TradeBot/crypto_spot_bot/time_series.py

from __future__ import annotations

import json
import math
from bisect import bisect_right
from dataclasses import asdict, dataclass, field
from typing import Any

from crypto_spot_bot.config import Settings
from crypto_spot_bot.models import Candle


DEFAULT_TORCH_FEATURES = (
    "return_1",
    "return_3",
    "return_6",
    "return_12",
    "return_24",
    "range_percent",
    "body_percent",
    "upper_wick_percent",
    "lower_wick_percent",
    "volume_change",
    "volume_ratio",
    "volume_percentile_20",
    "atr_percent",
    "atr_ratio_20",
    "realized_volatility_12",
    "realized_volatility_24",
    "rsi_centered",
    "rsi_slope_6",
    "macd_hist_percent",
    "macd_hist_slope_3",
    "ema50_gap_percent",
    "ema200_gap_percent",
    "ema20_slope_6",
    "ema50_slope_12",
    "ema200_slope_24",
    "ema50_ema200_gap_percent",
    "range_position_50",
    "trend_return_4h",
    "trend_return_24h",
    "daily_close_ema200_gap_percent",
    "daily_ema50_ema200_gap_percent",
    "daily_ema50_slope",
    "btc_return_1",
    "btc_return_3",
    "btc_return_6",
    "btc_return_24",
    "eth_return_1",
    "eth_return_3",
    "eth_return_6",
    "eth_return_24",
    "relative_btc_return_3",
    "relative_eth_return_3",
    "btc_eth_return_spread_3",
    "pattern_score",
    "pattern_bullish",
    "pattern_bearish",
    "pattern_range",
    "pattern_pullback",
    "pattern_oversold_reversal",
    "pattern_stabilized_drop",
    "pattern_breakout",
    "pattern_breakdown",
    "pattern_fast_drop",
    "pattern_volume_spike",
    "pattern_range_position_20",
)


FEATURE_DESCRIPTIONS: dict[str, tuple[str, str, str]] = {
    "return_1": ("Цена", "Доходность 1ч", "Изменение цены закрытия за последнюю 1h свечу."),
    "return_3": ("Цена", "Доходность 3ч", "Изменение цены закрытия за последние 3 часовые свечи."),
    "return_6": ("Цена", "Доходность 6ч", "Изменение цены закрытия за последние 6 часовых свечей."),
    "return_12": ("Цена", "Доходность 12ч", "Изменение цены закрытия за последние 12 часовых свечей."),
    "return_24": ("Цена", "Доходность 24ч", "Изменение цены закрытия за последние 24 часовые свечи."),
    "range_percent": ("Свеча", "Диапазон свечи", "Размер high-low последней свечи относительно цены закрытия."),
    "body_percent": ("Свеча", "Тело свечи", "Разница close-open относительно цены закрытия; знак показывает цвет свечи."),
    "upper_wick_percent": ("Свеча", "Верхняя тень", "Насколько далеко цена уходила выше тела свечи."),
    "lower_wick_percent": ("Свеча", "Нижняя тень", "Насколько далеко цена уходила ниже тела свечи."),
    "volume_change": ("Объем", "Изменение объема", "Изменение объема последней свечи относительно предыдущей."),
    "volume_ratio": ("Объем", "Объем к MA20", "Отклонение текущего объема от средней за 20 свечей."),
    "volume_percentile_20": ("Объем", "Процентиль объема", "Позиция текущего объема среди последних 20 свечей."),
    "atr_percent": ("Волатильность", "ATR14 %", "Средний торговый диапазон ATR14 относительно цены."),
    "atr_ratio_20": ("Волатильность", "ATR к среднему", "Отклонение текущего ATR от среднего ATR за 20 свечей."),
    "realized_volatility_12": ("Волатильность", "Реализованная вола 12ч", "Фактическая волатильность доходностей за 12 свечей."),
    "realized_volatility_24": ("Волатильность", "Реализованная вола 24ч", "Фактическая волатильность доходностей за 24 свечи."),
    "rsi_centered": ("Индикаторы", "RSI14 от 50", "RSI14, приведенный к центру 50: выше нуля сильнее покупатели."),
    "rsi_slope_6": ("Индикаторы", "Наклон RSI 6ч", "Изменение RSI14 за последние 6 свечей."),
    "macd_hist_percent": ("Индикаторы", "MACD histogram", "MACD histogram относительно цены; знак показывает импульс."),
    "macd_hist_slope_3": ("Индикаторы", "Наклон MACD hist", "Изменение MACD histogram за последние 3 свечи."),
    "ema50_gap_percent": ("EMA/тренд", "Цена к EMA50", "Расстояние цены закрытия до EMA50."),
    "ema200_gap_percent": ("EMA/тренд", "Цена к EMA200", "Расстояние цены закрытия до EMA200."),
    "ema20_slope_6": ("EMA/тренд", "Наклон EMA20", "Изменение EMA20 за последние 6 свечей."),
    "ema50_slope_12": ("EMA/тренд", "Наклон EMA50", "Изменение EMA50 за последние 12 свечей."),
    "ema200_slope_24": ("EMA/тренд", "Наклон EMA200", "Изменение EMA200 за последние 24 свечи."),
    "ema50_ema200_gap_percent": ("EMA/тренд", "EMA50 к EMA200", "Расстояние EMA50 относительно EMA200."),
    "range_position_50": ("Цена", "Позиция в диапазоне 50ч", "Где текущая цена внутри high-low диапазона последних 50 свечей."),
    "trend_return_4h": ("Цена", "Тренд 4ч", "Изменение цены за последние 4 свечи."),
    "trend_return_24h": ("Цена", "Тренд 24ч", "Изменение цены за последние 24 свечи."),
    "daily_close_ema200_gap_percent": ("Дневной тренд", "D цена к EMA200", "Расстояние дневного close до дневной EMA200."),
    "daily_ema50_ema200_gap_percent": ("Дневной тренд", "D EMA50 к EMA200", "Расстояние дневной EMA50 относительно дневной EMA200."),
    "daily_ema50_slope": ("Дневной тренд", "D наклон EMA50", "Изменение дневной EMA50 за последние несколько дневных свечей."),
    "btc_return_1": ("BTC/ETH контекст", "BTC 1ч", "Изменение BTCUSDT за последнюю 1h свечу."),
    "btc_return_3": ("BTC/ETH контекст", "BTC 3ч", "Изменение BTCUSDT за последние 3 часа."),
    "btc_return_6": ("BTC/ETH контекст", "BTC 6ч", "Изменение BTCUSDT за последние 6 часов."),
    "btc_return_24": ("BTC/ETH контекст", "BTC 24ч", "Изменение BTCUSDT за последние 24 часа."),
    "eth_return_1": ("BTC/ETH контекст", "ETH 1ч", "Изменение ETHUSDT за последнюю 1h свечу."),
    "eth_return_3": ("BTC/ETH контекст", "ETH 3ч", "Изменение ETHUSDT за последние 3 часа."),
    "eth_return_6": ("BTC/ETH контекст", "ETH 6ч", "Изменение ETHUSDT за последние 6 часов."),
    "eth_return_24": ("BTC/ETH контекст", "ETH 24ч", "Изменение ETHUSDT за последние 24 часа."),
    "relative_btc_return_3": ("BTC/ETH контекст", "Сила к BTC 3ч", "Доходность пары за 3 часа минус доходность BTC за 3 часа."),
    "relative_eth_return_3": ("BTC/ETH контекст", "Сила к ETH 3ч", "Доходность пары за 3 часа минус доходность ETH за 3 часа."),
    "btc_eth_return_spread_3": ("BTC/ETH контекст", "BTC-ETH 3ч", "Разница 3-часовой доходности BTC и ETH."),
    "pattern_score": ("Шаблон", "Оценка шаблона", "Числовая оценка текущего рыночного шаблона от 0 до 1."),
    "pattern_bullish": ("Шаблон", "Бычий шаблон", "Флаг, что текущий шаблон похож на бычий."),
    "pattern_bearish": ("Шаблон", "Медвежий шаблон", "Флаг, что текущий шаблон похож на медвежий."),
    "pattern_range": ("Шаблон", "Флэтовый шаблон", "Флаг, что рынок похож на диапазон/флэт."),
    "pattern_pullback": ("Шаблон", "Откат", "Флаг отката внутри восходящего движения."),
    "pattern_oversold_reversal": ("Шаблон", "Перепроданный разворот", "Флаг возможного разворота после перепроданности."),
    "pattern_stabilized_drop": ("Шаблон", "Падение стабилизировалось", "Флаг замедления падения и попытки стабилизации."),
    "pattern_breakout": ("Шаблон", "Пробой вверх", "Флаг пробоя верхней части диапазона с объемом."),
    "pattern_breakdown": ("Шаблон", "Пробой вниз", "Флаг пробоя нижней части диапазона с объемом."),
    "pattern_fast_drop": ("Шаблон", "Быстрое падение", "Флаг резкого снижения или сильной перепроданности."),
    "pattern_volume_spike": ("Шаблон", "Всплеск объема", "Флаг объема заметно выше обычного."),
    "pattern_range_position_20": ("Шаблон", "Позиция в диапазоне 20ч", "Где цена внутри high-low диапазона последних 20 свечей."),
}


@dataclass(slots=True)
class TimeSeriesForecast:
    enabled: bool
    usable: bool
    model: str
    volatility_model: str
    expected_return_percent: float
    expected_price: float
    volatility_percent: float
    probability_up: float
    confidence_adjustment: float
    block_entry: bool
    validation_mae_percent: float
    baseline_mae_percent: float
    skill: float
    horizon: int
    reason: str
    expected_gross_return_percent: float
    quantile_10_percent: float
    quantile_50_percent: float
    quantile_90_percent: float
    conservative_return_percent: float
    target_transform: str
    feature_snapshot: list[dict[str, Any]] = field(default_factory=list)
    horizon_forecasts: dict[str, Any] = field(default_factory=dict)
    candidates: list[dict[str, Any]] = field(default_factory=list)

    def as_dict(self) -> dict[str, Any]:
        return asdict(self)


class TimeSeriesForecaster:
    def __init__(self, settings: Settings):
        self.settings = settings
        self._lstm_artifact_mtime: float | None = None
        self._lstm_artifact: dict[str, Any] = {}

    def forecast(
        self,
        candles: list[Candle],
        symbol: str | None = None,
        *,
        market_candles: dict[str, list[Candle]] | None = None,
        trend_candles: list[Candle] | None = None,
    ) -> TimeSeriesForecast:
        if not self.settings.time_series_forecast_enabled:
            return _empty_forecast(False, "time-series forecast is disabled")
        closes = [float(candle.close) for candle in candles if candle.close > 0]
        min_candles = max(30, self.settings.time_series_min_candles)
        if len(closes) < min_candles:
            return _empty_forecast(True, "not enough candles for PyTorch forecast")
        returns = _log_returns(closes)
        if len(returns) < 20:
            return _empty_forecast(True, "not enough returns for PyTorch forecast")

        artifact = self._load_lstm_artifact()
        entry = _torch_recurrent_entry(symbol, artifact)
        model = _torch_recurrent_model_name(symbol, artifact)
        clip = _clamp(_float_entry(entry or {}, "clip", 8.0), 1.0, 50.0)
        feature_rows = (
            _feature_matrix(
                candles,
                _feature_names(entry),
                symbol=symbol,
                market_candles=market_candles,
                trend_candles=trend_candles,
            )
            if entry
            else []
        )
        feature_snapshot = _feature_snapshot(feature_rows, entry, clip)
        if not model or not _can_use_torch_recurrent(returns, symbol, artifact, feature_rows):
            return _empty_forecast(True, "no valid PyTorch LSTM/GRU model for symbol")

        prediction = _torch_recurrent_predict(
            returns,
            symbol,
            artifact,
            feature_rows=feature_rows,
            closes=closes,
            candles=candles,
        )
        if entry is None or prediction is None:
            return _empty_forecast(True, "PyTorch LSTM/GRU model could not build a forecast")

        if isinstance(prediction, dict):
            selected = _select_horizon_prediction(
                prediction,
                _entry_horizon(entry, self.settings.time_series_forecast_horizon),
            )
            if not selected:
                return _empty_forecast(True, "PyTorch LSTM/GRU model could not select a forecast horizon")
            expected_return = float(selected["expected_return"])
            expected_gross_return = float(selected.get("expected_gross_return", expected_return))
            expected_price = closes[-1] * math.exp(expected_gross_return)
            probability_up = _clamp(float(selected.get("probability_up", 0.5)), 0.0, 1.0)
            model_mae = max(float(selected.get("validation_mae", 0.0)), 1e-9)
            baseline_mae = max(float(selected.get("baseline_mae", model_mae)), model_mae)
            uncertainty = max(float(selected.get("uncertainty", model_mae)), 1e-9)
            volatility_percent = uncertainty * 100
            expected_return_percent = (math.exp(expected_return) - 1) * 100
            expected_gross_return_percent = (math.exp(expected_gross_return) - 1) * 100
            q10_percent = (math.exp(float(selected.get("q10", expected_return))) - 1) * 100
            q50_percent = (math.exp(float(selected.get("q50", expected_return))) - 1) * 100
            q90_percent = (math.exp(float(selected.get("q90", expected_return))) - 1) * 100
            skill = _clamp(_float_entry(entry, "skill", 0.0), -1.0, 1.0)
            horizon = int(selected.get("horizon", _entry_horizon(entry, self.settings.time_series_forecast_horizon)))
            min_edge = max(0.0, self.settings.time_series_min_edge_percent)
            confidence_adjustment = _confidence_adjustment(
                expected_return_percent=expected_return_percent,
                probability_up=probability_up,
                skill=skill,
                min_edge=min_edge,
                max_adjustment=self.settings.time_series_max_adjustment,
            )
            conservative_return_percent = min(expected_return_percent, q50_percent)
            block_entry = bool(
                (expected_return_percent <= -min_edge and probability_up <= 0.45)
                or (q50_percent <= -min_edge and probability_up <= 0.48)
            )
            reason = _reason(
                model=model,
                expected_return_percent=expected_return_percent,
                probability_up=probability_up,
                skill=skill,
                block_entry=block_entry,
            )
            return TimeSeriesForecast(
                enabled=True,
                usable=True,
                model=model,
                volatility_model="probabilistic multi-horizon after-cost quantile",
                expected_return_percent=round(expected_return_percent, 4),
                expected_price=round(expected_price, 8),
                volatility_percent=round(volatility_percent, 4),
                probability_up=round(probability_up, 4),
                confidence_adjustment=round(confidence_adjustment, 4),
                block_entry=block_entry,
                validation_mae_percent=round(model_mae * 100, 4),
                baseline_mae_percent=round(baseline_mae * 100, 4),
                skill=round(skill, 4),
                horizon=horizon,
                reason=reason,
                expected_gross_return_percent=round(expected_gross_return_percent, 4),
                quantile_10_percent=round(q10_percent, 4),
                quantile_50_percent=round(q50_percent, 4),
                quantile_90_percent=round(q90_percent, 4),
                conservative_return_percent=round(conservative_return_percent, 4),
                target_transform=str(entry.get("target_transform", "net_return_over_volatility")),
                feature_snapshot=feature_snapshot,
                horizon_forecasts=_public_horizon_forecasts(prediction),
                candidates=[{"model": model, "mae_percent": round(model_mae * 100, 4)}],
            )

        direct_horizon = _is_direct_horizon(entry)
        horizon = _entry_horizon(entry, self.settings.time_series_forecast_horizon)
        expected_return = prediction if direct_horizon else prediction * horizon
        expected_price = closes[-1] * math.exp(expected_return)
        model_mae = _torch_validation_mae(entry, returns)
        baseline_mae = max(_float_entry(entry, "baseline_mae_percent", model_mae * 100) / 100, model_mae)
        if direct_horizon:
            uncertainty = max(model_mae, _horizon_return_scale(closes, horizon) * 0.25, 1e-9)
            volatility_model = "direct horizon validation MAE"
        else:
            uncertainty_one_step = max(model_mae, _return_scale(returns) * 0.25, 1e-9)
            uncertainty = uncertainty_one_step * math.sqrt(horizon)
            volatility_model = "one-step validation MAE scaled by horizon"
        volatility_percent = uncertainty * 100
        expected_return_percent = (math.exp(expected_return) - 1) * 100
        probability_up = _normal_cdf(expected_return / max(uncertainty, 1e-9))
        skill = _clamp(_float_entry(entry, "skill", 0.0), -1.0, 1.0)
        min_edge = max(0.0, self.settings.time_series_min_edge_percent)
        confidence_adjustment = _confidence_adjustment(
            expected_return_percent=expected_return_percent,
            probability_up=probability_up,
            skill=skill,
            min_edge=min_edge,
            max_adjustment=self.settings.time_series_max_adjustment,
        )
        block_entry = bool(expected_return_percent <= -min_edge and probability_up <= 0.45)
        reason = _reason(
            model=model,
            expected_return_percent=expected_return_percent,
            probability_up=probability_up,
            skill=skill,
            block_entry=block_entry,
        )
        return TimeSeriesForecast(
            enabled=True,
            usable=True,
            model=model,
            volatility_model=volatility_model,
            expected_return_percent=round(expected_return_percent, 4),
            expected_price=round(expected_price, 8),
            volatility_percent=round(volatility_percent, 4),
            probability_up=round(probability_up, 4),
            confidence_adjustment=round(confidence_adjustment, 4),
            block_entry=block_entry,
            validation_mae_percent=round(model_mae * 100, 4),
            baseline_mae_percent=round(baseline_mae * 100, 4),
            skill=round(skill, 4),
            horizon=horizon,
            reason=reason,
            expected_gross_return_percent=round(expected_return_percent, 4),
            quantile_10_percent=round(expected_return_percent - volatility_percent, 4),
            quantile_50_percent=round(expected_return_percent, 4),
            quantile_90_percent=round(expected_return_percent + volatility_percent, 4),
            conservative_return_percent=round(expected_return_percent, 4),
            target_transform=str(entry.get("target_transform", "direct_log_return")),
            feature_snapshot=feature_snapshot,
            horizon_forecasts={},
            candidates=[{"model": model, "mae_percent": round(model_mae * 100, 4)}],
        )

    def _load_lstm_artifact(self) -> dict[str, Any]:
        if not self.settings.time_series_lstm_enabled:
            return {}
        path = self.settings.time_series_lstm_model_path
        try:
            stat = path.stat()
        except OSError:
            self._lstm_artifact_mtime = None
            self._lstm_artifact = {}
            return {}
        if self._lstm_artifact_mtime == stat.st_mtime:
            return self._lstm_artifact
        try:
            data = json.loads(path.read_text(encoding="utf-8"))
        except (OSError, json.JSONDecodeError):
            data = {}
        self._lstm_artifact = data if isinstance(data, dict) else {}
        self._lstm_artifact_mtime = stat.st_mtime
        return self._lstm_artifact


def _empty_forecast(enabled: bool, reason: str) -> TimeSeriesForecast:
    return TimeSeriesForecast(
        enabled=enabled,
        usable=False,
        model="none",
        volatility_model="none",
        expected_return_percent=0.0,
        expected_price=0.0,
        volatility_percent=0.0,
        probability_up=0.5,
        confidence_adjustment=0.0,
        block_entry=False,
        validation_mae_percent=0.0,
        baseline_mae_percent=0.0,
        skill=0.0,
        horizon=0,
        reason=reason,
        expected_gross_return_percent=0.0,
        quantile_10_percent=0.0,
        quantile_50_percent=0.0,
        quantile_90_percent=0.0,
        conservative_return_percent=0.0,
        target_transform="none",
        feature_snapshot=[],
        horizon_forecasts={},
    )


def _log_returns(closes: list[float]) -> list[float]:
    return [math.log(closes[index] / closes[index - 1]) for index in range(1, len(closes))]


def _feature_matrix(
    candles: list[Candle],
    feature_names: list[str] | tuple[str, ...] | None = None,
    *,
    symbol: str | None = None,
    market_candles: dict[str, list[Candle]] | None = None,
    trend_candles: list[Candle] | None = None,
) -> list[list[float]]:
    names = list(feature_names or DEFAULT_TORCH_FEATURES)
    context = _feature_context(
        candles,
        symbol=symbol,
        market_candles=market_candles,
        trend_candles=trend_candles,
    )
    rows: list[list[float]] = []
    for index, candle in enumerate(candles):
        rows.append([_feature_value(name, candles, index, candle, context) for name in names])
    return rows


def _feature_context(
    candles: list[Candle],
    *,
    symbol: str | None,
    market_candles: dict[str, list[Candle]] | None,
    trend_candles: list[Candle] | None,
) -> dict[str, Any]:
    market_candles = market_candles or {}
    normalized_market = {key.upper(): value for key, value in market_candles.items()}
    context_indexes = {
        key: {candle.timestamp: index for index, candle in enumerate(rows)}
        for key, rows in normalized_market.items()
    }
    trend_rows = trend_candles or []
    trend_timestamps = [candle.timestamp for candle in trend_rows]
    trend_positions = [
        bisect_right(trend_timestamps, candle.timestamp) - 1
        if trend_timestamps
        else -1
        for candle in candles
    ]
    return {
        "symbol": (symbol or "").upper(),
        "market_candles": normalized_market,
        "context_indexes": context_indexes,
        "trend_candles": trend_rows,
        "trend_positions": trend_positions,
    }


def _feature_value(name: str, candles: list[Candle], index: int, candle: Candle, context: dict[str, Any]) -> float:
    close = max(float(candle.close), 1e-12)
    previous = candles[index - 1] if index >= 1 else candle
    if name == "return_1":
        return _log_change(candle.close, previous.close)
    if name == "return_3":
        return _log_change(candle.close, candles[index - 3].close) if index >= 3 else 0.0
    if name == "return_6":
        return _log_change(candle.close, candles[index - 6].close) if index >= 6 else 0.0
    if name == "return_12":
        return _log_change(candle.close, candles[index - 12].close) if index >= 12 else 0.0
    if name == "return_24":
        return _log_change(candle.close, candles[index - 24].close) if index >= 24 else 0.0
    if name == "range_percent":
        return _safe_feature((candle.high - candle.low) / close)
    if name == "body_percent":
        return _safe_feature((candle.close - candle.open) / close)
    if name == "upper_wick_percent":
        return _safe_feature((candle.high - max(candle.open, candle.close)) / close)
    if name == "lower_wick_percent":
        return _safe_feature((min(candle.open, candle.close) - candle.low) / close)
    if name == "volume_change":
        return _log_change(max(candle.volume, 1e-12), max(previous.volume, 1e-12))
    if name == "volume_ratio":
        return _safe_feature((candle.volume / candle.volume_ma_20) - 1.0) if candle.volume_ma_20 else 0.0
    if name == "volume_percentile_20":
        return _rolling_percentile([row.volume for row in candles], index, 20)
    if name == "atr_percent":
        return _safe_feature(candle.atr_14 / close) if candle.atr_14 is not None else 0.0
    if name == "atr_ratio_20":
        return _ratio_to_recent_mean(
            [row.atr_14 if row.atr_14 is not None else 0.0 for row in candles],
            index,
            20,
        )
    if name == "realized_volatility_12":
        return _realized_volatility(candles, index, 12)
    if name == "realized_volatility_24":
        return _realized_volatility(candles, index, 24)
    if name == "rsi_centered":
        return _safe_feature((candle.rsi_14 - 50.0) / 50.0) if candle.rsi_14 is not None else 0.0
    if name == "rsi_slope_6":
        return _indicator_slope(candles, index, "rsi_14", 6, divisor=50.0)
    if name == "macd_hist_percent":
        return _safe_feature(candle.macd_hist / close) if candle.macd_hist is not None else 0.0
    if name == "macd_hist_slope_3":
        return _indicator_price_slope(candles, index, "macd_hist", 3)
    if name == "ema50_gap_percent":
        return _safe_feature((candle.close - candle.ema_50) / close) if candle.ema_50 is not None else 0.0
    if name == "ema200_gap_percent":
        return _safe_feature((candle.close - candle.ema_200) / close) if candle.ema_200 is not None else 0.0
    if name == "ema20_slope_6":
        return _ema_slope(candles, index, "ema_20", 6)
    if name == "ema50_slope_12":
        return _ema_slope(candles, index, "ema_50", 12)
    if name == "ema200_slope_24":
        return _ema_slope(candles, index, "ema_200", 24)
    if name == "ema50_ema200_gap_percent":
        if candle.ema_50 is not None and candle.ema_200 is not None and candle.ema_200 > 0:
            return _safe_feature((candle.ema_50 - candle.ema_200) / candle.ema_200)
        return 0.0
    if name == "range_position_50":
        return _range_position(candles, index, 50)
    if name == "trend_return_4h":
        return _log_change(candle.close, candles[index - 4].close) if index >= 4 else 0.0
    if name == "trend_return_24h":
        return _log_change(candle.close, candles[index - 24].close) if index >= 24 else 0.0
    if name.startswith("daily_"):
        return _daily_feature_value(name, context, index)
    if name.startswith("btc_") or name.startswith("eth_") or name.startswith("relative_"):
        return _cross_asset_feature_value(name, candles, index, candle, context)
    if name.startswith("pattern_"):
        return _pattern_feature_value(name, candles, index)
    return 0.0


def _pattern_feature_value(name: str, candles: list[Candle], index: int) -> float:
    pattern = _pattern_snapshot(candles, index)
    if name == "pattern_score":
        return pattern["score"]
    if name == "pattern_bullish":
        return pattern["bullish"]
    if name == "pattern_bearish":
        return pattern["bearish"]
    if name == "pattern_range":
        return pattern["range"]
    if name == "pattern_pullback":
        return pattern["pullback"]
    if name == "pattern_oversold_reversal":
        return pattern["oversold_reversal"]
    if name == "pattern_stabilized_drop":
        return pattern["stabilized_drop"]
    if name == "pattern_breakout":
        return pattern["breakout"]
    if name == "pattern_breakdown":
        return pattern["breakdown"]
    if name == "pattern_fast_drop":
        return pattern["fast_drop"]
    if name == "pattern_volume_spike":
        return pattern["volume_spike"]
    if name == "pattern_range_position_20":
        return pattern["range_position_20"]
    return 0.0


def _rolling_percentile(values: list[float], index: int, window: int) -> float:
    start = max(0, index - window + 1)
    sample = [float(value) for value in values[start : index + 1] if math.isfinite(float(value))]
    if not sample:
        return 0.5
    current = float(values[index])
    below_or_equal = sum(1 for value in sample if value <= current)
    return _clamp(below_or_equal / len(sample), 0.0, 1.0)


def _ratio_to_recent_mean(values: list[float], index: int, window: int) -> float:
    current = float(values[index]) if index < len(values) else 0.0
    start = max(0, index - window + 1)
    sample = [float(value) for value in values[start : index + 1] if math.isfinite(float(value)) and value > 0]
    if current <= 0 or not sample:
        return 0.0
    mean = sum(sample) / len(sample)
    return _safe_feature((current / mean) - 1.0) if mean > 0 else 0.0


def _realized_volatility(candles: list[Candle], index: int, window: int) -> float:
    if index < 1:
        return 0.0
    start = max(1, index - window + 1)
    returns = [
        _log_change(candles[position].close, candles[position - 1].close)
        for position in range(start, index + 1)
    ]
    if not returns:
        return 0.0
    return _safe_feature(math.sqrt(sum(value * value for value in returns) / len(returns)))


def _indicator_slope(candles: list[Candle], index: int, attr: str, steps: int, *, divisor: float) -> float:
    if index < steps:
        return 0.0
    current = getattr(candles[index], attr)
    previous = getattr(candles[index - steps], attr)
    if current is None or previous is None or divisor <= 0:
        return 0.0
    return _safe_feature((float(current) - float(previous)) / divisor)


def _indicator_price_slope(candles: list[Candle], index: int, attr: str, steps: int) -> float:
    if index < steps:
        return 0.0
    current = getattr(candles[index], attr)
    previous = getattr(candles[index - steps], attr)
    close = max(float(candles[index].close), 1e-12)
    if current is None or previous is None:
        return 0.0
    return _safe_feature((float(current) - float(previous)) / close)


def _ema_slope(candles: list[Candle], index: int, attr: str, steps: int) -> float:
    if index < steps:
        return 0.0
    current = getattr(candles[index], attr)
    previous = getattr(candles[index - steps], attr)
    if current is None or previous is None or previous <= 0:
        return 0.0
    return _safe_feature(math.log(float(current) / float(previous)))


def _range_position(candles: list[Candle], index: int, window: int) -> float:
    start = max(0, index - window + 1)
    rows = candles[start : index + 1]
    if not rows:
        return 0.5
    high = max(row.high for row in rows)
    low = min(row.low for row in rows)
    width = high - low
    if width <= 0:
        return 0.5
    return _clamp((candles[index].close - low) / width, 0.0, 1.0)


def _daily_feature_value(name: str, context: dict[str, Any], index: int) -> float:
    trend_rows: list[Candle] = context.get("trend_candles") or []
    trend_positions: list[int] = context.get("trend_positions") or []
    if index >= len(trend_positions):
        return 0.0
    trend_index = trend_positions[index]
    if trend_index < 0 or trend_index >= len(trend_rows):
        return 0.0
    trend = trend_rows[trend_index]
    if name == "daily_close_ema200_gap_percent":
        if trend.ema_200 is not None and trend.ema_200 > 0:
            return _safe_feature((trend.close - trend.ema_200) / trend.ema_200)
        return 0.0
    if name == "daily_ema50_ema200_gap_percent":
        if trend.ema_50 is not None and trend.ema_200 is not None and trend.ema_200 > 0:
            return _safe_feature((trend.ema_50 - trend.ema_200) / trend.ema_200)
        return 0.0
    if name == "daily_ema50_slope":
        previous_index = trend_index - 5
        if previous_index >= 0 and trend.ema_50 is not None and trend_rows[previous_index].ema_50:
            return _safe_feature(math.log(trend.ema_50 / trend_rows[previous_index].ema_50))
    return 0.0


def _cross_asset_feature_value(
    name: str,
    candles: list[Candle],
    index: int,
    candle: Candle,
    context: dict[str, Any],
) -> float:
    if name == "btc_eth_return_spread_3":
        return _safe_feature(
            _context_return("BTCUSDT", 3, candle.timestamp, context)
            - _context_return("ETHUSDT", 3, candle.timestamp, context)
        )
    if name.startswith("btc_return_"):
        steps = int(name.rsplit("_", 1)[1])
        return _context_return("BTCUSDT", steps, candle.timestamp, context)
    if name.startswith("eth_return_"):
        steps = int(name.rsplit("_", 1)[1])
        return _context_return("ETHUSDT", steps, candle.timestamp, context)
    if name == "relative_btc_return_3":
        return _safe_feature((_log_change(candle.close, candles[index - 3].close) if index >= 3 else 0.0) - _context_return("BTCUSDT", 3, candle.timestamp, context))
    if name == "relative_eth_return_3":
        return _safe_feature((_log_change(candle.close, candles[index - 3].close) if index >= 3 else 0.0) - _context_return("ETHUSDT", 3, candle.timestamp, context))
    return 0.0


def _context_return(symbol: str, steps: int, timestamp: int, context: dict[str, Any]) -> float:
    rows = (context.get("market_candles") or {}).get(symbol)
    indexes = (context.get("context_indexes") or {}).get(symbol)
    if not rows or not indexes:
        return 0.0
    index = indexes.get(timestamp)
    if index is None or index < steps:
        return 0.0
    return _log_change(rows[index].close, rows[index - steps].close)


def _pattern_snapshot(candles: list[Candle], index: int) -> dict[str, float]:
    if index < 29:
        return {
            "score": 0.0,
            "bullish": 0.0,
            "bearish": 0.0,
            "range": 0.0,
            "pullback": 0.0,
            "oversold_reversal": 0.0,
            "stabilized_drop": 0.0,
            "breakout": 0.0,
            "breakdown": 0.0,
            "fast_drop": 0.0,
            "volume_spike": 0.0,
            "range_position_20": 0.5,
        }

    window = candles[: index + 1]
    latest = window[-1]
    previous = window[-2]
    high20 = max(candle.high for candle in window[-20:])
    low20 = min(candle.low for candle in window[-20:])
    width20 = max(0.0, high20 - low20)
    range_position_20 = _clamp((latest.close - low20) / width20, 0.0, 1.0) if width20 else 0.5
    close_3 = window[-4].close if len(window) >= 4 else window[0].close
    close_10 = window[-11].close if len(window) >= 11 else window[0].close
    close_20 = window[-21].close if len(window) >= 21 else window[0].close
    ret_3 = _percent_change(latest.close, close_3)
    ret_10 = _percent_change(latest.close, close_10)
    ret_20 = _percent_change(latest.close, close_20)
    body = abs(latest.close - latest.open)
    lower_wick = max(0.0, min(latest.open, latest.close) - latest.low)
    atr_percent = (latest.atr_14 / latest.close * 100) if latest.atr_14 and latest.close else 0.0
    volume_ratio = (
        latest.volume / latest.volume_ma_20
        if latest.volume_ma_20 and latest.volume_ma_20 > 0
        else 0.0
    )
    ema_gap_percent = (
        (latest.ema_50 - latest.ema_200) / latest.ema_200 * 100
        if latest.ema_50 and latest.ema_200
        else 0.0
    )
    uptrend = bool(
        latest.ema_20
        and latest.ema_50
        and latest.ema_200
        and latest.ema_20 >= latest.ema_50 >= latest.ema_200
        and latest.close >= latest.ema_50
    )
    downtrend = bool(
        latest.ema_20
        and latest.ema_50
        and latest.ema_200
        and latest.ema_20 <= latest.ema_50 <= latest.ema_200
        and latest.close <= latest.ema_50
    )
    pullback = bool(
        latest.ema_20
        and uptrend
        and latest.close <= latest.ema_20 * 1.012
        and latest.rsi_14 is not None
        and 35 <= latest.rsi_14 <= 58
    )
    oversold_reversal = bool(
        latest.rsi_14 is not None
        and latest.rsi_14 <= 35
        and latest.close > previous.close
        and lower_wick >= body * 1.2
    )
    stabilized_drop = _pattern_stabilized_drop(
        candles=window,
        latest=latest,
        previous=previous,
        ret_3=ret_3,
        ret_10=ret_10,
        ret_20=ret_20,
        atr_percent=atr_percent,
        volume_ratio=volume_ratio,
        lower_wick=lower_wick,
        body=body,
    )
    breakout = bool(latest.close >= high20 * 0.995 and volume_ratio >= 1.15 and latest.close > latest.open)
    breakdown = bool(latest.close <= low20 * 1.005 and volume_ratio >= 1.1 and latest.close < latest.open)
    fast_drop = bool(ret_3 <= -max(1.2, atr_percent * 1.8) or (latest.rsi_14 or 100) <= 25)
    range_market = bool(abs(ret_20) <= max(0.8, atr_percent * 1.2) and abs(ema_gap_percent) <= 0.35)
    volume_spike = bool(volume_ratio >= 1.6)

    score = 0.50
    if fast_drop and breakdown:
        score = 0.18
    elif breakdown:
        score = 0.24
    elif pullback:
        score = 0.76
    elif oversold_reversal:
        score = 0.68
    elif stabilized_drop:
        score = 0.58
    elif breakout:
        score = 0.72
    elif uptrend:
        score = 0.64
    elif range_market:
        score = 0.48
    elif downtrend:
        score = 0.28
    bullish = float(pullback or oversold_reversal or stabilized_drop or breakout or uptrend)
    bearish = float((fast_drop and breakdown) or breakdown or downtrend)
    return {
        "score": score,
        "bullish": bullish,
        "bearish": bearish,
        "range": float(range_market),
        "pullback": float(pullback),
        "oversold_reversal": float(oversold_reversal),
        "stabilized_drop": float(stabilized_drop),
        "breakout": float(breakout),
        "breakdown": float(breakdown),
        "fast_drop": float(fast_drop),
        "volume_spike": float(volume_spike),
        "range_position_20": range_position_20,
    }


def _percent_change(current: float, previous: float) -> float:
    return ((current - previous) / previous * 100) if previous else 0.0


def _pattern_stabilized_drop(
    *,
    candles: list[Candle],
    latest: Candle,
    previous: Candle,
    ret_3: float,
    ret_10: float,
    ret_20: float,
    atr_percent: float,
    volume_ratio: float,
    lower_wick: float,
    body: float,
) -> bool:
    recent_drop = ret_10 <= -max(0.35, atr_percent * 1.1) or ret_20 <= -max(0.6, atr_percent * 1.6)
    if not recent_drop or latest.rsi_14 is None or latest.rsi_14 > 52:
        return False
    recent_lows = [candle.low for candle in candles[-5:-1]]
    no_new_low = bool(recent_lows) and latest.low >= min(recent_lows) * 0.999
    bounce_from_low = ((latest.close - min(candle.low for candle in candles[-6:])) / latest.close * 100) if latest.close else 0.0
    body_base = max(body, latest.close * 0.0001)
    absorption = lower_wick >= body_base * 0.6 or bounce_from_low >= max(0.08, atr_percent * 0.3)
    momentum_stabilized = latest.close >= previous.close or abs(ret_3) <= max(0.25, atr_percent * 0.8) or no_new_low
    volume_present = volume_ratio >= 0.55
    continuing_drop = latest.close < previous.close and not no_new_low and ret_3 <= -max(0.6, atr_percent * 1.2)
    return bool(momentum_stabilized and absorption and volume_present and not continuing_drop)


def _log_change(current: float, previous: float) -> float:
    if current <= 0 or previous <= 0:
        return 0.0
    return _safe_feature(math.log(current / previous))


def _safe_feature(value: float) -> float:
    if not math.isfinite(value):
        return 0.0
    return _clamp(float(value), -50.0, 50.0)


def _torch_recurrent_model_name(symbol: str | None, artifact: dict[str, Any]) -> str | None:
    entry = _torch_recurrent_entry(symbol, artifact)
    if not entry:
        return None
    architecture = str(entry.get("architecture", "")).strip().lower()
    if architecture in {"lstm", "gru"}:
        return f"torch_{architecture}"
    model = str(entry.get("model", "")).strip().lower()
    return model if model in {"torch_lstm", "torch_gru"} else None


def _torch_recurrent_entry(symbol: str | None, artifact: dict[str, Any]) -> dict[str, Any] | None:
    if artifact.get("type") != "pytorch_recurrent_forecaster":
        return None
    symbols = artifact.get("symbols")
    entry = symbols.get(symbol.upper()) if symbol and isinstance(symbols, dict) else None
    if not isinstance(entry, dict):
        default = artifact.get("default")
        entry = default if isinstance(default, dict) else None
    if not isinstance(entry, dict):
        return None
    if not isinstance(entry.get("state_dict"), dict):
        return None
    return entry


def _can_use_torch_recurrent(
    returns: list[float],
    symbol: str | None,
    artifact: dict[str, Any],
    feature_rows: list[list[float]] | None = None,
) -> bool:
    entry = _torch_recurrent_entry(symbol, artifact)
    if not entry:
        return False
    lookback = int(_clamp(_float_entry(entry, "lookback", 0.0), 4.0, 512.0))
    hidden_size = int(_clamp(_float_entry(entry, "hidden_size", 0.0), 1.0, 512.0))
    num_layers = int(_clamp(_float_entry(entry, "num_layers", 1.0), 1.0, 8.0))
    if hidden_size <= 0 or num_layers <= 0:
        return False
    if _is_direct_horizon(entry):
        return bool(feature_rows and len(feature_rows) >= lookback)
    return len(returns) >= lookback + 1


def _torch_recurrent_predict(
    returns: list[float],
    symbol: str | None,
    artifact: dict[str, Any],
    *,
    feature_rows: list[list[float]] | None = None,
    closes: list[float] | None = None,
    candles: list[Candle] | None = None,
) -> float | dict[str, Any] | None:
    entry = _torch_recurrent_entry(symbol, artifact)
    model_name = _torch_recurrent_model_name(symbol, artifact)
    if not entry or not model_name:
        return None
    lookback = int(_clamp(_float_entry(entry, "lookback", 0.0), 4.0, 512.0))
    hidden_size = int(_clamp(_float_entry(entry, "hidden_size", 0.0), 1.0, 512.0))
    num_layers = int(_clamp(_float_entry(entry, "num_layers", 1.0), 1.0, 8.0))
    clip = _clamp(_float_entry(entry, "clip", 8.0), 1.0, 50.0)
    direct_horizon = _is_direct_horizon(entry)

    if direct_horizon:
        rows = feature_rows or []
        if len(rows) < lookback:
            return None
        sequence = _normalize_feature_rows(rows[-lookback:], entry, clip)
        target_mean = _float_entry(entry, "target_mean", 0.0)
        target_scale = max(_float_entry(entry, "target_scale", _return_scale(returns)), 1e-8)
    else:
        mean = _float_entry(entry, "mean", 0.0)
        scale = max(_float_entry(entry, "scale", _return_scale(returns)), 1e-8)
        if len(returns) < lookback:
            return None
        sequence = [[_clamp((value - mean) / scale, -clip, clip)] for value in returns[-lookback:]]
        target_mean = mean
        target_scale = scale

    try:
        hidden = _torch_recurrent_hidden(
            sequence,
            entry=entry,
            model_name=model_name,
            hidden_size=hidden_size,
            num_layers=num_layers,
        )
        if hidden is None:
            return None
        head_outputs = _torch_head_outputs(hidden, entry, hidden_size)
        if not head_outputs:
            return None
        if _entry_target_horizons(entry):
            return _decode_multi_horizon_prediction(
                outputs=head_outputs,
                entry=entry,
                returns=returns,
                closes=closes or [],
                candles=candles or [],
                clip=clip,
            )
        normalized_prediction = head_outputs[0]
        if not math.isfinite(normalized_prediction):
            return None
        prediction = _clamp(normalized_prediction, -clip, clip) * target_scale + target_mean
    except (IndexError, KeyError, TypeError, ValueError, OverflowError):
        return None

    if direct_horizon and closes:
        horizon = _entry_horizon(entry, 1)
        recent_abs = sorted(abs(value) for value in _horizon_log_returns(closes, horizon)[-48:])
    else:
        recent_abs = sorted(abs(value) for value in returns[-48:]) if len(returns) >= 8 else [0.01]
    cap = max(recent_abs[int(len(recent_abs) * 0.9)] if recent_abs else 0.0, 0.0002)
    return _clamp(prediction, -cap, cap)


def _torch_head_outputs(context: list[float], entry: dict[str, Any], hidden_size: int) -> list[float]:
    context = _apply_context_norm(context, entry)
    raw_weight = entry.get("head_weight")
    if isinstance(raw_weight, list) and raw_weight and isinstance(raw_weight[0], list):
        matrix = _float_matrix(raw_weight)
        bias = _float_vector(entry.get("head_bias"))
        if len(bias) != len(matrix):
            bias = [0.0 for _ in matrix]
        return [
            _dot(row, context) + bias[index]
            for index, row in enumerate(matrix)
            if len(row) == hidden_size
        ]
    head_weight = _float_vector(raw_weight)
    head_bias = _float_entry(entry, "head_bias", 0.0)
    if len(head_weight) != hidden_size:
        return []
    return [sum(weight * value for weight, value in zip(head_weight, context)) + head_bias]


def _apply_context_norm(context: list[float], entry: dict[str, Any]) -> list[float]:
    weight = _float_vector(entry.get("context_norm_weight"))
    bias = _float_vector(entry.get("context_norm_bias"))
    if not weight or len(weight) != len(context):
        return context
    if len(bias) != len(context):
        bias = [0.0 for _ in context]
    mean = sum(context) / len(context)
    variance = sum((value - mean) ** 2 for value in context) / len(context)
    denominator = math.sqrt(variance + 1e-5)
    return [
        ((value - mean) / denominator) * weight[index] + bias[index]
        for index, value in enumerate(context)
    ]


def _decode_multi_horizon_prediction(
    *,
    outputs: list[float],
    entry: dict[str, Any],
    returns: list[float],
    closes: list[float],
    candles: list[Candle],
    clip: float,
) -> dict[str, Any] | None:
    horizons = _entry_target_horizons(entry)
    if not horizons:
        return None
    layout = _entry_output_layout(entry)
    group_size = len(layout)
    if len(outputs) < len(horizons) * group_size:
        return None
    target_means = _target_vector(entry, "target_means", "target_mean", len(horizons), 0.0)
    target_scales = _target_vector(entry, "target_scales", "target_scale", len(horizons), _return_scale(returns))
    validation_mae = _target_vector(entry, "validation_mae_by_horizon", "validation_mae_percent", len(horizons), 0.0)
    baseline_mae = _target_vector(entry, "baseline_mae_by_horizon", "baseline_mae_percent", len(horizons), 0.0)
    round_trip_cost = max(0.0, _float_entry(entry, "round_trip_cost", 0.0))
    result: dict[str, Any] = {"horizons": {}}
    for horizon_index, horizon in enumerate(horizons):
        base = horizon_index * group_size
        values = {layout[offset]: outputs[base + offset] for offset in range(group_size)}
        vol_scale = _current_volatility_scale(candles, closes, horizon)

        def decode(name: str, fallback: float = 0.0) -> float:
            normalized = _clamp(float(values.get(name, fallback)), -clip, clip)
            transformed = normalized * max(target_scales[horizon_index], 1e-8) + target_means[horizon_index]
            if str(entry.get("target_transform", "")) == "net_return_over_volatility":
                return transformed * vol_scale
            return transformed

        expected = decode("mean")
        q_values = sorted([decode("q10", expected), decode("q50", expected), decode("q90", expected)])
        probability_up = _sigmoid(float(values.get("logit_up", 0.0)))
        cap = _prediction_cap(closes, horizon, round_trip_cost)
        expected = _clamp(expected, -cap, cap)
        q10 = _clamp(q_values[0], -cap, cap)
        q50 = _clamp(q_values[1], -cap, cap)
        q90 = _clamp(q_values[2], -cap, cap)
        mae = validation_mae[horizon_index]
        if mae > 1.0:
            mae = mae / 100
        base_mae = baseline_mae[horizon_index]
        if base_mae > 1.0:
            base_mae = base_mae / 100
        if mae <= 0:
            mae = max(_horizon_return_scale(closes, horizon), 1e-9)
        if base_mae <= 0:
            base_mae = max(mae, _horizon_return_scale(closes, horizon))
        uncertainty = max(abs(q90 - q10) * 0.5, mae, 1e-9)
        result["horizons"][str(horizon)] = {
            "horizon": horizon,
            "expected_return": expected,
            "expected_gross_return": expected + round_trip_cost,
            "q10": q10,
            "q50": q50,
            "q90": q90,
            "probability_up": probability_up,
            "volatility_scale": vol_scale,
            "validation_mae": mae,
            "baseline_mae": base_mae,
            "uncertainty": uncertainty,
        }
    return result


def _normalize_feature_rows(rows: list[list[float]], entry: dict[str, Any], clip: float) -> list[list[float]]:
    means = _float_vector(entry.get("feature_means"))
    scales = _float_vector(entry.get("feature_scales"))
    input_size = int(_clamp(_float_entry(entry, "input_size", len(rows[-1]) if rows else 1), 1.0, 256.0))
    if len(means) != input_size:
        means = [0.0 for _ in range(input_size)]
    if len(scales) != input_size:
        scales = [1.0 for _ in range(input_size)]
    normalized = []
    for row in rows:
        normalized.append(
            [
                _clamp(((row[index] if index < len(row) else 0.0) - means[index]) / max(scales[index], 1e-8), -clip, clip)
                for index in range(input_size)
            ]
        )
    return normalized


def _feature_snapshot(
    feature_rows: list[list[float]],
    entry: dict[str, Any] | None,
    clip: float,
) -> list[dict[str, Any]]:
    if not entry or not feature_rows:
        return []
    names = _feature_names(entry)
    latest = feature_rows[-1]
    normalized_rows = _normalize_feature_rows([latest], entry, clip)
    normalized = normalized_rows[-1] if normalized_rows else []
    means = _float_vector(entry.get("feature_means"))
    scales = _float_vector(entry.get("feature_scales"))
    snapshot: list[dict[str, Any]] = []
    for index, name in enumerate(names):
        raw_value = float(latest[index]) if index < len(latest) else 0.0
        model_value = float(normalized[index]) if index < len(normalized) else 0.0
        group, label, meaning = FEATURE_DESCRIPTIONS.get(
            name,
            ("Прочее", name, "Технический входной признак модели."),
        )
        snapshot.append(
            {
                "name": name,
                "label": label,
                "group": group,
                "raw_value": round(raw_value, 10),
                "raw_display": _feature_raw_display(name, raw_value),
                "model_value": round(model_value, 4),
                "model_display": f"{model_value:+.2f}",
                "mean": round(float(means[index]), 10) if index < len(means) else 0.0,
                "scale": round(float(scales[index]), 10) if index < len(scales) else 1.0,
                "meaning": meaning,
                "interpretation": _feature_interpretation(name, raw_value, model_value),
            }
        )
    return snapshot


def _feature_raw_display(name: str, value: float) -> str:
    if _feature_is_log_percent(name):
        return f"{(math.exp(value) - 1) * 100:+.3f}%"
    if _feature_is_linear_percent(name):
        return f"{value * 100:+.3f}%"
    if name in {"rsi_centered"}:
        return f"RSI {value * 50 + 50:.1f}"
    if name in {"rsi_slope_6"}:
        return f"{value * 50:+.2f} RSI"
    if name in {"volume_percentile_20", "range_position_50", "pattern_range_position_20"}:
        return f"{value * 100:.1f}%"
    if name.startswith("pattern_") and name != "pattern_score":
        return "да" if value >= 0.5 else "нет"
    if name == "pattern_score":
        return f"{value:.2f}"
    return f"{value:+.4f}"


def _feature_interpretation(name: str, value: float, model_value: float) -> str:
    norm = _model_value_text(model_value)
    if name.startswith("pattern_") and name != "pattern_score" and name != "pattern_range_position_20":
        state = "шаблон активен" if value >= 0.5 else "шаблон не активен"
        return f"{state}; {norm}."
    if name in {"volume_percentile_20", "range_position_50", "pattern_range_position_20"}:
        if value >= 0.8:
            state = "значение находится в верхней части диапазона"
        elif value <= 0.2:
            state = "значение находится в нижней части диапазона"
        else:
            state = "значение около середины диапазона"
        return f"{state}; {norm}."
    if name in {"volume_ratio", "atr_ratio_20"}:
        state = "выше среднего" if value > 0 else "ниже среднего" if value < 0 else "около среднего"
        return f"{state}; {norm}."
    if name == "rsi_centered":
        rsi = value * 50 + 50
        if rsi >= 65:
            state = "RSI высокий"
        elif rsi <= 35:
            state = "RSI низкий"
        else:
            state = "RSI в средней зоне"
        return f"{state}; {norm}."
    if _feature_is_log_percent(name) or _feature_is_linear_percent(name) or name.endswith("_slope"):
        if value > 0:
            state = "положительное значение"
        elif value < 0:
            state = "отрицательное значение"
        else:
            state = "нейтральное значение"
        return f"{state}; {norm}."
    if name == "pattern_score":
        if value >= 0.65:
            state = "шаблон скорее поддерживает long"
        elif value <= 0.35:
            state = "шаблон скорее против long"
        else:
            state = "шаблон нейтральный"
        return f"{state}; {norm}."
    return norm + "."


def _model_value_text(value: float) -> str:
    magnitude = abs(value)
    if magnitude >= 2.0:
        return "для модели это сильное отклонение от обучающей нормы"
    if magnitude >= 1.0:
        return "для модели это заметное отклонение от обучающей нормы"
    return "для модели это близко к обычному диапазону"


def _feature_is_log_percent(name: str) -> bool:
    return (
        name.startswith("return_")
        or name.startswith("btc_return_")
        or name.startswith("eth_return_")
        or name.startswith("relative_")
        or name in {
            "volume_change",
            "trend_return_4h",
            "trend_return_24h",
            "ema20_slope_6",
            "ema50_slope_12",
            "ema200_slope_24",
            "daily_ema50_slope",
            "btc_eth_return_spread_3",
        }
    )


def _feature_is_linear_percent(name: str) -> bool:
    return name in {
        "range_percent",
        "body_percent",
        "upper_wick_percent",
        "lower_wick_percent",
        "volume_ratio",
        "atr_percent",
        "atr_ratio_20",
        "realized_volatility_12",
        "realized_volatility_24",
        "macd_hist_percent",
        "macd_hist_slope_3",
        "ema50_gap_percent",
        "ema200_gap_percent",
        "ema50_ema200_gap_percent",
        "daily_close_ema200_gap_percent",
        "daily_ema50_ema200_gap_percent",
    }


def _torch_recurrent_hidden(
    sequence: list[list[float]],
    *,
    entry: dict[str, Any],
    model_name: str,
    hidden_size: int,
    num_layers: int,
) -> list[float] | None:
    state = entry.get("state_dict")
    if not isinstance(state, dict):
        return None
    h_layers = [[0.0 for _ in range(hidden_size)] for _ in range(num_layers)]
    c_layers = [[0.0 for _ in range(hidden_size)] for _ in range(num_layers)]
    top_outputs: list[list[float]] = []
    for row in sequence:
        layer_input = list(row)
        for layer in range(num_layers):
            if model_name == "torch_lstm":
                next_hidden, next_cell = _torch_lstm_step(layer_input, h_layers[layer], c_layers[layer], state, layer)
                h_layers[layer] = next_hidden
                c_layers[layer] = next_cell
            elif model_name == "torch_gru":
                h_layers[layer] = _torch_gru_step(layer_input, h_layers[layer], state, layer)
            else:
                return None
            layer_input = h_layers[layer]
        top_outputs.append(list(layer_input))
    if not top_outputs:
        return None
    if bool(entry.get("attention_pooling")):
        return _attention_context(top_outputs, entry, hidden_size)
    return top_outputs[-1]


def _attention_context(outputs: list[list[float]], entry: dict[str, Any], hidden_size: int) -> list[float] | None:
    weight = _float_vector(entry.get("attention_weight"))
    if len(weight) != hidden_size:
        return outputs[-1] if outputs else None
    bias = _float_entry(entry, "attention_bias", 0.0)
    scores = [_dot(weight, row) + bias for row in outputs]
    if not scores:
        return None
    max_score = max(scores)
    exps = [math.exp(_clamp(score - max_score, -50.0, 50.0)) for score in scores]
    total = sum(exps)
    if total <= 0:
        return outputs[-1]
    attention = [value / total for value in exps]
    return [
        sum(attention[row_index] * outputs[row_index][hidden_index] for row_index in range(len(outputs)))
        for hidden_index in range(hidden_size)
    ]


def _torch_lstm_step(
    inputs: list[float],
    hidden: list[float],
    cell: list[float],
    state: dict[str, Any],
    layer: int,
) -> tuple[list[float], list[float]]:
    hidden_size = len(hidden)
    gates = _torch_gate_values(inputs, hidden, state, layer, gate_count=4)
    input_gate = [_sigmoid(value) for value in gates[0]]
    forget_gate = [_sigmoid(value) for value in gates[1]]
    cell_gate = [math.tanh(value) for value in gates[2]]
    output_gate = [_sigmoid(value) for value in gates[3]]
    next_cell = [
        forget_gate[index] * cell[index] + input_gate[index] * cell_gate[index]
        for index in range(hidden_size)
    ]
    next_hidden = [
        output_gate[index] * math.tanh(next_cell[index])
        for index in range(hidden_size)
    ]
    return next_hidden, next_cell


def _torch_gru_step(
    inputs: list[float],
    hidden: list[float],
    state: dict[str, Any],
    layer: int,
) -> list[float]:
    hidden_size = len(hidden)
    weight_ih = _float_matrix(state[f"weight_ih_l{layer}"])
    weight_hh = _float_matrix(state[f"weight_hh_l{layer}"])
    bias_ih = _float_vector(state[f"bias_ih_l{layer}"])
    bias_hh = _float_vector(state[f"bias_hh_l{layer}"])

    def gate_input(gate: int) -> list[float]:
        start = gate * hidden_size
        output = []
        for index in range(hidden_size):
            row = start + index
            output.append(_dot(weight_ih[row], inputs) + bias_ih[row])
        return output

    def gate_hidden(gate: int) -> list[float]:
        start = gate * hidden_size
        output = []
        for index in range(hidden_size):
            row = start + index
            output.append(_dot(weight_hh[row], hidden) + bias_hh[row])
        return output

    reset_input = gate_input(0)
    update_input = gate_input(1)
    new_input = gate_input(2)
    reset_hidden = gate_hidden(0)
    update_hidden = gate_hidden(1)
    new_hidden = gate_hidden(2)

    reset_gate = [_sigmoid(reset_input[index] + reset_hidden[index]) for index in range(hidden_size)]
    update_gate = [_sigmoid(update_input[index] + update_hidden[index]) for index in range(hidden_size)]
    candidate = [
        math.tanh(new_input[index] + reset_gate[index] * new_hidden[index])
        for index in range(hidden_size)
    ]
    return [
        (1 - update_gate[index]) * candidate[index] + update_gate[index] * hidden[index]
        for index in range(hidden_size)
    ]


def _torch_gate_values(
    inputs: list[float],
    hidden: list[float],
    state: dict[str, Any],
    layer: int,
    gate_count: int,
) -> list[list[float]]:
    hidden_size = len(hidden)
    weight_ih = _float_matrix(state[f"weight_ih_l{layer}"])
    weight_hh = _float_matrix(state[f"weight_hh_l{layer}"])
    bias_ih = _float_vector(state[f"bias_ih_l{layer}"])
    bias_hh = _float_vector(state[f"bias_hh_l{layer}"])
    gates: list[list[float]] = []
    for gate in range(gate_count):
        values = []
        start = gate * hidden_size
        for index in range(hidden_size):
            row = start + index
            values.append(_dot(weight_ih[row], inputs) + _dot(weight_hh[row], hidden) + bias_ih[row] + bias_hh[row])
        gates.append(values)
    return gates


def _torch_validation_mae(entry: dict[str, Any], returns: list[float]) -> float:
    mae_percent = _float_entry(entry, "validation_mae_percent", 0.0)
    if mae_percent > 0:
        return mae_percent / 100
    return _return_scale(returns)


def _feature_names(entry: dict[str, Any] | None) -> list[str]:
    if not entry:
        return list(DEFAULT_TORCH_FEATURES)
    names = entry.get("feature_names")
    if isinstance(names, list) and names:
        return [str(name) for name in names]
    return list(DEFAULT_TORCH_FEATURES)


def _is_direct_horizon(entry: dict[str, Any]) -> bool:
    return bool(entry.get("direct_horizon")) or "target_horizon" in entry


def _entry_horizon(entry: dict[str, Any], default: int) -> int:
    horizons = _entry_target_horizons(entry)
    requested = int(_clamp(_float_entry(entry, "target_horizon", float(max(1, default))), 1.0, 96.0))
    if horizons:
        if requested in horizons:
            return requested
        return min(horizons, key=lambda value: abs(value - requested))
    return requested


def _entry_target_horizons(entry: dict[str, Any]) -> list[int]:
    raw = entry.get("target_horizons")
    if not isinstance(raw, list):
        return []
    horizons = []
    for value in raw:
        try:
            horizon = int(value)
        except (TypeError, ValueError):
            continue
        if 1 <= horizon <= 96 and horizon not in horizons:
            horizons.append(horizon)
    return horizons


def _entry_output_layout(entry: dict[str, Any]) -> list[str]:
    raw = entry.get("output_layout")
    if isinstance(raw, list) and raw:
        return [str(value) for value in raw]
    return ["mean", "q10", "q50", "q90", "logit_up"]


def _target_vector(
    entry: dict[str, Any],
    plural_key: str,
    scalar_key: str,
    size: int,
    default: float,
) -> list[float]:
    raw = entry.get(plural_key)
    if isinstance(raw, dict):
        values = []
        for horizon in _entry_target_horizons(entry):
            values.append(float(raw.get(str(horizon), default)))
        if len(values) == size:
            return values
    if isinstance(raw, list) and len(raw) == size:
        return [float(value) for value in raw]
    scalar = _float_entry(entry, scalar_key, default)
    return [scalar for _ in range(size)]


def _select_horizon_prediction(prediction: dict[str, Any], horizon: int) -> dict[str, Any] | None:
    horizons = prediction.get("horizons")
    if not isinstance(horizons, dict) or not horizons:
        return None
    key = str(horizon)
    selected = horizons.get(key)
    if isinstance(selected, dict):
        return selected
    numeric = []
    for raw_key, value in horizons.items():
        try:
            numeric.append((abs(int(raw_key) - horizon), value))
        except (TypeError, ValueError):
            continue
    numeric.sort(key=lambda item: item[0])
    return numeric[0][1] if numeric and isinstance(numeric[0][1], dict) else None


def _public_horizon_forecasts(prediction: dict[str, Any]) -> dict[str, Any]:
    horizons = prediction.get("horizons")
    if not isinstance(horizons, dict):
        return {}
    public: dict[str, Any] = {}
    for key, row in horizons.items():
        if not isinstance(row, dict):
            continue
        public[key] = {
            "expected_return_percent": round((math.exp(float(row.get("expected_return", 0.0))) - 1) * 100, 4),
            "probability_up": round(_clamp(float(row.get("probability_up", 0.5)), 0.0, 1.0), 4),
            "quantile_10_percent": round((math.exp(float(row.get("q10", 0.0))) - 1) * 100, 4),
            "quantile_50_percent": round((math.exp(float(row.get("q50", 0.0))) - 1) * 100, 4),
            "quantile_90_percent": round((math.exp(float(row.get("q90", 0.0))) - 1) * 100, 4),
        }
    return public


def _float_entry(data: dict[str, Any], key: str, default: float) -> float:
    value = data.get(key)
    if isinstance(value, (int, float)):
        return float(value)
    if isinstance(value, str):
        try:
            return float(value)
        except ValueError:
            return default
    return default


def _float_vector(data: Any) -> list[float]:
    if not isinstance(data, list):
        return []
    return [float(value) for value in data]


def _float_matrix(data: Any) -> list[list[float]]:
    if not isinstance(data, list):
        return []
    return [_float_vector(row) for row in data]


def _dot(left: list[float], right: list[float]) -> float:
    return sum(left[index] * right[index] for index in range(min(len(left), len(right))))


def _return_scale(returns: list[float]) -> float:
    recent = returns[-120:] if len(returns) > 120 else returns
    values = sorted(abs(value) for value in recent if math.isfinite(value))
    if not values:
        return 0.0005
    median = values[len(values) // 2]
    mean = sum(values) / len(values)
    return max(max(median, mean * 0.5), 1e-5)


def _horizon_log_returns(closes: list[float], horizon: int) -> list[float]:
    horizon = max(1, horizon)
    values = []
    for index in range(0, len(closes) - horizon):
        current = closes[index]
        future = closes[index + horizon]
        if current > 0 and future > 0:
            values.append(math.log(future / current))
    return values


def _horizon_return_scale(closes: list[float], horizon: int) -> float:
    values = _horizon_log_returns(closes, horizon)
    return _return_scale(values) if values else 0.0005


def _current_volatility_scale(candles: list[Candle], closes: list[float], horizon: int) -> float:
    horizon = max(1, horizon)
    latest = candles[-1] if candles else None
    close = closes[-1] if closes else (latest.close if latest else 0.0)
    atr_scale = 0.0
    if latest and latest.atr_14 is not None and close > 0:
        atr_scale = (latest.atr_14 / close) * math.sqrt(horizon)
    realized = _horizon_return_scale(closes, horizon)
    one_step = _return_scale(_log_returns(closes)) * math.sqrt(horizon) if len(closes) > 2 else 0.0
    return max(atr_scale * 0.7, realized, one_step, 0.0005)


def _prediction_cap(closes: list[float], horizon: int, round_trip_cost: float) -> float:
    values = sorted(abs(value) for value in _horizon_log_returns(closes, horizon)[-96:])
    base = values[int(len(values) * 0.9)] if values else 0.0
    return max(base * 1.5 + round_trip_cost, 0.0005)


def _sigmoid(value: float) -> float:
    if value >= 40:
        return 1.0
    if value <= -40:
        return 0.0
    return 1 / (1 + math.exp(-value))


def _confidence_adjustment(
    *,
    expected_return_percent: float,
    probability_up: float,
    skill: float,
    min_edge: float,
    max_adjustment: float,
) -> float:
    edge = abs(expected_return_percent) - min_edge
    if edge <= 0:
        return 0.0
    direction = 1.0 if expected_return_percent > 0 and probability_up >= 0.55 else -1.0
    if direction < 0 and probability_up > 0.45:
        return 0.0
    strength = _clamp(edge / max(min_edge, 0.05), 0.0, 1.0)
    probability_strength = _clamp(abs(probability_up - 0.5) / 0.25, 0.0, 1.0)
    skill_strength = _clamp((skill + 0.03) / 0.18, 0.25, 1.0)
    return direction * _clamp(max_adjustment, 0.0, 0.18) * strength * probability_strength * skill_strength


def _reason(
    *,
    model: str,
    expected_return_percent: float,
    probability_up: float,
    skill: float,
    block_entry: bool,
) -> str:
    if block_entry:
        return f"model {model}: expected move down {expected_return_percent:.3f}%, P(up)={probability_up:.2f}"
    return f"model {model}: forecast {expected_return_percent:.3f}%, P(up)={probability_up:.2f}, skill={skill:.3f}"


def _normal_cdf(value: float) -> float:
    return 0.5 * (1 + math.erf(value / math.sqrt(2)))


def _clamp(value: float, low: float, high: float) -> float:
    return max(low, min(high, value))