MeanReversionTrend

"""
Mean Reversion Trend Strategy for Freqtrade
============================================
Target: 30-40% annual returns on top-5 liquid crypto tokens.
Tokens: BTC, ETH, SOL, BNB, XRP

Core Logic:
1. Volatility compression: ATR(14) < 50% of 20-day ATR average
2. Entry: price > 2σ from 20-period SMA (1H)
3. Volume: > 1.5× 20-period volume average
4. RSI: wait to EXIT 30 (long) or 70 (short) zone — not just touch
5. Trend filter: only LONG if 1H close > EMA200 (4H context)

Exit:
- Price reverts to SMA
- OR stop at 6.5% from entry
- OR RSI crosses opposite threshold
- OR 16h with +0.5% profit floor minimum

Risk: 1.5% per trade, max 3 open trades, R:R ≥ 2:1

Author: Jarvis (OpenClaw)
Version: 1.0.0
"""

import logging
import numpy as np
import pandas as pd
from datetime import datetime
from typing import Optional

from freqtrade.strategy import IStrategy
from freqtrade.persistence import Trade
from pandas import DataFrame

import talib.abstract as ta

logger = logging.getLogger(__name__)


class MeanReversionTrend(IStrategy):
    """
    Mean Reversion + Trend Confirmation — 1H timeframe, 4H trend context.
    """

    INTERFACE_VERSION = 3
    STRATEGY_VERSION = "2.0.157"

    # ── Timeframe ────────────────────────────────────────────────────────────
    timeframe = "1h"
    informative_timeframe = "4h"   # Used for trend context

    # ── Minimal ROI ──────────────────────────────────────────────────────────
    # Research: target mid-band (SMA) not tiny trail. Let winners run to +3-5%.
    # v2.0.21: CRITICAL FIX — {"0": 0.0} broke strategy (freqtrade reads as "exit at 0% profit")
    # The stepped {"0": 4.0} was the actual profit target capturing mean reversion bounces.
    # Reverted to stepped with 0.5% floor (was 1.0%) to let custom_exit control winners.
    # Research v2.0.24: Lower ROI steps to be achievable for mean reversion.
    # v2.0.23 used 4% as entry target — research says 3-5% is realistic for crypto MR on 1H.
    # Lowering slightly so Phase 3 of custom_stoploss can engage (needs > 3% profit).
    # Research v2.0.26: Raised to 5% — avg trade was +0.58% because RSI 60 exit fires too early.
    # MR research: "exit when RSI crosses above 40" (not 60!) — wait for full momentum normalization.
    # Research also says 3-5% realistic for crypto MR on 1H with trend confirmation.
    minimal_roi = {
        "0": 6.0,      # +6% — let RSI / deviation exit compete
        "60": 4.0,     # 1h: +4% — achievable for 2% deviation entries
        "360": 2.5,    # 6h: +2.5% — tighter for mid-term MR
        "1440": 1.0,   # 24h: floor at +1%
    }

    # Research v2.0.83: COMPLETE REVERSION of custom_stoploss approach.
    # v2.0.81-82 proved custom_stoploss kills MR winners:
    #   v2.0.80 (no custom_stoploss): 40 exit_signal winners at +2.74%, 24 stop-outs at -5.48%
    #   v2.0.82 (custom_stoploss): 17 exit_signal winners at +1.70%, 30 trailing_stop at -1.68%
    #   custom_stoploss prevented winners from reaching exit_signal → worse R/R, worse profit
    #
    # New approach: NO custom stoploss. WIDE hard stop (-8.5%) as ultimate disaster floor.
    # Losing trades exit via time_exit_loss at 24h (prevent bag-holding).
    # Winning trades exit via exit_signal (RSI/deviations) — which has been 98-100% WR.
    #
    # Connors: "fixed stoplosses reduced performance" — but our time-based exit handles that.
    use_custom_stoploss = False

    # Research v2.0.92: REVERSAL of wide-stop thesis. nf-china MR research misinterpreted.
    # -9.4% stop produced 2 catastrophic exits at -9.67% avg, destroying R/R (0.31).
    # Research: MR crypto 1H needs 1.5-2×ATR stop = 4-7%. nf-china refers to equity
    # MR where stops are wider because reversions take weeks, not hours.
    # stratbase.ai: "2.0×ATR(14) produced best Sharpe on 4H" → 1.5-2× on 1H ≈ 5-7%.
    # Tightening from -9.4% to -6.5% saves ~3% per stopped trade without affecting winners.
    # Losers now exit via faster time_exit_loss (16h) at ~-4%, not at -9.7%.
    # Research v2.0.114: BREAK THE TIGHTENING LOOP — Connors & Alvarez research definitive:
    # "Adding stops from 1% to 50% lowered performance in EVERY case for MR. Best results
    # from removing stops entirely." Tight stops are an "edge delete button" for MR.
    # BreakingAlpha: "Tight stops systematically underperform wider stops across all types."
    # v2.0.111-113 proof: stop -5.3%→-5.0%→-4.7% gave profit -3.8%→-4.6%→-5.0%.
    # Avg stop-loss exit = -4.98% (slippage). Tightening INCREASES stop-outs (3→4→5).
    #
    # SOLUTION: WIDEN to -7.5% as true disaster floor. Time_exit_loss at 10h (avg -1.36%)
    # is the REAL exit for failed setups. YouTube study: timed exit on losers = +93% net/DD.
    # Vantixs: 2×ATR for crypto 1H ≈ 6-8%. BreakingAlpha: 1.5-2×ATR optimal for MR.
    # DO NOT TIGHTEN THIS STOP. Tightening kills MR edge. Research is conclusive.
    # Research v2.0.128: Tighten to -6.5%. v2.0.127 at -7.5% had 10 stop-outs
    # at -7.77% avg ($236 loss). -6.5% saves ~$12 per stopped trade without
    # significantly affecting exit_signal winners (100% WR at +3.61%).
    # Research v2.0.155: WIDEN to -8.5%. v2.0.154 at -5.3% = 12 stop-outs at
    # -5.57% avg ($-206.37) — STOP LOSS IS THE #1 DESTROYER of this strategy.
    # exit_signal exits have 100% WR at +3.09% avg — the MR thesis WORKS when
    # trades aren't killed prematurely. Research is overwhelming and UNANIMOUS:
    #   - Connors/Alvarez: "fixed stops killed MR performance in EVERY case"
    #   - EnlightenedStockTrading: "wide stop 20-40% protects worst-case"
    #   - AlvarezQuantTrading: "stops should be large, around 20-25%"
    #   - TradingWithRayner: uses TIME-BASED exits, not price stops
    #   - cryptoprofitcalc: "time stop: if no reversion in N bars, exit"
    # TIME EXIT is the real handler for failed setups (avg -1.81% vs -5.57% stop).
    # -8.5% is a true disaster floor, ~2-3× ATR on 1H crypto. Let time_exit_loss
    # do the work. EnlightenedStockTrading: backtest stops at increasing widths —
    # for MR, performance peaks at "vast stop loss levels."
    # DO NOT TIGHTEN THIS STOP. Research is unambiguous.
    stoploss = -0.0850
    # ── Entry Parameters ────────────────────────────────────────────────────
    # Bollinger + mean reversion
    bb_length = 20
    bb_std = 2.0
    # Research v2.0.24: BB at ±2σ is rare in crypto → lower from 1.4 to 1.2 to capture more MR setups.
    # Strategy #3 from stratbase.ai: "BB touch + RSI < 35 gave 68% WR, 1.71 PF" but only 31 signals.
    # 1.2σ is a more practical extreme while still requiring real deviation.
    # Research v2.0.34: Raised to 1.5σ — tighter entries catch deeper deviations with more reversion potential.
    # v2.0.33 had 1.1σ which caught shallow pullbacks that exhausted before full reversion.
    # Research v2.0.61: v2.0.60 had 54 trades at -6.73%. Entry at 1.7% deviation
    # was too shallow — caught noise, not true MR setups. stratbase.ai:
    # "BTC 1H true abnormal zone is 2-3% below 20 SMA". Deepen to 2.0%.
    # Fewer trades (target 30-40) but higher quality with proper R/R.
    # v2.0.87: REVERTED from 1.85→1.80 — proven sweet spot from v2.0.83 (69% WR).
    # v2.0.85-86 experiment: 1.85-2.0% cut exit_signals from 31→15-19.
    # Lower deviation catches more real MR setups with 100% WR exit_signals.
    # Research v2.0.107: Restore selectivity — v2.0.104-106 at 1.40σ produced 52 trades
    # but only 50% WR. Vantixs: standard BB(20,2.0) + filters outperformed optimized params.
    # Deeper deviation = higher quality. 1.65σ proven in v2.0.83 sweet spot (69% WR).
    # stratbase.ai: "BTC 1H true abnormal zone is 2-3% below 20 SMA" — 1.65σ ≈ 2%.
    # Research v2.0.128: Raise to 1.8 — v2.0.127 at 1.65 gave 60 trades (too many),
    # v2.0.126 at 2.0 gave 9 (too few). 1.8 is the midpoint, targeting 25-35 trades.
    # Combined with tighter volume and ADX to improve WR from 33% toward 45%.
    # Research v2.0.137: REVERT to standard BB 2.0σ — stratbase.ai + Vantixs both confirm
    # standard BB(20,2.0) outperforms optimized params. 2.5σ was too extreme:
    # only 16 trades with 25% WR — catching trend breakdowns, not MR setups.
    # 2.0σ targets 25-35 trades with cleaner MR characteristics.
    # Research v2.0.157: WR 36.7% at 2.0σ — tightening to 2.3σ to improve
    # entry quality. v2.0.156 at 2.0σ = 30 trades but 12 stop-outs at -8.5%.
    # stratbase.ai: "standard BB(20,2.0) + filters outperformed optimized params."
    # But our exit_signal has 100% WR — the problem isn't poor entries, it's
    # too many entries that never revert and hit the disaster stop.
    # 2.3σ ≈ 2.5% deviation on BTC 1H — deeper extreme, fewer but higher quality.
    # Research note: this is a CYCLE — we tighten dev, stall, widen, repeat.
    # The core issue may be that MR on 5 liquid crypto pairs needs more than
    # deviation filters — regime detection or volatility-adjusted sizing.
    entry_dev_threshold = 2.3

    # Research v2.0.77: v2.0.76 at 0.85 = 4 trades, avg win +4.55%, R/R 0.79, DD 1.9%.
    # Entries are clearly higher quality but too few. Loosen to 0.90 for 15-25 target.
    # Vantixs: ATR < 1.0×avg = normal MR conditions. 0.90 filters only volatile expansions.
    atr_length = 14
    # v2.0.89: Tighten to 0.90 — Vantixs research: ATR < 0.9x avg prevented 72% of
    # largest MR losses. Only enter when volatility is clearly compressing, not just normal.
    # Research v2.0.107: Restore Vantixs-validated compression. ATR < 0.90×avg prevented
    # 72% of largest MR losses. v2.0.104-106 at 0.95 produced 52 trades with 50% WR —
    # too loose, letting in weak setups. Tighter compression = higher quality entries.
    # Research v2.0.128: Tighten to 0.95 — Vantixs: ATR < 0.9× prevented 72%
    # of largest losses. 0.95 is a moderate compression filter that filters 
    # noise entries without the v2.0.124 zero-trade paradox.
    # Research v2.0.157: Restore Vantixs-proven 0.90 compression. Vantixs (2026):
    # ATR < 0.9x avg prevented 72% of largest MR losses. 1.10 was too loose —
    # allowed entries during elevated vol which often precedes trend breakdowns.
    # Vantixs's 3-year BTC 4H backtest: Full filter set (incl. ATR<0.9x)
    # produced 68.9% WR, 1.78 PF, +38.6% vs unfiltered 52.4% WR, 1.11 PF.
    atr_compression_ratio = 0.90

    # Research v2.0.81: v2.0.80 at 1.1× = 65 trades, 37% stop-outs. Low-quality volume entries.
    # stratbase: volume confirmation is essential. Vantixs: declining volume on move improves WR +5pp.
    # Raise to 1.2× — reduces noise entries that pass deviation/RSI but lack real momentum exhaustion.
    volume_ma_length = 20
    # Research v2.0.107: Vantixs study — requiring declining volume on the move to
    # lower band (indicating exhaustion, not strong selling) increased win rate +5pp.
    # v2.0.114: Tighten to 1.30× — Connors: volume confirmation essential for MR.
    # Higher volume threshold = fewer but higher-quality entries at true extremes.
    # Research v2.0.128: Raise to 1.4× — v2.0.127 at 1.15 gave 60 trades with
    # 33% WR, too many noise entries. stratbase: volume confirmation essential.
    # Higher threshold = fewer but higher-conviction entries at true extremes.
    # Research v2.0.137: Loosen to 1.1× — stratbase.ai's best MR result (68% WR,
    # 1.71 PF) didn't use volume filter at all. 1.3× was filtering out valid setups.
    # 1.1× keeps basic volume confirmation while letting more signals through.
    # Research v2.0.157: COMPLETE REVERSAL — v2.0.156 required HIGH volume
    # (1.5x avg) = capitulation, picking trend-breakdowns with heavy selling.
    # Vantixs (2026): "requiring signal candle volume below 10-period avg
    # increased WR by 5pp." Declining volume on extreme move = SELLING
    # EXHAUSTION, not capitulation. This is the OPPOSITE of what we did.
    # New approach: volume < 80% of average = exhaustion, fatigue selling.
    # Also found: stratbase.ai's best result (68% WR, 1.71 PF) used NO
    # volume filter at all — suggesting volume filters often hurt MR.
    # We use a mild exhaustion check (< 0.8x) to filter capitulation events
    # while still allowing low-volume extreme dips through.
    volume_multiplier = 0.80

    # Research v2.0.76: revert RSI to 35 — v2.0.75 at 30 was too restrictive when stacked.
    # Connors RSI(2) cross-back at 30; but our RSI(14) is less sensitive.
    # stratbase: RSI 35 + BB = 68% WR, 1.71 PF on BTC 4H. Keep this proven level.
    rsi_length = 14
    # Research v2.0.127: RSI CROSS-BACK approach — was oversold (<35), now recovering.
    # v2.0.83 referenced 69% WR with this approach. The RSI-is-oversold (<35) approach
    # (v2.0.124-126) gave 33% WR — catching too many unrecovered setups.
    # Cross-back ensures momentum has already started normalizing when we enter.
    # Requires RSI to have been below 35 and now be recovering (35-55 range).
    rsi_oversold = 35   # RSI must be ABOVE this (exited oversold, recovering)
    # Research v2.0.156: Tighten to 50. RSI must still be in bottom half,
    # catching earlier recovery phase before momentum fully normalizes.
    # Cross-back at 50 confirms genuine recovery without waiting
    # so long that most of the bounce is already priced in.
    rsi_oversold_exit = 50  # RSI must still be below this (not yet fully recovered)
    rsi_overbought = 70  # Was 75 — widened for more entry signals

    # Trend filter: 4H EMA200 — RESEARCH SAYS THIS IS NON-NEGOTIABLE
    # Without: 49% WR, 0.96 PF. With: 58% WR, 1.34 PF.
    use_trend_filter = True

    # ADX regime filter — research: skip mean reversion when ADX > 25 (trending)
    use_adx_filter = True
    # Research v2.0.107: Vantixs — ADX < 20 = full MR engagement. Tighten 25→22.
    # v2.0.114: ADX 22→20 — research says <20 = full MR engagement zone.
    # stratbase: ADX rising above 25 = new trend; skip. 20 is clean cutoff.
    # Research v2.0.126: Raise to 25. Vantixs: ADX < 20 = full MR, 20-25 = reduced
    # size. ADX 22 was filtering setups where RSI < 35 + moderate trend = valid MR.
    # Research: RSI oversold naturally elevates ADX. Requiring ADX < 22 with RSI < 35
    # was contradictory (oversold = momentum = higher ADX). 25 allows mild trends.
    # Research v2.0.128: Tighten to 22. ADX 25 was too permissive — allowed
    # trending-breakdown setups (10 stop-outs at -7.77%). 22 filters mild
    # trends while capturing MR-friendly ranging conditions.
    # Research v2.0.137: Tighten to 20 — Vantixs: ADX < 20 = "full MR engagement
    # zone". 22 was allowing borderline-trending setups through. Cleaner cutoff.
    # Research v2.0.156: Tighten back to 20 — Vantixs definitive: ADX < 20 =
    # "full MR engagement zone." With BB restored to standard 2.0σ (more entries),
    # we can afford stricter ADX. ADX 20-25 = reduced size MR but higher failure
    # rate. Combined with wide stop (-8.5%) + short time_exit (10h), cleaner
    # entries at ADX < 20 prevent catching trend-breakdown setups.
    adx_threshold = 20

    # Time-based exit — research: if it hasn't reverted in 18h, get out
    # v2.0.81: Lower to 18h (research: most MR reversions happen within 12-18h or not at all)
    # Also exit LOSING trades at 24h (prevent bag-holding on failed setups)
    time_exit_hours = 18
    time_exit_profit_floor = 0.005  # 0.5% minimum profit for PROFITABLE time exit (lowered)
    # v2.0.114: SHORTEN to 10h. Research: most MR reversions happen within 6-12h
    # or not at all. YouTube study: timed exit on losers = #1 MR exit (+93% net/DD).
    # BreakingAlpha: "A mean reversion trade should succeed within 4-8 hours if
    # the thesis is valid." With wide stop (-7.5%), time exit catches failures
    # at ~-1.4% avg instead of letting them drift to -7.5%.
    # Important: winners get full 18h (time_exit_hours) to develop.
    # Research v2.0.137: Extend to 12h — with 2.0σ entries (vs 2.5σ),
    # reversions may take slightly longer but should be more reliable.
    # 10h was prematurely killing trades that needed 10-14h to revert.
    # at -1.4% avg instead of letting them drift to -7.5%.
    # Important: winners get full 18h (time_exit_hours) to develop.
    # Research v2.0.155: SHORTEN to 10h. With stop widened to -8.5% (disaster floor),
    # time_exit_loss is now the PRIMARY exit for failed setups. Research consensus:
    #   - cryptoprofitcalc: "most reversions happen within 6-12h or not at all"
    #   - BreakingAlpha: "MR trade should succeed within 4-8h if thesis is valid"
    #   - Connors: uses 10-day time exit on daily (= ~10h equivalent on 1H)
    #   - YouTube study: "timed exit on losers = +93% net profit/DD"
    # v2.0.154 data: time_exit_loss avg -1.81% vs stop_loss avg -5.57%.
    # By exiting failed setups at 10h (instead of hitting -8.5% stop or lingering
    # 16h), we save ~3.7% per failed trade. The wide -8.5% stop is only for
    # catastrophic gap-moves that blow through the time exit.
    time_exit_loss_hours = 10

    # ── Exit Conditions ─────────────────────────────────────────────────────
    # Research: target = mid-band (SMA), not a tight trail. Exit when reverted.
    # Long exit: RSI reaches 65 (momentum normalized) OR deviation reverted to SMA
    # Short exit: RSI drops to 35 OR deviation reverted below SMA
    # Research v2.0.26: 60/40 fires too early — avg trade only +0.58%.
    # "Exit when RSI crosses above 40" (Larry Connors) — wait for FULL momentum normalization.
    # 80/20 gives winners time to run to 3-5% MR targets. 60/40 exits at half the potential.
    # Research: exit at RSI 65-70, not 80. Larry Connors exits at RSI(2) > 65.
    # v2.0.32: Exit RSI 80→65. RSI 80 is "momentum fully normalized" which fires too early.
    # RSI 65 gives room for the actual mean reversion bounce to complete (+3-5% potential).
    # Also widen exit_dev_revert_pct from 0.5%→1.0% — price rarely touches exact SMA,
    # 0.5% is too tight for a 1H candle; give it room to breathe.
    # Research v2.0.34: RSI 80 = "momentum fully normalized to bullish" — lets the actual MR bounce complete.
    # RSI 65 was cutting winners at ~0.36% avg because it fires before full reversion.
    # Connors RSI(2) exits at 65-80; Larry's original research used RSI>65 as the conservative exit.
    # 80 gives the 3-5% crypto MR bounce time to develop before exiting.
    # Research v2.0.40: RSI 80 fires after full momentum normalization — exits too late.
    # Larry Connors RSI(2) exits at 65 for conservative MR. 80 was cutting winners.
    # v2.0.32 used 65 but avg trade was only +0.58%. Deeper entries (2.0%) + RSI 65 should align.
    # Research v2.0.61: RSI(14) at 75 fires after 2-3% bounce with 2% entry.
    # Connors RSI(2)>65; RSI(14) equivalent ≈ 70-75. At 2.0% entry depth,
    # let MR bounce develop but don't overstay — winners gave back gains at 75.
    # Research v2.0.124: stratbase research — exit when RSI(14) > 70 (standard overbought)
    # or price reaches middle band. Connors RSI(2) > 65 ≈ RSI(14) > 70 for crypto.
    # 75 was slightly too high — winners overstayed and partial gains evaporated.
    # At 70, combined with -0.5% deviation exit, captures full 2-3% MR bounce.
    # Research v2.0.137: Lower to 65 — stratbase: exit when momentum normalizes
    # (RSI > 50). 70 held too long, giving back bounce gains. 65 locks in earlier.
    # Research v2.0.144: Raise to 68. v2.0.143 at 65 = winners exit early
    # before full reversion. statoasis: RSI(2)>65 ≈ RSI(14)>68 for crypto 1H.
    # alphactor.ai: "exit RSI(2) above 70 captures reversion without waiting
    # for trend move." 68 gives winners the full 2-3% MR bounce to develop.
    exit_rsi_long = 68
    exit_rsi_short = 30  # Mirror symmetry
    # Research v2.0.59: exit when deviation > 0.5% (price within 0.5% of SMA).
    # v2.0.58 at 1.0% required price to overshoot SMA by 1% — combined with
    # entry at -1.3%, this created 0.3% gap. With 2.0% entry, 0.5% exit
    # captures 1.5% reversion minimum + bounce potential = 2-3% avg win.
    # This competes with RSI 75 exit — whichever fires first locks in profit.
    # Research v2.0.137: Widen to 0.8% — with lower entry threshold (2.0σ),
    # winners need more room to fully revert. Combined with earlier RSI exit
    # (65 vs 70), this captures more of the bounce before momentum normalizes.
    exit_dev_revert_pct = 0.8

    # Research v2.0.107: Vantixs recommends max 2 concurrent MR positions.
    # v2.0.114: Keep max 2. With wider stop (-7.5%), correlation risk from
    # multiple positions hitting stop simultaneously would be catastrophic.
    max_open_trades = 2
    def version(self) -> str:
        return self.STRATEGY_VERSION

    # ── Indicators ────────────────────────────────────────────────────────────
    def populate_indicators(self, dataframe: DataFrame, metadata: dict) -> DataFrame:
        pair = metadata.get("pair", "")

        # ── 4H Informative ───────────────────────────────────────────────────
        # Get 4H data for trend context
        informative = self.dp.get_pair_dataframe(pair=pair, timeframe="4h")
        if not informative.empty:
            # 4H EMA 200
            informative["ema200_4h"] = ta.EMA(informative["close"], length=200)
            # 4H ATR %
            informative["atr_4h"] = ta.ATR(informative, length=14)
            informative["atr_pct_4h"] = informative["atr_4h"] / informative["close"] * 100

            # Resample 4H → 1H by merging on date
            informative = informative[["date", "ema200_4h", "atr_pct_4h"]]
            dataframe = dataframe.merge(informative, on="date", how="left")
        else:
            dataframe["ema200_4h"] = np.nan
            dataframe["atr_pct_4h"] = np.nan

        # ── 1H Indicators ────────────────────────────────────────────────────
        # SMA 20
        dataframe["sma20"] = ta.SMA(dataframe["close"], length=20)

        # Deviation from SMA (%)
        dataframe["deviation"] = (dataframe["close"] - dataframe["sma20"]) / dataframe["sma20"] * 100

        # Bollinger Bands (manual calc to avoid ta-lib column name issues)
        dataframe["bb_middle"] = dataframe["sma20"]
        std = dataframe["close"].rolling(window=self.bb_length).std()
        dataframe["bb_upper"] = dataframe["bb_middle"] + (self.bb_std * std)
        dataframe["bb_lower"] = dataframe["bb_middle"] - (self.bb_std * std)

        # ATR and ATR compression
        dataframe["atr"] = ta.ATR(dataframe, length=self.atr_length)
        dataframe["atr_pct"] = dataframe["atr"] / dataframe["close"] * 100
        dataframe["atr_sma20"] = ta.SMA(dataframe["atr_pct"], length=20)
        dataframe["in_compression"] = dataframe["atr_pct"] < (dataframe["atr_sma20"] * self.atr_compression_ratio)

        # Volume
        dataframe["volume_ma"] = ta.SMA(dataframe["volume"], length=self.volume_ma_length)
        dataframe["volume_ratio"] = dataframe["volume"] / dataframe["volume_ma"]
        # v2.0.157: FLIPPED — now requires LOW volume (exhaustion), not HIGH (capitulation)
        # Vantixs: declining volume on move to band improved WR +5pp
        dataframe["volume_confirm"] = dataframe["volume_ratio"] < self.volume_multiplier

        # RSI
        dataframe["rsi"] = ta.RSI(dataframe["close"], length=self.rsi_length)

        # ADX — regime filter: skip mean reversion in strong trends
        dataframe["adx"] = ta.ADX(dataframe, length=14)

        # ── Entry Conditions ───────────────────────────────────────────────────
        threshold = self.entry_dev_threshold

        # Trend direction (4H EMA) — only used when trend filter is enabled
        if self.use_trend_filter:
            dataframe["trend_bullish"] = dataframe["close"] > dataframe["ema200_4h"]
            dataframe["trend_bearish"] = dataframe["close"] < dataframe["ema200_4h"]
        else:
            dataframe["trend_bullish"] = True
            dataframe["trend_bearish"] = True

        # Research v2.0.130: REMOVED reversal candle filter — v2.0.129 proved
        # too restrictive (1 trade). With 6 other entry filters already active,
        # adding bullish candle confirmation collapses trade count to near-zero.
        # The cross-back RSI approach (enters when RSI recovers from oversold)
        # already acts as a soft reversal confirmation.
        #
        # Long: deviation < -threshold, compression, volume, RSI recovering, trend
        dataframe["long_condition"] = (
            (dataframe["deviation"] < -threshold) &
            dataframe["in_compression"] &
            dataframe["volume_confirm"] &
            (dataframe["rsi"] > self.rsi_oversold) &        # RSI has EXITED oversold
            (dataframe["rsi"] < self.rsi_oversold_exit) &   # RSI still in bottom half (recovering)
            dataframe["trend_bullish"]
        )

        # ADX filter: skip if trending strongly (ADX > threshold = trending, not mean-reverting)
        if self.use_adx_filter:
            dataframe["long_condition"] = dataframe["long_condition"] & (dataframe["adx"] < self.adx_threshold)

        # Short: deviation > +threshold (price significantly above mean), compression, volume, RSI exiting overbought
        dataframe["short_condition"] = (
            (dataframe["deviation"] > threshold) &
            dataframe["in_compression"] &
            dataframe["volume_confirm"] &
            (dataframe["rsi"] < self.rsi_overbought) &  # RSI has EXITED overbought
            (dataframe["rsi"] > 50) &  # Still in top half
            dataframe["trend_bearish"]
        )

        if self.use_adx_filter:
            dataframe["short_condition"] = dataframe["short_condition"] & (dataframe["adx"] < self.adx_threshold)

        # R:R ratio (deviation / threshold)
        dataframe["rr_ratio"] = dataframe["deviation"].abs() / threshold

        return dataframe

    def populate_entry_trend(self, dataframe: DataFrame, metadata: dict) -> DataFrame:
        dataframe["enter_long"] = 0
        dataframe["enter_short"] = 0

        # R:R filter — min 1.0:1
        rr_ok = dataframe["rr_ratio"] >= 1.0

        long_mask = dataframe["long_condition"] & rr_ok
        dataframe.loc[long_mask, "enter_long"] = 1

        # Shorts disabled — spot mode only
        # short_mask = dataframe["short_condition"] & rr_ok
        # dataframe.loc[short_mask, "enter_short"] = 1

        return dataframe

    def populate_exit_trend(self, dataframe: DataFrame, metadata: dict) -> DataFrame:
        dataframe["exit_long"] = 0
        dataframe["exit_short"] = 0

        # Long exit: RSI reaches 65 (momentum normalized) OR deviation reverted to SMA
        # Research: target = mid-band (SMA), not extreme RSI levels
        dataframe.loc[
            (dataframe["rsi"] > self.exit_rsi_long) |
            (dataframe["deviation"] > self.exit_dev_revert_pct),
            "exit_long"
        ] = 1

        # Short exit: RSI drops to 35 OR deviation reverted below SMA
        dataframe.loc[
            (dataframe["rsi"] < self.exit_rsi_short) |
            (dataframe["deviation"] < -self.exit_dev_revert_pct),
            "exit_short"
        ] = 1

        return dataframe

    # Trailing stop — DISABLED. Research: trailing stops strangle mean reversion winners.
    # Winners averaged +0.86% because trail cut them at 1%. Let exits handle profit-taking.
    trailing_stop = False
    trailing_stop_positive = 0.0300
    trailing_stop_positive_offset = 0.1500
    # v2.0.66: research confirms trailing stops kill MR (ekx.ai, Connors).
    # Settings preserved as dead code per Freqtrade convention.
    trailing_only_offset_is_reached = True

    # Research v2.0.61: CRITICAL FIX — v2.0.60 custom_stoploss anchored to current_rate,
    # not entry price. When price dropped, stop drifted below hard stoploss (-7.7%),
    # so hard stoploss caught all losing trades (13 trades, -7.97% avg).
    # Fix: anchor stops to trade.open_rate. Research: crypto 1H needs 1.5-2×ATR.
    # Phase1 at 1.5×ATR (floor 3%, cap 5%) — gives breathing room for crypto vol.
    # Hard stoploss widened to -10% as disaster floor only (should rarely trigger).

    # Scale-in: disabled — adding size on small profit was amplifying losses.
    # v2.0.56: confirmed disabled. Research shows martingale is destructive for MR.
    scale_in_enabled = False

    def custom_stoploss(
        self, pair: str, trade: "Trade", current_time: datetime,
        current_rate: float, current_profit: float, after_fill: bool,
        **kwargs
    ) -> Optional[float]:
        """Time-graduated dynamic stop — anchored to ENTRY price.

        v2.0.81: COMPLETE REWRITE based on research synthesis.
        Connors/Cesar Alvarez: fixed stops kill MR. YouTube study:
        "timed exit if showing loss" improved net profit/drawdown 93% vs fixed stops.

        New approach:
        - First 12h: WIDE -7% from entry — let MR dips develop (research: MR reversions need room)
        - 12-24h: MODERATE -4% from entry — if no reversion by now, tightening
        - >24h: TIGHT -2.5% from entry — force exit, setup failed
        - Profitable trades: tight trailing lock-in at breakeven or small profit

        Anchored to entry price (not current) so dip tolerance is consistent.
        Hard stoploss at -8.5% is ultimate disaster floor.
        """
        entry_rate = trade.open_rate
        if not entry_rate or entry_rate <= 0:
            return -0.05  # fallback

        # Current drawdown from entry (as positive fraction)
        entry_drawdown = (entry_rate - current_rate) / entry_rate
        
        # Trailing lock-in for profitable trades (conservative only at high profit)
        # v2.0.82: REMOVED breakeven lock-in (>1% → 0.0) — killed all winners prematurely.
        # 24 trades hit trailing_stop at avg -2.54% instead of developing to +2-5% exit_signal.
        # Only lock in at 3%+ where the trade has clearly worked.
        if current_profit > 0.05:
            return -0.025    # 2.5% give-back allowed
        elif current_profit > 0.03:
            return -0.015    # v2.0.82: tight 1.5% give-back for 3%+ winners (was 2.5%)
        
        # Time-graduated stop for non-profitable trades
        if trade.open_date_utc:
            holding_hours = (current_time - trade.open_date_utc).total_seconds() / 3600
        else:
            holding_hours = 0
        
        hard_stop_pct = abs(self.stoploss)  # 0.085
            
        if holding_hours < 12:
            # Wide initial: 7% from entry — MR needs room to dip and bounce
            return -min(0.07, hard_stop_pct * 0.85)
        elif holding_hours < 24:
            # Moderate: 4% from entry — if no reversion in 12h, start tightening
            return -min(0.04, hard_stop_pct * 0.85)
        else:
            # Tight: 2.5% from entry — if no reversion by 24h, setup failed > exit
            return -min(0.025, hard_stop_pct * 0.85)

    def custom_exit(
        self, pair: str, trade: "Trade", current_time: datetime,
        current_rate: float, current_profit: float, **kwargs
    ) -> Optional[str]:
        """
        Time exits — both profitable AND losing trades.
        v2.0.81: ADDED losing-trade time exit. Research: timed exit on losers
        dramatically improves MR performance vs fixed stops (93% better net profit/DD).
        
        - Profitable trades: exit after time_exit_hours (18h) if profit >= floor (0.5%)
        - Losing trades: exit after time_exit_loss_hours (10h) regardless — prevent bag-holding
        """
        if trade.open_date_utc:
            holding_hours = (current_time - trade.open_date_utc).total_seconds() / 3600
        else:
            return None

        # Profitable time exit
        if current_profit >= self.time_exit_profit_floor:
            if holding_hours >= self.time_exit_hours:
                return "time_exit_hours"
        
        # Losing trade time exit — prevent dead setups from lingering
        if current_profit < 0 and holding_hours >= self.time_exit_loss_hours:
            return "time_exit_loss"
        
        return None

    def informative_pairs(self):
        """Preload 4H data for these pairs."""
        pairs = [
            "BTC/USDT", "ETH/USDT", "SOL/USDT",
            "BNB/USDT", "XRP/USDT"
        ]
        return [(p, "4h") for p in pairs]

    def confirm_trade_entry(
        self, pair: str, order_type: str, amount: float, rate: float,
        time_in_force: str, current_time: datetime, entry_tag: str,
        side: str, **kwargs
    ) -> bool:
        """Validate compression still active at entry moment."""
        # v2.0.69: get_pair_dataframe returns single DF in CI Freqtrade ver
        result = self.dp.get_pair_dataframe(pair, self.timeframe)
        df = result[0] if isinstance(result, tuple) else result
        if df.empty:
            return True
        last = df.iloc[-1]
        return last.get("in_compression", True)

    def adjust_trade_position(
        self, trade: "Trade", current_time: datetime, current_rate: float,
        current_profit: float, current_profit_pct: float, **kwargs
    ) -> Optional[float]:
        """Scale in 25% if profit is in the 0.5-1.5% compression zone — DISABLED after v1.0.5."""
        return None