Merge pull request #9179 from freqtrade/chore/better-freqai-reload

chore: fix bug associated with leaving FreqAI offline for more than 1candle

freqtrade/freqai/data_drawer.py

@@ -263,23 +263,55 @@ class FreqaiDataDrawer:
             self.pair_dict[metadata["pair"]] = self.empty_pair_dict.copy()
             return
 
-    def set_initial_return_values(self, pair: str, pred_df: DataFrame) -> None:
+    def set_initial_return_values(self, pair: str,
+                                  pred_df: DataFrame,
+                                  dataframe: DataFrame
+                                  ) -> None:
         """
         Set the initial return values to the historical predictions dataframe. This avoids needing
         to repredict on historical candles, and also stores historical predictions despite
         retrainings (so stored predictions are true predictions, not just inferencing on trained
-        data)
+        data).
+
+        We also aim to keep the date from historical predictions so that the FreqUI displays
+        zeros during any downtime (between FreqAI reloads).
         """
 
-        hist_df = self.historic_predictions
-        len_diff = len(hist_df[pair].index) - len(pred_df.index)
-        if len_diff < 0:
-            df_concat = pd.concat([pred_df.iloc[:abs(len_diff)], hist_df[pair]],
-                                  ignore_index=True, keys=hist_df[pair].keys())
+        new_pred = pred_df.copy()
+        # set new_pred values to nans (we want to signal to user that there was nothing
+        # historically made during downtime. The newest pred will get appeneded later in
+        # append_model_predictions)
+        new_pred.iloc[:, :] = np.nan
+        new_pred["date"] = dataframe["date"]
+
+        hist_preds = self.historic_predictions[pair].copy()
+        # rename date_pred column to date so that we can merge on date
+        hist_preds = hist_preds.rename(columns={"date_pred": "date"})
+
+        # find the closest common date between new_pred and historic predictions
+        # and cut off the new_pred dataframe at that date
+        common_dates = pd.merge(new_pred, hist_preds, on="date", how="inner")
+        if len(common_dates.index) > 0:
+            new_pred = new_pred.iloc[len(common_dates):]
         else:
-            df_concat = hist_df[pair].tail(len(pred_df.index)).reset_index(drop=True)
+            logger.warning("No common dates found between new predictions and historic "
+                           "predictions. You likely left your FreqAI instance offline "
+                           f"for more than {len(dataframe.index)} candles.")
+
+        df_concat = pd.concat([hist_preds, new_pred], ignore_index=True, keys=hist_preds.keys())
+
+        # remove last row because we will append that later in append_model_predictions()
+        df_concat = df_concat.iloc[:-1]
+        # any missing values will get zeroed out so users can see the exact
+        # downtime in FreqUI
         df_concat = df_concat.fillna(0)
-        self.model_return_values[pair] = df_concat
+
+        # rename date column back to date_pred
+        df_concat = df_concat.rename(columns={"date": "date_pred"})
+
+        self.historic_predictions[pair] = df_concat
+
+        self.model_return_values[pair] = df_concat.tail(len(dataframe.index)).reset_index(drop=True)
 
     def append_model_predictions(self, pair: str, predictions: DataFrame,
                                  do_preds: NDArray[np.int_],

freqtrade/freqai/freqai_interface.py

@@ -453,7 +453,7 @@ class IFreqaiModel(ABC):
             pred_df, do_preds = self.predict(dataframe, dk)
             if pair not in self.dd.historic_predictions:
                 self.set_initial_historic_predictions(pred_df, dk, pair, dataframe)
-            self.dd.set_initial_return_values(pair, pred_df)
+            self.dd.set_initial_return_values(pair, pred_df, dataframe)
 
             dk.return_dataframe = self.dd.attach_return_values_to_return_dataframe(pair, dataframe)
             return
@@ -645,11 +645,11 @@ class IFreqaiModel(ABC):
         If the user reuses an identifier on a subsequent instance,
         this function will not be called. In that case, "real" predictions
         will be appended to the loaded set of historic predictions.
-        :param df: DataFrame = the dataframe containing the training feature data
-        :param model: Any = A model which was `fit` using a common library such as
-                      catboost or lightgbm
+        :param pred_df: DataFrame = the dataframe containing the predictions coming
+            out of a model
         :param dk: FreqaiDataKitchen = object containing methods for data analysis
         :param pair: str = current pair
+        :param strat_df: DataFrame = dataframe coming from strategy
         """
 
         self.dd.historic_predictions[pair] = pred_df

freqtrade/templates/FreqaiExampleStrategy.py

@@ -31,7 +31,7 @@ class FreqaiExampleStrategy(IStrategy):
     plot_config = {
         "main_plot": {},
         "subplots": {
-            "&-s_close": {"prediction": {"color": "blue"}},
+            "&-s_close": {"&-s_close": {"color": "blue"}},
             "do_predict": {
                 "do_predict": {"color": "brown"},
             },

tests/freqai/test_freqai_datadrawer.py

@@ -1,7 +1,9 @@
 
 import shutil
 from pathlib import Path
+from unittest.mock import patch
 
+import pandas as pd
 import pytest
 
 from freqtrade.configuration import TimeRange
@@ -135,3 +137,113 @@ def test_get_timerange_from_backtesting_live_df_pred_not_found(mocker, freqai_co
             match=r'Historic predictions not found.*'
             ):
         freqai.dd.get_timerange_from_live_historic_predictions()
+
+
+def test_set_initial_return_values(mocker, freqai_conf):
+    """
+    Simple test of the set initial return values that ensures
+    we are concatening and ffilling values properly.
+    """
+
+    strategy = get_patched_freqai_strategy(mocker, freqai_conf)
+    exchange = get_patched_exchange(mocker, freqai_conf)
+    strategy.dp = DataProvider(freqai_conf, exchange)
+    freqai = strategy.freqai
+    freqai.live = False
+    freqai.dk = FreqaiDataKitchen(freqai_conf)
+    # Setup
+    pair = "BTC/USD"
+    end_x = "2023-08-31"
+    start_x_plus_1 = "2023-08-30"
+    end_x_plus_5 = "2023-09-03"
+
+    historic_data = {
+        'date_pred': pd.date_range(end=end_x, periods=5),
+        'value': range(1, 6)
+    }
+    new_data = {
+        'date': pd.date_range(start=start_x_plus_1, end=end_x_plus_5),
+        'value': range(6, 11)
+    }
+
+    freqai.dd.historic_predictions[pair] = pd.DataFrame(historic_data)
+
+    new_pred_df = pd.DataFrame(new_data)
+    dataframe = pd.DataFrame(new_data)
+
+    # Action
+    with patch('logging.Logger.warning') as mock_logger_warning:
+        freqai.dd.set_initial_return_values(pair, new_pred_df, dataframe)
+
+    # Assertions
+    hist_pred_df = freqai.dd.historic_predictions[pair]
+    model_return_df = freqai.dd.model_return_values[pair]
+
+    assert (hist_pred_df['date_pred'].iloc[-1] ==
+            pd.Timestamp(end_x_plus_5) - pd.Timedelta(days=1))
+    assert 'date' not in hist_pred_df.columns
+    assert 'date_pred' in hist_pred_df.columns
+    assert hist_pred_df.shape[0] == 7  # Total rows: 5 from historic and 2 new zeros
+
+    # compare values in model_return_df with hist_pred_df
+    assert (model_return_df["value"].values ==
+            hist_pred_df.tail(len(dataframe))["value"].values).all()
+    assert model_return_df.shape[0] == len(dataframe)
+
+    # Ensure logger error is not called
+    mock_logger_warning.assert_not_called()
+
+
+def test_set_initial_return_values_warning(mocker, freqai_conf):
+    """
+    Simple test of set_initial_return_values that hits the warning
+    associated with leaving a FreqAI bot offline so long that the
+    exchange candles have no common date with the historic predictions
+    """
+
+    strategy = get_patched_freqai_strategy(mocker, freqai_conf)
+    exchange = get_patched_exchange(mocker, freqai_conf)
+    strategy.dp = DataProvider(freqai_conf, exchange)
+    freqai = strategy.freqai
+    freqai.live = False
+    freqai.dk = FreqaiDataKitchen(freqai_conf)
+    # Setup
+    pair = "BTC/USD"
+    end_x = "2023-08-31"
+    start_x_plus_1 = "2023-09-01"
+    end_x_plus_5 = "2023-09-05"
+
+    historic_data = {
+        'date_pred': pd.date_range(end=end_x, periods=5),
+        'value': range(1, 6)
+    }
+    new_data = {
+        'date': pd.date_range(start=start_x_plus_1, end=end_x_plus_5),
+        'value': range(6, 11)
+    }
+
+    freqai.dd.historic_predictions[pair] = pd.DataFrame(historic_data)
+
+    new_pred_df = pd.DataFrame(new_data)
+    dataframe = pd.DataFrame(new_data)
+
+    # Action
+    with patch('logging.Logger.warning') as mock_logger_warning:
+        freqai.dd.set_initial_return_values(pair, new_pred_df, dataframe)
+
+    # Assertions
+    hist_pred_df = freqai.dd.historic_predictions[pair]
+    model_return_df = freqai.dd.model_return_values[pair]
+
+    assert hist_pred_df['date_pred'].iloc[-1] == pd.Timestamp(end_x_plus_5) - pd.Timedelta(days=1)
+    assert 'date' not in hist_pred_df.columns
+    assert 'date_pred' in hist_pred_df.columns
+    assert hist_pred_df.shape[0] == 9  # Total rows: 5 from historic and 4 new zeros
+
+    # compare values in model_return_df with hist_pred_df
+    assert (model_return_df["value"].values == hist_pred_df.tail(
+        len(dataframe))["value"].values).all()
+    assert model_return_df.shape[0] == len(dataframe)
+
+    # Ensure logger error is not called
+    mock_logger_warning.assert_called()