join("| " + " | ".join(format_cell(value) for value in row) + " |" for row in rows)


def write_report(*, summary: pd.DataFrame, horizon: pd.DataFrame, first_ts: int, last_ts: int, requested_years: float, command: str) -> str:
    primary = summary[summary["cost"] == PRIMARY_COST]
    top = primary.head(10)
    horizon_top = (
        horizon[horizon["cost"] == PRIMARY_COST]
        .sort_values(["horizon", "net_calmar", "net_annualized_return"], ascending=[True, False, False])
        .groupby("horizon", observed=True)
        .head(3)
    )
    lines = [
        "# ETH BB squeeze adaptive state-exit exploration",
        "",
        f"Run command: `{command}`",
        f"Requested years: {requested_years:g}",
        f"Actual continuous local history: `{_format_ts(first_ts)}` to `{_format_ts(last_ts)}`.",
        "",
        "State machine:",
        "- `trend`: ETH move remains aligned with the breakout and close stays on the favorable side of the BB middle; middle exit uses a wider buffer and more confirmations.",
        "- `protect`: trade reaches floating-profit trigger; a lock/trailing stop and faster middle exit protect profit.",
        "- `fast`: realized volatility or bandwidth expansion crosses the variant threshold; middle exit is immediate.",
        "- `neutral`: default state, between trend continuation and protection.",
        "",
        "Top 10 by maker_taker full-period Calmar:",
        markdown_table(
            top[
                [
                    "name",
                    "trades",
                    "win_rate",
                    "net_total_return",
                    "net_annualized_return",
                    "net_max_drawdown",
                    "net_calmar",
                    "profit_factor",
                    "payoff_ratio",
                    "avg_return_pct",
                    "avg_mfe_pct",
                    "protect_exits",
                    "trend_middle_exits",
                    "fast_middle_exits",
                    "neutral_middle_exits",
                ]
            ]
        ),
        "",
        "Horizon leaders:",
        markdown_table(
            horizon_top[
                [
                    "horizon",
                    "name",
                    "trades",
                    "win_rate",
                    "net_total_return",
                    "net_annualized_return",
                    "net_max_drawdown",
                    "net_calmar",
                    "profit_factor",
                    "payoff_ratio",
                    "protect_exits",
                    "fast_middle_exits",
                ]
            ]
        ),
    ]
    return "\n".join(lines) + "\n"


def main() -> int:
    parser = argparse.ArgumentParser()
    parser.add_argument("--bar", default=BAR)
    parser.add_argument("--years", type=float, default=YEARS)
    parser.add_argument("--output-dir", type=Path, default=OUTPUT_DIR)
    args = parser.parse_args()

    eth = _load_candles(ETH_SYMBOL, args.bar)
    btc = _load_candles(BTC_SYMBOL, args.bar)
    eth, btc = _align_pair(eth, btc)
    requested_bars = int(args.years * 365 * 24 * 60 / 15)
    eth = eth[-requested_bars:]
    btc = btc[-requested_bars:]

    summary_rows: list[dict[str, object]] = []
    horizon_rows_out: list[dict[str, object]] = []
    variants = build_variants()
    for index, variant in enumerate(variants, start=1):
        result, exit_counts = run_variant(eth, btc, variant)
        if not result.equity_curve:
            print(f"skip {index}/{len(variants)} {variant.name}", flush=True)
            continue
        stats = trade_stats(result.trades)
        reason_counts = exit_reason_counts(result.trades)
        for cost_name, cost in COSTS:
            frame = cost_equity_frame(result, cost)
            metrics = equity_metrics(frame, eth[0].ts, eth[-1].ts)
            month, month_return = worst_month(frame)
            row = {
                "family": "eth_adaptive_state_exit",
                "cost": cost_name,
                "symbol": ETH_SYMBOL,
                "signal_symbol": BTC_SYMBOL if variant.btc_filter != "none" else "",
                "bar": args.bar,
                "name": variant.name,
                "band_length": variant.band_length,
                "bandwidth_lookback": variant.bandwidth_lookback,
                "bandwidth_quantile": variant.bandwidth_quantile,
                "side_mode": variant.side_mode,
                "btc_filter": variant.btc_filter,
                "eth_vol_cap": variant.eth_vol_cap,
                "cooldown_bars": variant.cooldown_bars,
                "stop_loss_pct": variant.stop_loss_pct,
                "trend_momentum_bars": variant.trend_momentum_bars,
                "trend_middle_buffer_pct": variant.trend_middle_buffer_pct,
                "trend_middle_confirm_bars": variant.trend_middle_confirm_bars,
                "neutral_middle_buffer_pct": variant.neutral_middle_buffer_pct,
                "neutral_middle_confirm_bars": variant.neutral_middle_confirm_bars,
                "protect_trigger_pct": variant.protect_trigger_pct,
                "protect_lock_pct": variant.protect_lock_pct,
                "protect_trail_giveback_pct": variant.protect_trail_giveback_pct,
                "protect_middle_confirm_bars": variant.protect_middle_confirm_bars,
                "fast_vol_pct": variant.fast_vol_pct,
                "fast_bandwidth_ratio": variant.fast_bandwidth_ratio,
                "fast_middle_buffer_pct": variant.fast_middle_buffer_pct,
                "fast_middle_confirm_bars": variant.fast_middle_confirm_bars,
                "max_hold_bars": variant.max_hold_bars,
                "first_candle": _format_ts(eth[0].ts),
                "last_candle": _format_ts(eth[-1].ts),
                "years": (eth[-1].ts - eth[0].ts) / 86_400_000 / 365,
                "gross_total_return": result.total_return,
                "gross_max_drawdown_mark_to_market": result.max_drawdown,
                "worst_month": month,
                "worst_month_return": month_return,
                **stats,
                **reason_counts,
                **metrics,
            }
            summary_rows.append(row)
            for horizon_row in horizon_rows(result, frame, eth[0].ts, eth[-1].ts):
                horizon_rows_out.append(
                    {
                        "family": "eth_adaptive_state_exit",
                        "cost": cost_name,
                        "symbol": ETH_SYMBOL,
                        "signal_symbol": BTC_SYMBOL if variant.btc_filter != "none" else "",
                        "bar": args.bar,
                        "name": variant.name,
                        **horizon_row,
                    }
                )
        print(f"done {index}/{len(variants)} {variant.name} exits={exit_counts}", flush=True)

    summary = pd.DataFrame(summary_rows).sort_values(
        ["cost", "net_calmar", "net_annualized_return", "profit_factor"],
        ascending=[True, False, False, False],
    )
    primary = summary[summary["cost"] == PRIMARY_COST]
    summary = pd.concat([primary, summary[summary["cost"] != PRIMARY_COST]], ignore_index=True)
    horizon = pd.DataFrame(horizon_rows_out)
    horizon["horizon"] = pd.Categorical(horizon["horizon"], categories=[label for label, _ in HORIZONS], ordered=True)
    horizon = horizon.sort_values(["cost", "horizon", "net_calmar", "net_annualized_return"], ascending=[True, True, False, False])

    args.output_dir.mkdir(parents=True, exist_ok=True)
    summary_path = args.output_dir / "eth-adaptive-state-exit-summary.csv"
    horizon_path = args.output_dir / "eth-adaptive-state-exit-horizon.csv"
    report_path = args.output_dir / "eth-adaptive-state-exit-report.md"
    summary.to_csv(summary_path, index=False)
    horizon.to_csv(horizon_path, index=False)
    command = f"rtk .venv/bin/python {Path(__file__).as_posix()} --bar {args.bar} --years {args.years}"
    report_path.write_text(
        write_report(summary=summary, horizon=horizon, first_ts=eth[0].ts, last_ts=eth[-1].ts, requested_years=args.years, command=command),
        encoding="utf-8",
    )
    print(primary.head(10).to_string(index=False))
    return 0


if __name__ == "__main__":
    raise SystemExit(main())