diff --git a/docs/freqai-reinforcement-learning.md b/docs/freqai-reinforcement-learning.md
index 58eb638b8..7e4110b0b 100644
--- a/docs/freqai-reinforcement-learning.md
+++ b/docs/freqai-reinforcement-learning.md
@@ -177,7 +177,7 @@ As you begin to modify the strategy and the prediction model, you will quickly r
                 factor = 100
 
             # you can use feature values from dataframe
-            rsi_now = self.df[f"%-rsi-period-10_shift-1_{self.pair}_"
+            rsi_now = self.raw_features[f"%-rsi-period-10_shift-1_{self.pair}_"
                               f"{self.config['timeframe']}"].iloc[self._current_tick]
 
             # reward agent for entering trades
diff --git a/freqtrade/freqai/RL/BaseEnvironment.py b/freqtrade/freqai/RL/BaseEnvironment.py
index 3d9527d4f..7a4467bf7 100644
--- a/freqtrade/freqai/RL/BaseEnvironment.py
+++ b/freqtrade/freqai/RL/BaseEnvironment.py
@@ -45,7 +45,8 @@ class BaseEnvironment(gym.Env):
     def __init__(self, df: DataFrame = DataFrame(), prices: DataFrame = DataFrame(),
                  reward_kwargs: dict = {}, window_size=10, starting_point=True,
                  id: str = 'baseenv-1', seed: int = 1, config: dict = {}, live: bool = False,
-                 fee: float = 0.0015, can_short: bool = False, pair: str = ""):
+                 fee: float = 0.0015, can_short: bool = False, pair: str = "",
+                 df_raw: DataFrame = DataFrame()):
         """
         Initializes the training/eval environment.
         :param df: dataframe of features
@@ -67,6 +68,7 @@ class BaseEnvironment(gym.Env):
         self.max_drawdown: float = 1 - self.rl_config.get('max_training_drawdown_pct', 0.8)
         self.compound_trades: bool = config['stake_amount'] == 'unlimited'
         self.pair: str = pair
+        self.raw_features: DataFrame = df_raw
         if self.config.get('fee', None) is not None:
             self.fee = self.config['fee']
         else:
@@ -94,8 +96,7 @@ class BaseEnvironment(gym.Env):
         :param reward_kwargs: extra config settings assigned by user in `rl_config`
         :param starting_point: start at edge of window or not
         """
-        self.df: DataFrame = df
-        self.signal_features: DataFrame = self.df
+        self.signal_features: DataFrame = df
         self.prices: DataFrame = prices
         self.window_size: int = window_size
         self.starting_point: bool = starting_point
diff --git a/freqtrade/freqai/RL/BaseReinforcementLearningModel.py b/freqtrade/freqai/RL/BaseReinforcementLearningModel.py
index 8987c35fe..a8ef69394 100644
--- a/freqtrade/freqai/RL/BaseReinforcementLearningModel.py
+++ b/freqtrade/freqai/RL/BaseReinforcementLearningModel.py
@@ -1,3 +1,4 @@
+import copy
 import importlib
 import logging
 from abc import abstractmethod
@@ -50,6 +51,7 @@ class BaseReinforcementLearningModel(IFreqaiModel):
         self.eval_callback: Optional[EvalCallback] = None
         self.model_type = self.freqai_info['rl_config']['model_type']
         self.rl_config = self.freqai_info['rl_config']
+        self.df_raw: DataFrame = DataFrame()
         self.continual_learning = self.freqai_info.get('continual_learning', False)
         if self.model_type in SB3_MODELS:
             import_str = 'stable_baselines3'
@@ -107,6 +109,7 @@ class BaseReinforcementLearningModel(IFreqaiModel):
 
         data_dictionary: Dict[str, Any] = dk.make_train_test_datasets(
             features_filtered, labels_filtered)
+        self.df_raw = copy.deepcopy(data_dictionary["train_features"])
         dk.fit_labels()  # FIXME useless for now, but just satiating append methods
 
         # normalize all data based on train_dataset only
@@ -167,7 +170,8 @@ class BaseReinforcementLearningModel(IFreqaiModel):
                     "config": self.config,
                     "live": self.live,
                     "can_short": self.can_short,
-                    "pair": pair}
+                    "pair": pair,
+                    "df_raw": self.df_raw}
         if self.data_provider:
             env_info["fee"] = self.data_provider._exchange \
                 .get_fee(symbol=self.data_provider.current_whitelist()[0])  # type: ignore
@@ -365,8 +369,8 @@ class BaseReinforcementLearningModel(IFreqaiModel):
             factor = 100.
 
             # you can use feature values from dataframe
-            rsi_now = self.df[f"%-rsi-period-10_shift-1_{self.pair}_"
-                              f"{self.config['timeframe']}"].iloc[self._current_tick]
+            rsi_now = self.raw_features[f"%-rsi-period-10_shift-1_{self.pair}_"
+                                        f"{self.config['timeframe']}"].iloc[self._current_tick]
 
             # reward agent for entering trades
             if (action in (Actions.Long_enter.value, Actions.Short_enter.value)