Merge pull request #7296 from th0rntwig/dbscan

Improve MinPts calculation in DBSCAN, add outlier protection, and add data_kitchen tests

docs/freqai.md

@@ -113,6 +113,7 @@ Mandatory parameters are marked as **Required**, which means that they are requi
 | `use_SVM_to_remove_outliers` | Train a support vector machine to detect and remove outliers from the training data set, as well as from incoming data points. See details about how it works [here](#removing-outliers-using-a-support-vector-machine-svm). <br> **Datatype:** Boolean.
 | `svm_params` | All parameters available in Sklearn's `SGDOneClassSVM()`. See details about some select parameters [here](#removing-outliers-using-a-support-vector-machine-svm). <br> **Datatype:** Dictionary.
 | `use_DBSCAN_to_remove_outliers` | Cluster data using DBSCAN to identify and remove outliers from training and prediction data. See details about how it works [here](#removing-outliers-with-dbscan). <br> **Datatype:** Boolean. 
+| `outlier_protection_percentage` | If more than `outlier_protection_percentage` fraction of points are removed as outliers, FreqAI will log a warning message and ignore outlier detection while keeping the original dataset intact. <br> **Datatype:** float. Default: `30`
 |  |  **Data split parameters**
 | `data_split_parameters` | Include any additional parameters available from Scikit-learn `test_train_split()`, which are shown [here](https://scikit-learn.org/stable/modules/generated/sklearn.model_selection.train_test_split.html) (external website). <br> **Datatype:** Dictionary.
 | `test_size` | Fraction of data that should be used for testing instead of training. <br> **Datatype:** Positive float < 1.

freqtrade/freqai/data_drawer.py

@@ -566,7 +566,6 @@ class FreqaiDataDrawer:
         for training according to user defined train_period_days
         metadata: dict = strategy furnished pair metadata
         """
-
         with self.history_lock:
             corr_dataframes: Dict[Any, Any] = {}
             base_dataframes: Dict[Any, Any] = {}

freqtrade/freqai/data_kitchen.py

@@ -513,6 +513,19 @@ class FreqaiDataKitchen:
 
         return avg_mean_dist
 
+    def get_outlier_percentage(self, dropped_pts: npt.NDArray) -> float:
+        """
+        Check if more than X% of points werer dropped during outlier detection.
+        """
+        outlier_protection_pct = self.freqai_config["feature_parameters"].get(
+            "outlier_protection_percentage", 30)
+        outlier_pct = (dropped_pts.sum() / len(dropped_pts)) * 100
+        if outlier_pct >= outlier_protection_pct:
+            self.svm_model = None
+            return outlier_pct
+        else:
+            return 0.0
+
     def use_SVM_to_remove_outliers(self, predict: bool) -> None:
         """
         Build/inference a Support Vector Machine to detect outliers
@@ -550,8 +563,16 @@ class FreqaiDataKitchen:
                 self.data_dictionary["train_features"]
             )
             y_pred = self.svm_model.predict(self.data_dictionary["train_features"])
-            dropped_points = np.where(y_pred == -1, 0, y_pred)
+            kept_points = np.where(y_pred == -1, 0, y_pred)
             # keep_index = np.where(y_pred == 1)
+            outlier_pct = self.get_outlier_percentage(1 - kept_points)
+            if outlier_pct:
+                logger.warning(
+                        f"SVM detected {outlier_pct:.2f}% of the points as outliers. "
+                        f"Keeping original dataset."
+                )
+                return
+
             self.data_dictionary["train_features"] = self.data_dictionary["train_features"][
                 (y_pred == 1)
             ]
@@ -563,7 +584,7 @@ class FreqaiDataKitchen:
             ]
 
             logger.info(
-                f"SVM tossed {len(y_pred) - dropped_points.sum()}"
+                f"SVM tossed {len(y_pred) - kept_points.sum()}"
                 f" train points from {len(y_pred)} total points."
             )
 
@@ -572,7 +593,7 @@ class FreqaiDataKitchen:
             # to reduce code duplication
             if self.freqai_config['data_split_parameters'].get('test_size', 0.1) != 0:
                 y_pred = self.svm_model.predict(self.data_dictionary["test_features"])
-                dropped_points = np.where(y_pred == -1, 0, y_pred)
+                kept_points = np.where(y_pred == -1, 0, y_pred)
                 self.data_dictionary["test_features"] = self.data_dictionary["test_features"][
                     (y_pred == 1)
                 ]
@@ -583,7 +604,7 @@ class FreqaiDataKitchen:
                 ]
 
             logger.info(
-                f"SVM tossed {len(y_pred) - dropped_points.sum()}"
+                f"SVM tossed {len(y_pred) - kept_points.sum()}"
                 f" test points from {len(y_pred)} total points."
             )
 
@@ -635,8 +656,8 @@ class FreqaiDataKitchen:
                     cos(angle) * (point[1] - origin[1])
                 return (x, y)
 
-            MinPts = len(self.data_dictionary['train_features'].columns) * 2
-            # measure pairwise distances to train_features.shape[1]*2 nearest neighbours
+            MinPts = int(len(self.data_dictionary['train_features'].index) * 0.25)
+            # measure pairwise distances to nearest neighbours
             neighbors = NearestNeighbors(
                 n_neighbors=MinPts, n_jobs=self.thread_count)
             neighbors_fit = neighbors.fit(self.data_dictionary['train_features'])
@@ -667,6 +688,14 @@ class FreqaiDataKitchen:
             self.data['DBSCAN_min_samples'] = MinPts
             dropped_points = np.where(clustering.labels_ == -1, 1, 0)
 
+            outlier_pct = self.get_outlier_percentage(dropped_points)
+            if outlier_pct:
+                logger.warning(
+                        f"DBSCAN detected {outlier_pct:.2f}% of the points as outliers. "
+                        f"Keeping original dataset."
+                )
+                return
+
             self.data_dictionary['train_features'] = self.data_dictionary['train_features'][
                 (clustering.labels_ != -1)
             ]
@@ -722,6 +751,14 @@ class FreqaiDataKitchen:
             0,
         )
 
+        outlier_pct = self.get_outlier_percentage(1 - do_predict)
+        if outlier_pct:
+            logger.warning(
+                    f"DI detected {outlier_pct:.2f}% of the points as outliers. "
+                    f"Keeping original dataset."
+            )
+            return
+
         if (len(do_predict) - do_predict.sum()) > 0:
             logger.info(
                 f"DI tossed {len(do_predict) - do_predict.sum()} predictions for "

tests/freqai/conftest.py

@@ -1,5 +1,6 @@
 from copy import deepcopy
 from pathlib import Path
+from unittest.mock import MagicMock
 
 import pytest
 
@@ -81,6 +82,51 @@ def get_patched_freqaimodel(mocker, freqaiconf):
     return freqaimodel
 
 
+def make_data_dictionary(mocker, freqai_conf):
+    freqai_conf.update({"timerange": "20180110-20180130"})
+
+    strategy = get_patched_freqai_strategy(mocker, freqai_conf)
+    exchange = get_patched_exchange(mocker, freqai_conf)
+    strategy.dp = DataProvider(freqai_conf, exchange)
+    strategy.freqai_info = freqai_conf.get("freqai", {})
+    freqai = strategy.freqai
+    freqai.live = True
+    freqai.dk = FreqaiDataKitchen(freqai_conf)
+    freqai.dk.pair = "ADA/BTC"
+    timerange = TimeRange.parse_timerange("20180110-20180130")
+    freqai.dd.load_all_pair_histories(timerange, freqai.dk)
+
+    freqai.dd.pair_dict = MagicMock()
+
+    data_load_timerange = TimeRange.parse_timerange("20180110-20180130")
+    new_timerange = TimeRange.parse_timerange("20180120-20180130")
+
+    corr_dataframes, base_dataframes = freqai.dd.get_base_and_corr_dataframes(
+            data_load_timerange, freqai.dk.pair, freqai.dk
+        )
+
+    unfiltered_dataframe = freqai.dk.use_strategy_to_populate_indicators(
+                strategy, corr_dataframes, base_dataframes, freqai.dk.pair
+            )
+
+    unfiltered_dataframe = freqai.dk.slice_dataframe(new_timerange, unfiltered_dataframe)
+
+    freqai.dk.find_features(unfiltered_dataframe)
+
+    features_filtered, labels_filtered = freqai.dk.filter_features(
+            unfiltered_dataframe,
+            freqai.dk.training_features_list,
+            freqai.dk.label_list,
+            training_filter=True,
+        )
+
+    data_dictionary = freqai.dk.make_train_test_datasets(features_filtered, labels_filtered)
+
+    data_dictionary = freqai.dk.normalize_data(data_dictionary)
+
+    return freqai
+
+
 def get_freqai_live_analyzed_dataframe(mocker, freqaiconf):
     strategy = get_patched_freqai_strategy(mocker, freqaiconf)
     exchange = get_patched_exchange(mocker, freqaiconf)

tests/freqai/test_freqai_datakitchen.py

@@ -5,7 +5,8 @@ from pathlib import Path
 import pytest
 
 from freqtrade.exceptions import OperationalException
-from tests.freqai.conftest import get_patched_data_kitchen
+from tests.conftest import log_has_re
+from tests.freqai.conftest import get_patched_data_kitchen, make_data_dictionary
 
 
 @pytest.mark.parametrize(
@@ -66,3 +67,30 @@ def test_check_if_model_expired(mocker, freqai_conf, timestamp, expected):
     dk = get_patched_data_kitchen(mocker, freqai_conf)
     assert dk.check_if_model_expired(timestamp) == expected
     shutil.rmtree(Path(dk.full_path))
+
+
+def test_use_DBSCAN_to_remove_outliers(mocker, freqai_conf, caplog):
+    freqai = make_data_dictionary(mocker, freqai_conf)
+    # freqai_conf['freqai']['feature_parameters'].update({"outlier_protection_percentage": 1})
+    freqai.dk.use_DBSCAN_to_remove_outliers(predict=False)
+    assert log_has_re(
+        "DBSCAN found eps of 2.42.",
+        caplog,
+    )
+
+
+def test_compute_distances(mocker, freqai_conf):
+    freqai = make_data_dictionary(mocker, freqai_conf)
+    freqai_conf['freqai']['feature_parameters'].update({"DI_threshold": 1})
+    avg_mean_dist = freqai.dk.compute_distances()
+    assert round(avg_mean_dist, 2) == 2.56
+
+
+def test_use_SVM_to_remove_outliers_and_outlier_protection(mocker, freqai_conf, caplog):
+    freqai = make_data_dictionary(mocker, freqai_conf)
+    freqai_conf['freqai']['feature_parameters'].update({"outlier_protection_percentage": 0.1})
+    freqai.dk.use_SVM_to_remove_outliers(predict=False)
+    assert log_has_re(
+        "SVM detected 8.46%",
+        caplog,
+    )