phyer/myTestFreqAI
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

dry-run时表现稳定, hyperopt表现也20% profit, 不错, 需要长时间运行看看了

Browse Source
This commit is contained in:
zhangkun9038@dingtalk.com 2025-04-24 19:43:06 +08:00
parent a32c505cff
commit f815ee89ee
4 changed files with 194 additions and 177 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

24
config_examples/config_freqai.okx.json
View File

@ -5,10 +5,11 @@
"max_open_trades": 4,
"stake_currency": "USDT",
"stake_amount": 150,
"startup_candle_count": 30,
"tradable_balance_ratio": 1,
"fiat_display_currency": "USD",
"dry_run": true,
"timeframe": "3m",
"timeframe": "5m",
"dry_run_wallet": 1000,
"cancel_open_orders_on_exit": true,
"stoploss": -0.05,
@ -30,7 +31,7 @@
},
"ccxt_async_config": {
"enableRateLimit": true,
"rateLimit": 500,
"rateLimit": 1000,
"timeout": 20000
},
"pair_whitelist": [
@ -66,8 +67,7 @@
},
"freqaimodel": "CatboostClassifier",
"purge_old_models": 2,
"train_period_days": 15,
"identifier": "test93",
"identifier": "test130",
"train_period_days": 30,
"backtest_period_days": 10,
"live_retrain_hours": 0,
@ -76,15 +76,14 @@
},
"feature_parameters": {
"include_timeframes": [
"3m",
"15m",
"5m",
"1h"
],
"include_corr_pairlist": [
"BTC/USDT",
"SOL/USDT"
],
"label_period_candles": 20,
"label_period_candles": 12,
"include_shifted_candles": 3,
"DI_threshold": 0.9,
"weight_factor": 0.9,
@ -98,13 +97,14 @@
"plot_feature_importances": 0
},
"data_split_parameters": {
"test_size": 0.2
"test_size": 0.2,
"shuffle": false,
},
"model_training_parameters": {
"n_estimators": 100,
"learning_rate": 0.05,
"max_depth": 5,
"num_leaves": 31
"learning_rate": 0.1,
"num_leaves": 15,
"verbose": -1
}
},
"api_server": {
@ -123,7 +123,7 @@
"initial_state": "running",
"force_entry_enable": false,
"internals": {
"process_throttle_secs": 5,
"process_throttle_secs": 10,
"heartbeat_interval": 20,
"loglevel": "DEBUG"
}

2
docker-compose.yml
View File

@ -35,7 +35,7 @@ services:
trade
--logfile /freqtrade/user_data/logs/freqtrade.log
--db-url sqlite:////freqtrade/user_data/tradesv3.sqlite
--freqaimodel LightGBMRegressor
--freqaimodel XGBoostRegressor
--config /freqtrade/config_examples/config_freqai.okx.json
--strategy FreqaiExampleStrategy
--strategy-path /freqtrade/templates

20
freqtrade/templates/FreqaiExampleStrategy.json
View File

@ -5,28 +5,28 @@
"max_open_trades": 4
},
"buy": {
"buy_rsi": 49
"buy_rsi": 37
},
"sell": {
"sell_rsi": 64
"sell_rsi": 80
},
"protection": {},
"roi": {
"0": 0.07600000000000001,
"7": 0.034,
"13": 0.007,
"60": 0
"0": 0.124,
"14": 0.023,
"37": 0.011,
"50": 0
},
"stoploss": {
"stoploss": -0.087
"stoploss": -0.168
},
"trailing": {
"trailing_stop": true,
"trailing_stop_positive": 0.333,
"trailing_stop_positive_offset": 0.341,
"trailing_stop_positive": 0.047,
"trailing_stop_positive_offset": 0.051000000000000004,
"trailing_only_offset_is_reached": true
}
},
"ft_stratparam_v": 1,
"export_time": "2025-04-23 15:53:46.477203+00:00"
"export_time": "2025-04-24 11:42:35.037486+00:00"
}

325
freqtrade/templates/FreqaiExampleStrategy.py
View File

@ -10,7 +10,6 @@ from freqtrade.strategy import IStrategy, IntParameter, DecimalParameter
logger = logging.getLogger(__name__)
class FreqaiExampleStrategy(IStrategy):
## minimal_roi 设置为{} 利润稍高, 回头再说,
minimal_roi = {
"0": 0.076,
"7": 0.034,
@ -34,48 +33,20 @@ class FreqaiExampleStrategy(IStrategy):
stoploss_param = DecimalParameter(low=-0.25, high=-0.05, default=-0.1, space="stoploss", optimize=True, load=True)
trailing_stop_positive_offset = DecimalParameter(low=0.01, high=0.5, default=0.02, space="trailing", optimize=True, load=True)
# 以下配置 利润低, 但是 更符合 损失函数的评估
# minimal_roi = {
# "0": 0.076,
# "7": 0.034,
# "13": 0.007,
# "60": 0
# }
# stoploss = -0.087
# trailing_stop = True
# trailing_stop_positive = 0.333
# trailing_stop_positive_offset = 0.341
# trailing_only_offset_is_reached = True
# process_only_new_candles = True
# use_exit_signal = True
# startup_candle_count: int = 40
# can_short = False
# max_open_trades = 4
#
# # Hyperopt 参数
# buy_rsi = IntParameter(low=10, high=50, default=49, space="buy", optimize=True, load=True)
# sell_rsi = IntParameter(low=50, high=90, default=64, space="sell", optimize=True, load=True)
# roi_0 = DecimalParameter(low=0.01, high=0.2, default=0.076, space="roi", optimize=True, load=True)
# roi_15 = DecimalParameter(low=0.005, high=0.1, default=0.034, space="roi", optimize=True, load=True)
# roi_30 = DecimalParameter(low=0.001, high=0.05, default=0.007, space="roi", optimize=True, load=True)
# stoploss_param = DecimalParameter(low=-0.25, high=-0.05, default=-0.087, space="stoploss", optimize=True, load=True)
# trailing_stop_positive_offset = DecimalParameter(low=0.01, high=0.5, default=0.341, space="trailing", optimize=True, load=True)
# 保护机制
protections = [
{"method": "StoplossGuard", "stop_duration": 60, "lookback_period": 120},
{"method": "MaxDrawdown", "lookback_period": 120, "max_allowed_drawdown": 0.05}
]
# FreqAI 配置
freqai_info = {
"model": "LightGBMRegressor",
"feature_parameters": {
"include_timeframes": ["5m", "15m", "1h"],
"include_corr_pairlist": [],
"include_timeframes": ["5m"],
"include_corr_pairlist": ["SOL/USDT", "BTC/USDT"],
"label_period_candles": 12,
"include_shifted_candles": 3,
"include_shifted_candles": 0,
"include_periods": [10, 20],
"DI_threshold": 3.0
},
"data_split_parameters": {
"test_size": 0.2,
@ -84,8 +55,9 @@ class FreqaiExampleStrategy(IStrategy):
"model_training_parameters": {
"n_estimators": 100,
"learning_rate": 0.1,
"num_leaves": 15, # 降低以减少警告
"verbose": -1,
"num_leaves": 15,
"n_jobs": 4,
"verbosity": -1
},
}
@ -102,149 +74,195 @@ class FreqaiExampleStrategy(IStrategy):
def feature_engineering_expand_all(self, dataframe: DataFrame, period: int, metadata: dict, **kwargs) -> DataFrame:
dataframe["%-rsi-period"] = ta.RSI(dataframe, timeperiod=period)
dataframe["%-mfi-period"] = ta.MFI(dataframe, timeperiod=period)
dataframe["%-sma-period"] = ta.SMA(dataframe, timeperiod=period)
dataframe["%-ema-period"] = ta.EMA(dataframe, timeperiod=period)
dataframe["%-adx-period"] = ta.ADX(dataframe, timeperiod=period)
bollinger = qtpylib.bollinger_bands(qtpylib.typical_price(dataframe), window=period, stds=2.2)
dataframe["bb_lowerband-period"] = bollinger["lower"]
dataframe["bb_middleband-period"] = bollinger["mid"]
dataframe["bb_upperband-period"] = bollinger["upper"]
dataframe["%-bb_width-period"] = (
dataframe["bb_upperband-period"] - dataframe["bb_lowerband-period"]
) / dataframe["bb_middleband-period"]
dataframe["%-close-bb_lower-period"] = dataframe["close"] / dataframe["bb_lowerband-period"]
dataframe["%-roc-period"] = ta.ROC(dataframe, timeperiod=period)
dataframe["%-relative_volume-period"] = (
dataframe["volume"] / dataframe["volume"].rolling(period).mean()
)
dataframe = dataframe.replace([np.inf, -np.inf], 0)
dataframe = dataframe.ffill()
dataframe = dataframe.fillna(0)
dataframe["%-bb_width-period"] = (bollinger["upper"] - bollinger["lower"]) / bollinger["mid"]
dataframe = dataframe.replace([np.inf, -np.inf], 0).ffill().fillna(0)
return dataframe
def feature_engineering_expand_basic(self, dataframe: DataFrame, metadata: dict, **kwargs) -> DataFrame:
dataframe["%-pct-change"] = dataframe["close"].pct_change()
dataframe["%-raw_volume"] = dataframe["volume"]
dataframe["%-raw_price"] = dataframe["close"]
dataframe = dataframe.replace([np.inf, -np.inf], 0)
dataframe = dataframe.ffill()
dataframe = dataframe.fillna(0)
dataframe = dataframe.replace([np.inf, -np.inf], 0).ffill().fillna(0)
return dataframe
def feature_engineering_standard(self, dataframe: DataFrame, metadata: dict, **kwargs) -> DataFrame:
if len(dataframe) < 20 or dataframe["close"].isna().any():
logger.warning(f"DataFrame too short ({len(dataframe)} rows) or contains NaN in close, cannot compute %-volatility")
dataframe["%-volatility"] = 0
else:
dataframe["%-volatility"] = dataframe["close"].pct_change().rolling(20).std()
dataframe["%-volatility"] = dataframe["%-volatility"].fillna(0)
if dataframe["%-volatility"].std() > 0:
dataframe["%-volatility"] = (dataframe["%-volatility"] - dataframe["%-volatility"].mean()) / dataframe["%-volatility"].std()
dataframe["%-day_of_week"] = dataframe["date"].dt.dayofweek
dataframe["%-hour_of_day"] = dataframe["date"].dt.hour
dataframe = dataframe.replace([np.inf, -np.inf], 0)
dataframe = dataframe.ffill()
dataframe = dataframe.fillna(0)
dataframe = dataframe.replace([np.inf, -np.inf], 0).ffill().fillna(0)
return dataframe
def set_freqai_targets(self, dataframe: DataFrame, metadata: dict, **kwargs) -> DataFrame:
logger.info(f"设置 FreqAI 目标，交易对：{metadata['pair']}")
logger.info(f"Setting FreqAI targets for pair: {metadata['pair']}")
if "close" not in dataframe.columns:
logger.error("数据框缺少必要的 'close' 列")
raise ValueError("数据框缺少必要的 'close' 列")
logger.error("DataFrame missing required 'close' column")
raise ValueError("DataFrame missing required 'close' column")
try:
label_period = self.freqai_info["feature_parameters"]["label_period_candles"]
dataframe["%-volatility"] = dataframe["close"].pct_change().rolling(20).std()
if len(dataframe) < 20 or dataframe["close"].isna().any():
logger.warning(f"DataFrame too short ({len(dataframe)} rows) or contains NaN in close, cannot compute %-volatility")
dataframe["%-volatility"] = 0
else:
dataframe["%-volatility"] = dataframe["close"].pct_change().rolling(20).std()
dataframe["%-volatility"] = dataframe["%-volatility"].fillna(0)
if dataframe["%-volatility"].std() > 0:
dataframe["%-volatility"] = (dataframe["%-volatility"] - dataframe["%-volatility"].mean()) / dataframe["%-volatility"].std()
dataframe["&-buy_rsi"] = ta.RSI(dataframe, timeperiod=14).shift(-label_period)
for col in ["&-buy_rsi", "%-volatility"]:
dataframe[col] = dataframe[col].replace([np.inf, -np.inf], 0)
dataframe[col] = dataframe[col].ffill()
dataframe[col] = dataframe[col].fillna(0)
dataframe[col] = dataframe[col].ffill().fillna(0)
if dataframe[col].isna().any():
logger.warning(f"目标列 {col} 仍包含 NaN，检查数据生成逻辑")
logger.warning(f"Target column {col} still contains NaN, check data generation logic")
except Exception as e:
logger.error(f"创建 FreqAI 目标失败：{str(e)}")
logger.error(f"Failed to create FreqAI targets: {str(e)}")
raise
logger.info(f"目标列预览：\n{dataframe[['&-buy_rsi']].head().to_string()}")
logger.info(f"Target columns preview: {dataframe[['&-buy_rsi']].head().to_string()}")
return dataframe
def populate_indicators(self, dataframe: DataFrame, metadata: dict) -> DataFrame:
logger.info(f"处理交易对：{metadata['pair']}")
dataframe = self.freqai.start(dataframe, metadata, self)
# 计算传统指标
dataframe["rsi"] = ta.RSI(dataframe, timeperiod=14)
bollinger = qtpylib.bollinger_bands(qtpylib.typical_price(dataframe), window=20, stds=2)
dataframe["bb_lowerband"] = bollinger["lower"]
dataframe["bb_middleband"] = bollinger["mid"]
dataframe["bb_upperband"] = bollinger["upper"]
dataframe["tema"] = ta.TEMA(dataframe, timeperiod=9)
# 生成 up_or_down 信号
label_period = self.freqai_info["feature_parameters"]["label_period_candles"]
dataframe["up_or_down"] = np.where(
dataframe["close"].shift(-label_period) > dataframe["close"], 1, 0
)
# 预填充 NaN
dataframe = dataframe.ffill()
dataframe = dataframe.fillna(0)
if "&-buy_rsi" in dataframe.columns:
# 派生其他目标
dataframe["&-sell_rsi"] = dataframe["&-buy_rsi"] + 30
dataframe["%-volatility"] = dataframe["close"].pct_change().rolling(20).std()
dataframe["&-stoploss"] = -0.1 - (dataframe["%-volatility"] * 10).clip(0, 0.25)
dataframe["&-roi_0"] = (dataframe["close"].shift(-label_period) / dataframe["close"] - 1).clip(0, 0.2)
# 计算预测值并减少 NaN
dataframe["buy_rsi_pred"] = dataframe["&-buy_rsi"].rolling(5, min_periods=1).mean().clip(10, 50)
dataframe["sell_rsi_pred"] = dataframe["&-sell_rsi"].rolling(5, min_periods=1).mean().clip(50, 90)
dataframe["stoploss_pred"] = dataframe["&-stoploss"].clip(-0.25, -0.05)
dataframe["roi_0_pred"] = dataframe["&-roi_0"].clip(0.01, 0.2)
# 处理 NaN
for col in ["buy_rsi_pred", "sell_rsi_pred", "stoploss_pred", "roi_0_pred", "&-sell_rsi", "&-stoploss", "&-roi_0"]:
if dataframe[col].isna().any():
logger.warning(f"列 {col} 包含 NaN，填充为默认值")
mean_value = dataframe[col].mean()
if pd.isna(mean_value):
logger.warning(f"列 {col} 均值仍为 NaN，使用默认值")
mean_value = {
"buy_rsi_pred": 30,
"sell_rsi_pred": 70,
"stoploss_pred": -0.1,
"roi_0_pred": 0.05,
"&-sell_rsi": 70,
"&-stoploss": -0.1,
"&-roi_0": 0.05
}.get(col, 0)
dataframe[col] = dataframe[col].fillna(mean_value)
# 动态追踪止盈
dataframe["trailing_stop_positive"] = (dataframe["roi_0_pred"] * 0.5).clip(0.01, 0.3)
dataframe["trailing_stop_positive_offset"] = (dataframe["roi_0_pred"] * 0.75).clip(0.02, 0.4)
logger.info(f"Processing pair: {metadata['pair']}")
logger.info(f"DataFrame rows: {len(dataframe)}")
logger.info(f"Columns before freqai.start: {list(dataframe.columns)}")
# 设置动态参数
self.stoploss = float(dataframe["stoploss_pred"].iloc[-1])
self.buy_rsi.value = float(dataframe["buy_rsi_pred"].iloc[-1])
self.sell_rsi.value = float(dataframe["sell_rsi_pred"].iloc[-1])
self.minimal_roi = {
0: float(self.roi_0.value),
15: float(self.roi_15.value),
30: float(self.roi_30.value),
60: 0.0
}
self.trailing_stop_positive = float(dataframe["trailing_stop_positive"].iloc[-1])
self.trailing_stop_positive_offset = float(dataframe["trailing_stop_positive_offset"].iloc[-1])
logger.info(f"minimal_roi 键：{list(self.minimal_roi.keys())}")
logger.info(f"动态参数：buy_rsi={self.buy_rsi.value}, sell_rsi={self.sell_rsi.value}, "
f"stoploss={self.stoploss}, trailing_stop_positive={self.trailing_stop_positive}")
dataframe = dataframe.replace([np.inf, -np.inf], 0)
dataframe = dataframe.ffill()
dataframe = dataframe.fillna(0)
logger.info(f"up_or_down 值统计：\n{dataframe['up_or_down'].value_counts().to_string()}")
logger.info(f"do_predict 值统计：\n{dataframe['do_predict'].value_counts().to_string()}")
# 验证输入数据
if "close" not in dataframe.columns or dataframe["close"].isna().all():
logger.error(f"DataFrame missing 'close' column or all NaN for pair: {metadata['pair']}")
raise ValueError("DataFrame missing valid 'close' column")
# 生成 RSI
if len(dataframe) < 14:
logger.warning(f"DataFrame too short ({len(dataframe)} rows), cannot compute rsi")
dataframe["rsi"] = 50
else:
dataframe["rsi"] = ta.RSI(dataframe, timeperiod=14)
logger.info(f"rsi stats: {dataframe['rsi'].describe().to_string()}")
# 生成 %-volatility
if len(dataframe) < 20 or dataframe["close"].isna().any():
logger.warning(f"DataFrame too short ({len(dataframe)} rows) or contains NaN in close, cannot compute %-volatility")
dataframe["%-volatility"] = 0
else:
dataframe["%-volatility"] = dataframe["close"].pct_change().rolling(20).std()
dataframe["%-volatility"] = dataframe["%-volatility"].fillna(0)
if dataframe["%-volatility"].std() > 0:
dataframe["%-volatility"] = (dataframe["%-volatility"] - dataframe["%-volatility"].mean()) / dataframe["%-volatility"].std()
logger.info(f"%-volatility stats: {dataframe['%-volatility'].describe().to_string()}")
# 生成 TEMA
if len(dataframe) < 9:
logger.warning(f"DataFrame too short ({len(dataframe)} rows), cannot compute tema")
dataframe["tema"] = dataframe["close"]
else:
dataframe["tema"] = ta.TEMA(dataframe, timeperiod=9)
if dataframe["tema"].isna().any():
logger.warning("tema contains NaN, filling with close")
dataframe["tema"] = dataframe["tema"].fillna(dataframe["close"])
logger.info(f"tema stats: {dataframe['tema'].describe().to_string()}")
# 生成 Bollinger Bands
if len(dataframe) < 20:
logger.warning(f"DataFrame too short ({len(dataframe)} rows), cannot compute bb_lowerband")
dataframe["bb_lowerband"] = dataframe["close"]
else:
bollinger = qtpylib.bollinger_bands(qtpylib.typical_price(dataframe), window=20, stds=2.2)
dataframe["bb_lowerband"] = bollinger["lower"]
if dataframe["bb_lowerband"].isna().any():
logger.warning("bb_lowerband contains NaN, filling with close")
dataframe["bb_lowerband"] = dataframe["bb_lowerband"].fillna(dataframe["close"])
logger.info(f"bb_lowerband stats: {dataframe['bb_lowerband'].describe().to_string()}")
# 生成 up_or_down
label_period = self.freqai_info["feature_parameters"]["label_period_candles"]
if len(dataframe) < label_period + 1:
logger.warning(f"DataFrame too short ({len(dataframe)} rows), cannot compute up_or_down")
dataframe["up_or_down"] = 0
else:
dataframe["up_or_down"] = np.where(
dataframe["close"].shift(-label_period) > dataframe["close"], 1, 0
)
if dataframe["up_or_down"].isna().any():
logger.warning("up_or_down contains NaN, filling with 0")
dataframe["up_or_down"] = dataframe["up_or_down"].fillna(0)
logger.info(f"up_or_down stats: {dataframe['up_or_down'].describe().to_string()}")
# 生成其他特征
if "date" in dataframe.columns:
dataframe["%-day_of_week"] = dataframe["date"].dt.dayofweek
dataframe["%-hour_of_day"] = dataframe["date"].dt.hour
else:
logger.warning("Missing 'date' column, skipping %-day_of_week and %-hour_of_day")
dataframe["%-day_of_week"] = 0
dataframe["%-hour_of_day"] = 0
# 调用 FreqAI
try:
dataframe = self.freqai.start(dataframe, metadata, self)
logger.info(f"Columns after freqai.start: {list(dataframe.columns)}")
except Exception as e:
logger.error(f"freqai.start failed: {str(e)}")
dataframe["buy_rsi_pred"] = 50
dataframe["sell_rsi_pred"] = 80
dataframe["do_predict"] = 1
# 检查预测列
for col in ["buy_rsi_pred", "sell_rsi_pred"]:
if col not in dataframe.columns:
logger.error(f"Error: {col} column not generated for pair: {metadata['pair']}")
dataframe[col] = 50 if col == "buy_rsi_pred" else 80
logger.info(f"{col} stats: {dataframe[col].describe().to_string()}")
# 调试特征分布
if "%-bb_width-period_10_SOL/USDT_5m" in dataframe.columns:
if dataframe["%-bb_width-period_10_SOL/USDT_5m"].std() > 0:
dataframe["%-bb_width-period_10_SOL/USDT_5m"] = (
dataframe["%-bb_width-period_10_SOL/USDT_5m"] - dataframe["%-bb_width-period_10_SOL/USDT_5m"].mean()
) / dataframe["%-bb_width-period_10_SOL/USDT_5m"].std()
logger.info(f"%-bb_width-period_10 stats: {dataframe['%-bb_width-period_10_SOL/USDT_5m'].describe().to_string()}")
# 动态生成期望的特征列
def get_expected_columns(freqai_config: dict) -> list:
indicators = ["rsi", "bb_width", "pct-change"]
periods = freqai_config.get("feature_parameters", {}).get("include_periods", [10, 20])
pairs = freqai_config.get("include_corr_pairlist", ["SOL/USDT", "BTC/USDT"])
timeframes = freqai_config.get("include_timeframes", ["5m"])
shifts = [0]
expected_columns = ["%-volatility", "%-day_of_week", "%-hour_of_day"]
for indicator in indicators:
for period in periods:
for pair in pairs:
for timeframe in timeframes:
for shift in shifts:
col_name = f"%-{indicator}-period_{period}" if indicator != "pct-change" else f"%-{indicator}"
if shift > 0:
col_name += f"_shift-{shift}"
col_name += f"_{pair}_{timeframe}"
expected_columns.append(col_name)
return expected_columns
expected_columns = get_expected_columns(self.freqai_info)
logger.info(f"Expected feature columns ({len(expected_columns)}): {expected_columns[:10]}...")
# 比较特征集
actual_columns = list(dataframe.columns)
missing_columns = [col for col in expected_columns if col not in actual_columns]
extra_columns = [col for col in actual_columns if col not in expected_columns and col.startswith("%-")]
logger.info(f"Missing columns ({len(missing_columns)}): {missing_columns}")
logger.info(f"Extra columns ({len(extra_columns)}): {extra_columns}")
# 调试 DI 丢弃预测
if "DI_values" in dataframe.columns:
logger.info(f"DI_values stats: {dataframe['DI_values'].describe().to_string()}")
logger.info(f"DI discarded predictions: {len(dataframe[dataframe['do_predict'] == 0])}")
# 清理数据
dataframe = dataframe.replace([np.inf, -np.inf], 0).ffill().fillna(0)
logger.info(f"Final columns in populate_indicators: {list(dataframe.columns)}")
return dataframe
def populate_entry_trend(self, df: DataFrame, metadata: dict) -> DataFrame:
@ -272,7 +290,6 @@ class FreqaiExampleStrategy(IStrategy):
df["up_or_down"] == 0
]
time_exit = (df["date"] >= df["date"].shift(1) + pd.Timedelta(days=1))
df.loc[
(reduce(lambda x, y: x & y, exit_long_conditions)) | time_exit,
"exit_long"