From f0cd5e00bbfed5243d7029e0682d6774cdfd7790 Mon Sep 17 00:00:00 2001 From: "zhangkun9038@dingtalk.com" Date: Mon, 2 Jun 2025 01:14:48 +0000 Subject: [PATCH] =?UTF-8?q?=E5=8A=A8=E6=80=81=E8=AE=A1=E7=AE=97=E9=98=88?= =?UTF-8?q?=E5=80=BCfuture=5Fadx=5Fthreshold?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- freqtrade/templates/freqaiprimer.py | 266 +++++++++++++++++++--------- 1 file changed, 185 insertions(+), 81 deletions(-) diff --git a/freqtrade/templates/freqaiprimer.py b/freqtrade/templates/freqaiprimer.py index 5278c7b0..e4802a97 100644 --- a/freqtrade/templates/freqaiprimer.py +++ b/freqtrade/templates/freqaiprimer.py @@ -8,11 +8,10 @@ from datetime import datetime import os import json import glob -from functools import reduce logger = logging.getLogger(__name__) -class FreqaiPrimer(IStrategy): +class HybridStrategy(IStrategy): minimal_roi = { "0": 0.02, "30": 0.01, @@ -42,7 +41,7 @@ class FreqaiPrimer(IStrategy): "verbose": -1, }, "fit_live_predictions_candles": 100, - "live_retrain_candles": 100, + "live_retrain_candles": 50, } def __init__(self, config: dict, *args, **kwargs): @@ -56,6 +55,105 @@ class FreqaiPrimer(IStrategy): self.stats_logged = False self.buy_threshold = -0.01 self.sell_threshold = 0.01 + self.prev_future_adx_mean = None + self.prev_future_adx_std = None + self.prev_price_value_divergence_mean = None + self.prev_price_value_divergence_std = None + + def load_metadata_stats(self, pair: str, identifier: str, dataframe: DataFrame) -> tuple[dict, dict]: + """ + 从 metadata.json 文件读取 labels_mean 和 labels_std，并应用平滑逻辑。 + + Args: + pair (str): 币对名称，例如 "BTC/USDT"。 + identifier (str): FreqAI 模型标识，例如 "test58"。 + dataframe (DataFrame): 用于回退计算的数据框架。 + + Returns: + tuple: (平滑后的 labels_mean, 平滑后的 labels_std) + """ + labels_mean = None + labels_std = None + pair_base = pair.split('/')[0] if '/' in pair else pair + + try: + model_base_dir = os.path.join(self.config["user_data_dir"], "models", identifier) + sub_dirs = glob.glob(os.path.join(model_base_dir, f"sub-train-{pair_base}_*")) + + if not sub_dirs: + logger.warning(f"[{pair}] 未找到任何子目录：{model_base_dir}/sub-train-{pair_base}_*") + else: + latest_sub_dir = max(sub_dirs, key=lambda x: int(x.split('_')[-1])) + pair_base_lower = pair_base.lower() + timestamp = latest_sub_dir.split('_')[-1] + metadata_file = os.path.join(latest_sub_dir, f"cb_{pair_base_lower}_{timestamp}_metadata.json") + + if os.path.exists(metadata_file): + with open(metadata_file, "r") as f: + metadata = json.load(f) + labels_mean = metadata["labels_mean"] + labels_std = metadata["labels_std"] + logger.info(f"[{pair}] 从最新子目录 {latest_sub_dir} 读取 labels_mean：{labels_mean}, labels_std：{labels_std}") + else: + logger.warning(f"[{pair}] 最新的 metadata.json 文件 {metadata_file} 不存在") + except Exception as e: + logger.warning(f"[{pair}] 无法从子目录读取 labels_mean 和 labels_std：{e}，重新计算") + + # 回退计算 + if labels_mean is None or labels_std is None: + logger.warning(f"[{pair}] 无法获取 labels_mean 和 labels_std，重新计算") + dataframe["&-price_value_divergence_actual"] = (dataframe["close"] - dataframe["ema200"]) / dataframe["ema200"] + dataframe["&-price_value_divergence_actual"] = dataframe["&-price_value_divergence_actual"].replace([np.inf, -np.inf], 0).ffill().fillna(0) + recent_data = dataframe["&-price_value_divergence_actual"].tail(self.fit_live_predictions_candles) + labels_mean = { + "&-future_adx": dataframe["adx"].tail(self.fit_live_predictions_candles).mean(), + "&-price_value_divergence": recent_data.mean() + } + labels_std = { + "&-future_adx": dataframe["adx"].tail(self.fit_live_predictions_candles).std(), + "&-price_value_divergence": recent_data.std() + } + if np.isnan(labels_std["&-future_adx"]) or labels_std["&-future_adx"] == 0: + labels_std["&-future_adx"] = 1.0 + if np.isnan(labels_std["&-price_value_divergence"]) or labels_std["&-price_value_divergence"] == 0: + labels_std["&-price_value_divergence"] = 0.01 + logger.warning(f"[{pair}] labels_std 计算异常，使用默认值") + + # 平滑"&-future_adx"的均值和标准差 + current_future_adx_mean = labels_mean["&-future_adx"] + current_future_adx_std = labels_std["&-future_adx"] + if self.prev_future_adx_mean is not None and self.prev_future_adx_std is not None: + smoothed_future_adx_mean = 0.7 * current_future_adx_mean + 0.3 * self.prev_future_adx_mean + smoothed_future_adx_std = 0.7 * current_future_adx_std + 0.3 * self.prev_future_adx_std + else: + smoothed_future_adx_mean = current_future_adx_mean + smoothed_future_adx_std = current_future_adx_std + self.prev_future_adx_mean = smoothed_future_adx_mean + self.prev_future_adx_std = smoothed_future_adx_std + + # 平滑"&-price_value_divergence"的均值和标准差 + current_price_value_divergence_mean = labels_mean["&-price_value_divergence"] + current_price_value_divergence_std = labels_std["&-price_value_divergence"] + if self.prev_price_value_divergence_mean is not None and self.prev_price_value_divergence_std is not None: + smoothed_price_value_divergence_mean = 0.7 * current_price_value_divergence_mean + 0.3 * self.prev_price_value_divergence_mean + smoothed_price_value_divergence_std = 0.7 * current_price_value_divergence_std + 0.3 * self.prev_price_value_divergence_std + else: + smoothed_price_value_divergence_mean = current_price_value_divergence_mean + smoothed_price_value_divergence_std = current_price_value_divergence_std + self.prev_price_value_divergence_mean = smoothed_price_value_divergence_mean + self.prev_price_value_divergence_std = smoothed_price_value_divergence_std + + # 更新平滑后的值 + smoothed_labels_mean = { + "&-future_adx": smoothed_future_adx_mean, + "&-price_value_divergence": smoothed_price_value_divergence_mean + } + smoothed_labels_std = { + "&-future_adx": smoothed_future_adx_std, + "&-price_value_divergence": smoothed_price_value_divergence_std + } + + return smoothed_labels_mean, smoothed_labels_std def feature_engineering_expand_all(self, dataframe: DataFrame, period: int, metadata: dict, **kwargs) -> DataFrame: dataframe["%-rsi-period"] = ta.RSI(dataframe, timeperiod=period) @@ -148,83 +246,60 @@ class FreqaiPrimer(IStrategy): for col in ["ema200", "ema200_slope", "ema20", "ema50", "adx", "atr", "atr_median", "bb_upperband", "bb_middleband", "bb_lowerband", "rsi", "volume_z_score", "&-future_adx", "&-price_value_divergence"]: dataframe[col] = dataframe[col].replace([np.inf, -np.inf], 0).ffill().fillna(0) - # 动态阈值计算 - labels_mean = None - labels_std = None - try: - model_base_dir = os.path.join(self.config["user_data_dir"], "models", self.freqai_info["identifier"]) - pair_base = pair.split('/')[0] if '/' in pair else pair - sub_dirs = glob.glob(os.path.join(model_base_dir, f"sub-train-{pair_base}_*")) - - if not sub_dirs: - logger.warning(f"[{pair}] 未找到任何子目录：{model_base_dir}/sub-train-{pair_base}_*") - else: - latest_sub_dir = max(sub_dirs, key=lambda x: int(x.split('_')[-1])) - pair_base_lower = pair_base.lower() - timestamp = latest_sub_dir.split('_')[-1] - metadata_file = os.path.join(latest_sub_dir, f"cb_{pair_base_lower}_{timestamp}_metadata.json") - - if os.path.exists(metadata_file): - with open(metadata_file, "r") as f: - metadata = json.load(f) - labels_mean = metadata["labels_mean"]["&-price_value_divergence"] - labels_std = metadata["labels_std"]["&-price_value_divergence"] - logger.info(f"[{pair}] 从最新子目录 {latest_sub_dir} 读取 labels_mean：{labels_mean}, labels_std：{labels_std}") - else: - logger.warning(f"[{pair}] 最新的 metadata.json 文件 {metadata_file} 不存在") - except Exception as e: - logger.warning(f"[{pair}] 无法从子目录读取 labels_mean 和 labels_std：{e}，重新计算") - - if labels_mean is None or labels_std is None: - logger.warning(f"[{pair}] 无法获取 labels_mean 和 labels_std，重新计算") - dataframe["&-price_value_divergence_actual"] = (dataframe["close"] - dataframe["ema200"]) / dataframe["ema200"] - dataframe["&-price_value_divergence_actual"] = dataframe["&-price_value_divergence_actual"].replace([np.inf, -np.inf], 0).ffill().fillna(0) - recent_data = dataframe["&-price_value_divergence_actual"].tail(self.fit_live_predictions_candles) - labels_mean = recent_data.mean() - labels_std = recent_data.std() - if np.isnan(labels_std) or labels_std == 0: - labels_std = 0.01 - logger.warning(f"[{pair}] labels_std 计算异常，使用默认值 0.01") - + # 使用抽离的函数读取并平滑统计数据 + labels_mean, labels_std = self.load_metadata_stats(pair, self.freqai_info["identifier"], dataframe) self.pair_stats[pair] = {"labels_mean": labels_mean, "labels_std": labels_std} - # 根据市场状态调整系数 - market_state = "trending" if ((dataframe["adx"].iloc[-1] > 25) or (dataframe["&-future_adx"].iloc[-1] > 25)) and \ - (dataframe["ema200_slope"].iloc[-1] > 0.0005) and \ + # 动态调整"&-future_adx"阈值 + smoothed_future_adx_mean = labels_mean["&-future_adx"] + smoothed_future_adx_std = labels_std["&-future_adx"] + + if smoothed_future_adx_std > 9: # 高波动 + k_adx = 0.3 + elif smoothed_future_adx_std <= 5: # 低波动 + k_adx = 0.7 + else: # 中等波动 + k_adx = 0.5 + + future_adx_threshold = smoothed_future_adx_mean + k_adx * smoothed_future_adx_std + future_adx_threshold = max(future_adx_threshold, 20) + future_adx_threshold = min(future_adx_threshold, 35) + + # 市场状态切换 + market_state = "trending" if ((dataframe["adx"].iloc[-1] > 25) or (dataframe["&-future_adx"].iloc[-1] > future_adx_threshold)) and \ + (dataframe["ema200_slope"].iloc[-1] > 0) and \ (dataframe["atr"].iloc[-1] > dataframe["atr_median"].iloc[-1]) else "ranging" - if labels_std > 0.015: - k_buy = 1.2 - k_sell = 1.5 - elif labels_std < 0.010: - k_buy = 0.8 - k_sell = 1.0 - else: - k_buy = 1.0 - k_sell = 1.2 + # 动态调整"&-price_value_divergence"阈值 + labels_std_pvd = labels_std["&-price_value_divergence"] + labels_mean_pvd = labels_mean["&-price_value_divergence"] - if labels_mean > 0.015: + if labels_std_pvd > 0.01: # 高波动 + k_buy = 1.4 if market_state == "trending" else 1.2 + k_sell = 1.6 if market_state == "trending" else 1.4 + elif labels_std_pvd < 0.005: # 低波动 + k_buy = 1.1 if market_state == "trending" else 0.9 + k_sell = 1.3 if market_state == "trending" else 1.1 + else: # 中等波动 + k_buy = 1.2 if market_state == "trending" else 1.0 + k_sell = 1.4 if market_state == "trending" else 1.2 + + if labels_mean_pvd > 0.015: k_sell += 0.5 - logger.info(f"[{pair}] labels_mean 较高（{labels_mean:.4f}），增加 k_sell 到 {k_sell:.2f}") + logger.info(f"[{pair}] labels_mean_pvd 较高（{labels_mean_pvd:.4f}），增加 k_sell 到 {k_sell:.2f}") - # 趋势市场放宽阈值，震荡市场更严格 - if market_state == "trending": - k_buy *= 1.2 # 趋势市场允许更低的买入点 - k_sell *= 0.9 # 趋势市场更早退出 - else: - k_buy *= 0.9 # 震荡市场更谨慎买入 - k_sell *= 1.1 # 震荡市场更晚退出 + self.buy_threshold = labels_mean_pvd - k_buy * labels_std_pvd + self.sell_threshold = labels_mean_pvd + k_sell * labels_std_pvd - self.buy_threshold = labels_mean - k_buy * labels_std - self.sell_threshold = labels_mean + k_sell * labels_std + # 收紧阈值范围 + self.buy_threshold = max(self.buy_threshold, -0.03) + self.buy_threshold = min(self.buy_threshold, -0.002) + self.sell_threshold = min(self.sell_threshold, 0.03) + self.sell_threshold = max(self.sell_threshold, 0.002) - # 设置阈值范围 - self.buy_threshold = max(self.buy_threshold, -0.05) - self.buy_threshold = min(self.buy_threshold, -0.005) - self.sell_threshold = min(self.sell_threshold, 0.05) - self.sell_threshold = max(self.sell_threshold, 0.005) - - logger.info(f"[{pair}] 市场状态：{market_state}, labels_mean：{labels_mean:.4f}, labels_std：{labels_std:.4f}") + # 日志记录 + logger.info(f"[{pair}] 市场状态：{market_state}, labels_mean：{labels_mean}, labels_std：{labels_std}") + logger.info(f"[{pair}] k_adx：{k_adx:.2f}, &-future_adx 动态阈值：{future_adx_threshold:.2f}") logger.info(f"[{pair}] k_buy：{k_buy:.2f}, k_sell：{k_sell:.2f}") logger.info(f"[{pair}] 动态买入阈值：{self.buy_threshold:.4f}") logger.info(f"[{pair}] 动态卖出阈值：{self.sell_threshold:.4f}") @@ -232,7 +307,7 @@ class FreqaiPrimer(IStrategy): if not self.stats_logged: logger.info("===== 所有币对的 labels_mean 和 labels_std 汇总 =====") for p, stats in self.pair_stats.items(): - logger.info(f"[{p}] labels_mean：{stats['labels_mean']:.4f}, labels_std：{stats['labels_std']:.4f}") + logger.info(f"[{p}] labels_mean：{stats['labels_mean']}, labels_std：{stats['labels_std']}") logger.info("==============================================") self.stats_logged = True @@ -243,10 +318,18 @@ class FreqaiPrimer(IStrategy): def populate_entry_trend(self, dataframe: DataFrame, metadata: dict) -> DataFrame: pair = metadata.get('pair', 'Unknown') - # 使用ADX、FreqAI预测的ADX、EMA 200斜率和ATR判断市场状态 + # 动态阈值已在 populate_indicators 中计算，此处直接使用 + labels_mean = self.pair_stats.get(pair, {}).get("labels_mean", {"&-future_adx": 25}) + labels_std = self.pair_stats.get(pair, {}).get("labels_std", {"&-future_adx": 1}) + k_adx = 0.5 if labels_std["&-future_adx"] <= 9 and labels_std["&-future_adx"] > 5 else (0.3 if labels_std["&-future_adx"] > 9 else 0.7) + future_adx_threshold = labels_mean["&-future_adx"] + k_adx * labels_std["&-future_adx"] + future_adx_threshold = max(future_adx_threshold, 20) + future_adx_threshold = min(future_adx_threshold, 35) + + # 使用动态阈值判断市场状态 dataframe["market_state"] = np.where( - ((dataframe["adx"] > 25) | (dataframe["&-future_adx"] > 25)) & - (dataframe["ema200_slope"] > 0.0005) & + ((dataframe["adx"] > 25) | (dataframe["&-future_adx"] > future_adx_threshold)) & + (dataframe["ema200_slope"] > 0) & (dataframe["atr"] > dataframe["atr_median"]), "trending", "ranging" @@ -317,8 +400,15 @@ class FreqaiPrimer(IStrategy): hold_time = (current_time - trade.open_date_utc).total_seconds() / 60 # 确定市场状态 - market_state = "trending" if ((trade.dataframe["adx"].iloc[-1] > 25) or (trade.dataframe["&-future_adx"].iloc[-1] > 25)) and \ - (trade.dataframe["ema200_slope"].iloc[-1] > 0.0005) and \ + labels_mean = self.pair_stats.get(trade.pair, {}).get("labels_mean", {"&-future_adx": 25}) + labels_std = self.pair_stats.get(trade.pair, {}).get("labels_std", {"&-future_adx": 1}) + k_adx = 0.5 if labels_std["&-future_adx"] <= 9 and labels_std["&-future_adx"] > 5 else (0.3 if labels_std["&-future_adx"] > 9 else 0.7) + future_adx_threshold = labels_mean["&-future_adx"] + k_adx * labels_std["&-future_adx"] + future_adx_threshold = max(future_adx_threshold, 20) + future_adx_threshold = min(future_adx_threshold, 35) + + market_state = "trending" if ((trade.dataframe["adx"].iloc[-1] > 25) or (trade.dataframe["&-future_adx"].iloc[-1] > future_adx_threshold)) and \ + (trade.dataframe["ema200_slope"].iloc[-1] > 0) and \ (trade.dataframe["atr"].iloc[-1] > trade.dataframe["atr_median"].iloc[-1]) else "ranging" # 最小持仓时间 @@ -351,8 +441,15 @@ class FreqaiPrimer(IStrategy): def confirm_trade_entry(self, pair: str, order_type: str, amount: float, rate: float, time_in_force: str, current_time: datetime, **kwargs) -> bool: # 确定市场状态 - market_state = "trending" if ((self.dataframe["adx"].iloc[-1] > 25) or (self.dataframe["&-future_adx"].iloc[-1] > 25)) and \ - (self.dataframe["ema200_slope"].iloc[-1] > 0.0005) and \ + labels_mean = self.pair_stats.get(pair, {}).get("labels_mean", {"&-future_adx": 25}) + labels_std = self.pair_stats.get(pair, {}).get("labels_std", {"&-future_adx": 1}) + k_adx = 0.5 if labels_std["&-future_adx"] <= 9 and labels_std["&-future_adx"] > 5 else (0.3 if labels_std["&-future_adx"] > 9 else 0.7) + future_adx_threshold = labels_mean["&-future_adx"] + k_adx * labels_std["&-future_adx"] + future_adx_threshold = max(future_adx_threshold, 20) + future_adx_threshold = min(future_adx_threshold, 35) + + market_state = "trending" if ((self.dataframe["adx"].iloc[-1] > 25) or (self.dataframe["&-future_adx"].iloc[-1] > future_adx_threshold)) and \ + (self.dataframe["ema200_slope"].iloc[-1] > 0) and \ (self.dataframe["atr"].iloc[-1] > self.dataframe["atr_median"].iloc[-1]) else "ranging" # 根据市场状态调整冷却期 @@ -374,8 +471,15 @@ class FreqaiPrimer(IStrategy): rate: float, time_in_force: str, exit_reason: str, current_time: datetime, **kwargs) -> bool: # 确定市场状态 - market_state = "trending" if ((trade.dataframe["adx"].iloc[-1] > 25) or (trade.dataframe["&-future_adx"].iloc[-1] > 25)) and \ - (trade.dataframe["ema200_slope"].iloc[-1] > 0.0005) and \ + labels_mean = self.pair_stats.get(pair, {}).get("labels_mean", {"&-future_adx": 25}) + labels_std = self.pair_stats.get(pair, {}).get("labels_std", {"&-future_adx": 1}) + k_adx = 0.5 if labels_std["&-future_adx"] <= 9 and labels_std["&-future_adx"] > 5 else (0.3 if labels_std["&-future_adx"] > 9 else 0.7) + future_adx_threshold = labels_mean["&-future_adx"] + k_adx * labels_std["&-future_adx"] + future_adx_threshold = max(future_adx_threshold, 20) + future_adx_threshold = min(future_adx_threshold, 35) + + market_state = "trending" if ((trade.dataframe["adx"].iloc[-1] > 25) or (trade.dataframe["&-future_adx"].iloc[-1] > future_adx_threshold)) and \ + (trade.dataframe["ema200_slope"].iloc[-1] > 0) and \ (trade.dataframe["atr"].iloc[-1] > trade.dataframe["atr_median"].iloc[-1]) else "ranging" # 调整卖出价格