# source: https://raw.githubusercontent.com/stefanoviana/deepalpha/b31d0a1b321262dcb2576ca8b706ebbd90808085/freqai-plugin/example_strategy.py """ DeepAlpha Example Strategy for Freqtrade ========================================= A minimal Freqtrade strategy demonstrating how to use the DeepAlpha FreqAI model. This strategy: - Defines technical indicator features for FreqAI - Uses Triple Barrier labels generated by the DeepAlpha model - Filters entries using meta-labeling confidence scores - Implements basic risk management via stop-loss and ROI Usage: freqtrade trade --strategy Github_stefanoviana_deepalpha__example_strategy__20260512_162149 --freqaimodel DeepAlphaModel Author: DeepAlpha Team License: MIT """ from __future__ import annotations import logging from typing import Dict import numpy as np import pandas as pd import talib.abstract as ta from pandas import DataFrame from freqtrade.strategy import IStrategy, merge_informative_pair from freqtrade.strategy.parameters import IntParameter, DecimalParameter logger = logging.getLogger(__name__) class Github_stefanoviana_deepalpha__example_strategy__20260512_162149(IStrategy): """ Freqtrade strategy powered by DeepAlpha's ML pipeline. The strategy delegates all signal generation to the FreqAI model. Features are defined in `feature_engineering_*` methods and passed to the DeepAlpha model, which returns predictions and a confidence mask from meta-labeling. """ # --------------------------------------------------------------- # Strategy configuration # --------------------------------------------------------------- INTERFACE_VERSION = 3 # Minimal ROI table tuned for 1h bars — less aggressive take-profit # so winners have room to compound over multiple bars. minimal_roi = { "0": 0.06, "240": 0.04, "480": 0.025, "1440": 0.015, } # 1h timeframe: reduces intraday noise that dominated 5m tests and # lets Triple Barrier capture real directional moves. stoploss = -0.04 trailing_stop = False trailing_stop_positive = 0.015 trailing_stop_positive_offset = 0.03 timeframe = "1h" # FreqAI configuration freqai_info = { "enabled": True, } # Strategy parameters (optimizable). # SHORT threshold higher than LONG (0.60 vs 0.55) to compensate for # the residual downside skew of Triple Barrier on volatile intraday data. buy_prediction_threshold = DecimalParameter( 0.5, 0.8, default=0.55, decimals=2, space="buy", optimize=True, load=True, ) sell_prediction_threshold = DecimalParameter( 0.5, 0.8, default=0.60, decimals=2, space="sell", optimize=True, load=True, ) meta_confidence_threshold = DecimalParameter( 0.5, 0.8, default=0.55, decimals=2, space="buy", optimize=True, load=True, ) # Confidence level that unlocks counter-trend entries (e.g. take a # short while EMA24 > EMA96 only if P_short exceeds this). counter_trend_override = DecimalParameter( 0.6, 0.9, default=0.70, decimals=2, space="buy", optimize=True, load=True, ) # --------------------------------------------------------------- # Feature engineering # --------------------------------------------------------------- def feature_engineering_expand_all( self, dataframe: DataFrame, period: int, metadata: dict, **kwargs ) -> DataFrame: """ Create features that will be expanded across multiple periods. FreqAI automatically generates these features for each period defined in the config (e.g., 10, 20, 50 candles). """ # Momentum indicators dataframe[f"%-rsi-period_{period}"] = ta.RSI(dataframe, timeperiod=period) dataframe[f"%-mfi-period_{period}"] = ta.MFI(dataframe, timeperiod=period) dataframe[f"%-adx-period_{period}"] = ta.ADX(dataframe, timeperiod=period) dataframe[f"%-cci-period_{period}"] = ta.CCI(dataframe, timeperiod=period) # Volatility bollinger = ta.BBANDS(dataframe, timeperiod=period, nbdevup=2.0, nbdevdn=2.0) dataframe[f"%-bb-width-period_{period}"] = ( (bollinger["upperband"] - bollinger["lowerband"]) / bollinger["middleband"] ) dataframe[f"%-bb-pct-period_{period}"] = ( (dataframe["close"] - bollinger["lowerband"]) / (bollinger["upperband"] - bollinger["lowerband"] + 1e-10) ) dataframe[f"%-atr-period_{period}"] = ta.ATR(dataframe, timeperiod=period) # Trend dataframe[f"%-sma-period_{period}"] = ta.SMA(dataframe, timeperiod=period) dataframe[f"%-ema-period_{period}"] = ta.EMA(dataframe, timeperiod=period) dataframe[f"%-close-sma-ratio_{period}"] = ( dataframe["close"] / (dataframe[f"%-sma-period_{period}"] + 1e-10) ) # Volume dataframe[f"%-volume-sma-ratio_{period}"] = ( dataframe["volume"] / (dataframe["volume"].rolling(period).mean() + 1e-10) ) return dataframe def feature_engineering_expand_basic( self, dataframe: DataFrame, metadata: dict, **kwargs ) -> DataFrame: """ Create features that are not expanded across multiple periods. """ # MACD macd = ta.MACD(dataframe, fastperiod=12, slowperiod=26, signalperiod=9) dataframe["%-macd"] = macd["macd"] dataframe["%-macd-signal"] = macd["macdsignal"] dataframe["%-macd-hist"] = macd["macdhist"] # Stochastic stoch = ta.STOCH(dataframe, fastk_period=14, slowk_period=3, slowd_period=3) dataframe["%-slowk"] = stoch["slowk"] dataframe["%-slowd"] = stoch["slowd"] # Ichimoku dataframe["%-tenkan"] = ( dataframe["high"].rolling(9).max() + dataframe["low"].rolling(9).min() ) / 2 dataframe["%-kijun"] = ( dataframe["high"].rolling(26).max() + dataframe["low"].rolling(26).min() ) / 2 # Price-based features dataframe["%-pct-change"] = dataframe["close"].pct_change() dataframe["%-log-return"] = np.log( dataframe["close"] / dataframe["close"].shift(1) ) # Candle patterns dataframe["%-body-ratio"] = ( abs(dataframe["close"] - dataframe["open"]) / (dataframe["high"] - dataframe["low"] + 1e-10) ) dataframe["%-upper-shadow"] = ( dataframe["high"] - dataframe[["close", "open"]].max(axis=1) ) / (dataframe["high"] - dataframe["low"] + 1e-10) # --- Sequential / memory features (poor-man's LSTM) -------------- # Rolling log-return statistics over multiple horizons give the # tree model temporal context without needing a recurrent network. log_ret = dataframe["%-log-return"] for win in (4, 12, 24, 48): dataframe[f"%-logret-mean-{win}"] = log_ret.rolling(win).mean() dataframe[f"%-logret-std-{win}"] = log_ret.rolling(win).std() dataframe[f"%-logret-sum-{win}"] = log_ret.rolling(win).sum() dataframe[f"%-logret-skew-{win}"] = log_ret.rolling(win).skew() # Return z-score: how extreme is the current bar's move vs recent history roll_std_24 = log_ret.rolling(24).std() dataframe["%-logret-zscore-24"] = log_ret / (roll_std_24 + 1e-10) # Volatility-of-volatility: regime-change signal dataframe["%-vol-of-vol-48"] = ( roll_std_24.rolling(48).std() / (roll_std_24.rolling(48).mean() + 1e-10) ) # Momentum-of-momentum: acceleration features for lag in (1, 3, 6, 12): dataframe[f"%-close-lag-{lag}"] = ( dataframe["close"] / dataframe["close"].shift(lag) - 1 ) dataframe[f"%-vol-lag-{lag}"] = ( dataframe["volume"] / (dataframe["volume"].shift(lag) + 1e-10) ) # High-Low range momentum hl_range = (dataframe["high"] - dataframe["low"]) / dataframe["close"] dataframe["%-hl-range-ratio-24"] = hl_range / (hl_range.rolling(24).mean() + 1e-10) return dataframe def feature_engineering_standard( self, dataframe: DataFrame, metadata: dict, **kwargs ) -> DataFrame: """ Features computed after all expansion. Also defines the target label column that the model will train on. """ # Day-of-week and hour features dataframe["%-day-of-week"] = dataframe["date"].dt.dayofweek dataframe["%-hour"] = dataframe["date"].dt.hour # NOTE: Do NOT define a "&-label" column here. Target generation is # handled exclusively in set_freqai_targets() with Triple Barrier # labels. If both "&-label" and "&-target" exist, FreqAI's # BaseClassifierModel.fit uses the first one alphabetically # ("&-label"), which would bypass our Triple Barrier labels. return dataframe def set_freqai_targets(self, dataframe: DataFrame, metadata: Dict, **kwargs) -> DataFrame: """ Generate regime-aware Triple Barrier labels. Class mapping (3-class, LightGBM-friendly non-negative ints): 0 = SHORT (-1), 1 = FLAT (0), 2 = LONG (+1) The previous symmetric (1.5 / 1.5 ATR) version produced too many SHORT labels in bull markets because intraday pullbacks hit the lower barrier first even when the overall drift was positive. This version detects the local regime via EMA24 vs EMA96 on the training timeframe and adapts the barrier multipliers: * BULL (EMA24 > EMA96 * 1.002): pt=1.2, sl=2.0 (favour LONG) * BEAR (EMA24 < EMA96 * 0.998): pt=2.0, sl=1.2 (favour SHORT) * SIDEWAYS: pt=1.5, sl=1.5 (symmetric) Horizon raised 12 -> 24 candles (2h on 5m) to let directional moves play out before the time-barrier fallback triggers. """ import numpy as np closes = dataframe["close"].values highs = dataframe["high"].values lows = dataframe["low"].values # Regime detection via EMA slope close_s = pd.Series(closes) ema_fast = close_s.ewm(span=24, adjust=False).mean().values ema_slow = close_s.ewm(span=96, adjust=False).mean().values regime = np.zeros(len(closes), dtype=int) regime[ema_fast > ema_slow * 1.002] = 1 # BULL regime[ema_fast < ema_slow * 0.998] = -1 # BEAR # ATR for adaptive barrier sizing atr_period = 14 tr = np.maximum.reduce([ highs - lows, np.abs(highs - np.roll(closes, 1)), np.abs(lows - np.roll(closes, 1)) ]) atr = np.zeros(len(closes)) if len(tr) >= atr_period: atr[atr_period:] = np.convolve( tr, np.ones(atr_period) / atr_period, mode='valid' )[:len(closes) - atr_period] horizon = 12 # 12 hours on a 1h timeframe labels = np.zeros(len(closes), dtype=int) for i in range(len(closes) - horizon): if atr[i] <= 0: continue entry = closes[i] atr_pct = atr[i] / entry if entry > 0 else 0 if atr_pct == 0: continue # Regime-aware asymmetric barriers if regime[i] == 1: pt_mult, sl_mult = 1.2, 2.0 elif regime[i] == -1: pt_mult, sl_mult = 2.0, 1.2 else: pt_mult, sl_mult = 1.5, 1.5 pt_long = entry * (1 + pt_mult * atr_pct) sl_long = entry * (1 - sl_mult * atr_pct) end = min(i + horizon, len(closes) - 1) hit = 0 for j in range(i + 1, end + 1): if highs[j] >= pt_long: hit = 1 # LONG wins break if lows[j] <= sl_long: hit = -1 # SHORT wins break if hit == 0: # Time barrier: use sign of return vs regime-aware dead-zone ret = (closes[end] - entry) / entry if ret > atr_pct * 0.3: hit = 1 elif ret < -atr_pct * 0.3: hit = -1 labels[i] = hit target_3class = (labels + 1).astype(int) dataframe["&-target"] = target_3class return dataframe # --------------------------------------------------------------- # Signal generation # --------------------------------------------------------------- def populate_indicators(self, dataframe: DataFrame, metadata: dict) -> DataFrame: """ FreqAI handles indicator generation via feature_engineering methods. This method is intentionally minimal. """ # FreqAI populates predictions automatically dataframe = self.freqai.start(dataframe, metadata, self) return dataframe def populate_entry_trend(self, dataframe: DataFrame, metadata: dict) -> DataFrame: """ Generate entry signals from DeepAlpha 3-class model predictions, gated by a same-timeframe EMA regime filter. Class mapping (LightGBM multi-class, non-negative ints): "0" -> P(SHORT) "1" -> P(FLAT) "2" -> P(LONG) Entry rules: * LONG if P_long > long_thr AND P_long > P_short AND (uptrend OR P_long > counter_trend_override) * SHORT if P_short > short_thr AND P_short > P_long AND (downtrend OR P_short > counter_trend_override) """ n = len(dataframe) p_short = dataframe["0"] if "0" in dataframe.columns else pd.Series([0.0] * n, index=dataframe.index) p_long = dataframe["2"] if "2" in dataframe.columns else pd.Series([0.0] * n, index=dataframe.index) do_predict = ( dataframe["do_predict"] if "do_predict" in dataframe.columns else pd.Series([1] * n, index=dataframe.index) ) # Regime filter on the trading timeframe (EMA24 vs EMA96 on 5m) ema_fast = dataframe["close"].ewm(span=24, adjust=False).mean() ema_slow = dataframe["close"].ewm(span=96, adjust=False).mean() uptrend = ema_fast > ema_slow downtrend = ema_fast < ema_slow long_thr = self.buy_prediction_threshold.value short_thr = self.sell_prediction_threshold.value override = self.counter_trend_override.value # Hard regime gate: never counter-trend. Counter-trend shorts in # bull markets produced a 3% stop-loss carnage in our Q1 2024 # tests (112 stop-losses = -36.6% total). The override was too # permissive even at 0.70 confidence. Force alignment with # same-timeframe EMA regime. long_entry = ( (p_long > long_thr) & (p_long > p_short) & (do_predict == 1) & uptrend ) short_entry = ( (p_short > short_thr) & (p_short > p_long) & (do_predict == 1) & downtrend ) _ = override # kept for hyperopt compatibility; currently unused dataframe.loc[long_entry, "enter_long"] = 1 dataframe.loc[long_entry, "enter_tag"] = "DeepAlpha-LONG" dataframe.loc[short_entry, "enter_short"] = 1 dataframe.loc[short_entry, "enter_tag"] = "DeepAlpha-SHORT" return dataframe # --------------------------------------------------------------- # Dynamic position sizing based on prediction confidence # --------------------------------------------------------------- def custom_stake_amount( self, pair: str, current_time, current_rate: float, proposed_stake: float, min_stake, max_stake: float, leverage: float, entry_tag, side: str, **kwargs, ) -> float: """ Scale the stake linearly with the winning-class probability. A P_max of 0.55 (just above the gate) uses 50% of the default stake; a P_max of 0.80+ uses 150%. This concentrates risk on high-confidence signals and reduces bleed on marginal calls. """ try: df, _ = self.dp.get_analyzed_dataframe(pair=pair, timeframe=self.timeframe) if df is None or df.empty: return proposed_stake last = df.iloc[-1] p_short = float(last.get("0", 0.0) or 0.0) p_long = float(last.get("2", 0.0) or 0.0) p_max = p_long if side == "long" else p_short # Linear scaling: 0.55 -> 0.5x, 0.80 -> 1.5x (clipped) scale = 0.5 + (p_max - 0.55) * 4.0 scale = max(0.25, min(scale, 1.75)) stake = proposed_stake * scale if min_stake is not None: stake = max(stake, float(min_stake)) stake = min(stake, float(max_stake)) return stake except Exception as e: logger.warning("custom_stake_amount fallback (%s): %s", pair, e) return proposed_stake def populate_exit_trend(self, dataframe: DataFrame, metadata: dict) -> DataFrame: """ 3-class exits: close longs when P_short dominates, close shorts when P_long dominates. ROI / stoploss / trailing stop remain the primary risk-management layer. """ n = len(dataframe) p_short = dataframe["0"] if "0" in dataframe.columns else pd.Series([0.0] * n, index=dataframe.index) p_long = dataframe["2"] if "2" in dataframe.columns else pd.Series([0.0] * n, index=dataframe.index) dataframe.loc[(p_short > 0.50) & (p_short > p_long), "exit_long"] = 1 dataframe.loc[(p_long > 0.50) & (p_long > p_short), "exit_short"] = 1 return dataframe