import copy import logging import shutil from datetime import datetime, timezone from math import cos, sin from pathlib import Path from typing import Any, Dict, List, Tuple import numpy as np import numpy.typing as npt import pandas as pd from pandas import DataFrame from scipy import stats from sklearn import linear_model from sklearn.cluster import DBSCAN from sklearn.metrics.pairwise import pairwise_distances from sklearn.model_selection import train_test_split from sklearn.neighbors import NearestNeighbors from freqtrade.configuration import TimeRange from freqtrade.constants import Config from freqtrade.exceptions import OperationalException from freqtrade.exchange import timeframe_to_seconds from freqtrade.strategy.interface import IStrategy SECONDS_IN_DAY = 86400 SECONDS_IN_HOUR = 3600 logger = logging.getLogger(__name__) class FreqaiDataKitchen: """ Class designed to analyze data for a single pair. Employed by the IFreqaiModel class. Functionalities include holding, saving, loading, and analyzing the data. This object is not persistent, it is reinstantiated for each coin, each time the coin model needs to be inferenced or trained. Record of contribution: FreqAI was developed by a group of individuals who all contributed specific skillsets to the project. Conception and software development: Robert Caulk @robcaulk Theoretical brainstorming: Elin Törnquist @th0rntwig Code review, software architecture brainstorming: @xmatthias Beta testing and bug reporting: @bloodhunter4rc, Salah Lamkadem @ikonx, @ken11o2, @longyu, @paranoidandy, @smidelis, @smarm Juha Nykänen @suikula, Wagner Costa @wagnercosta, Johan Vlugt @Jooopieeert """ def __init__( self, config: Config, live: bool = False, pair: str = "", ): self.data: Dict[str, Any] = {} self.data_dictionary: Dict[str, DataFrame] = {} self.config = config self.freqai_config: Dict[str, Any] = config["freqai"] self.full_df: DataFrame = DataFrame() self.append_df: DataFrame = DataFrame() self.data_path = Path() self.label_list: List = [] self.training_features_list: List = [] self.model_filename: str = "" self.backtesting_results_path = Path() self.backtest_predictions_folder: str = "backtesting_predictions" self.live = live self.pair = pair self.svm_model: linear_model.SGDOneClassSVM = None self.keras: bool = self.freqai_config.get("keras", False) self.set_all_pairs() if not self.live: if not self.config["timerange"]: raise OperationalException( 'Please pass --timerange if you intend to use FreqAI for backtesting.') self.full_timerange = self.create_fulltimerange( self.config["timerange"], self.freqai_config.get("train_period_days", 0) ) (self.training_timeranges, self.backtesting_timeranges) = self.split_timerange( self.full_timerange, config["freqai"]["train_period_days"], config["freqai"]["backtest_period_days"], ) self.data['extra_returns_per_train'] = self.freqai_config.get('extra_returns_per_train', {}) self.thread_count = self.freqai_config.get("data_kitchen_thread_count", -1) self.train_dates: DataFrame = pd.DataFrame() self.unique_classes: Dict[str, list] = {} self.unique_class_list: list = [] def set_paths( self, pair: str, trained_timestamp: int = None, ) -> None: """ Set the paths to the data for the present coin/botloop :param metadata: dict = strategy furnished pair metadata :param trained_timestamp: int = timestamp of most recent training """ self.full_path = Path( self.config["user_data_dir"] / "models" / str(self.freqai_config.get("identifier")) ) self.data_path = Path( self.full_path / f"sub-train-{pair.split('/')[0]}_{trained_timestamp}" ) return def make_train_test_datasets( self, filtered_dataframe: DataFrame, labels: DataFrame ) -> Dict[Any, Any]: """ Given the dataframe for the full history for training, split the data into training and test data according to user specified parameters in configuration file. :param filtered_dataframe: cleaned dataframe ready to be split. :param labels: cleaned labels ready to be split. """ feat_dict = self.freqai_config["feature_parameters"] if 'shuffle' not in self.freqai_config['data_split_parameters']: self.freqai_config["data_split_parameters"].update({'shuffle': False}) weights: npt.ArrayLike if feat_dict.get("weight_factor", 0) > 0: weights = self.set_weights_higher_recent(len(filtered_dataframe)) else: weights = np.ones(len(filtered_dataframe)) if self.freqai_config.get('data_split_parameters', {}).get('test_size', 0.1) != 0: ( train_features, test_features, train_labels, test_labels, train_weights, test_weights, ) = train_test_split( filtered_dataframe[: filtered_dataframe.shape[0]], labels, weights, **self.config["freqai"]["data_split_parameters"], ) else: test_labels = np.zeros(2) test_features = pd.DataFrame() test_weights = np.zeros(2) train_features = filtered_dataframe train_labels = labels train_weights = weights # Simplest way to reverse the order of training and test data: if self.freqai_config['feature_parameters'].get('reverse_train_test_order', False): return self.build_data_dictionary( test_features, train_features, test_labels, train_labels, test_weights, train_weights ) else: return self.build_data_dictionary( train_features, test_features, train_labels, test_labels, train_weights, test_weights ) def filter_features( self, unfiltered_df: DataFrame, training_feature_list: List, label_list: List = list(), training_filter: bool = True, ) -> Tuple[DataFrame, DataFrame]: """ Filter the unfiltered dataframe to extract the user requested features/labels and properly remove all NaNs. Any row with a NaN is removed from training dataset or replaced with 0s in the prediction dataset. However, prediction dataset do_predict will reflect any row that had a NaN and will shield user from that prediction. :param unfiltered_df: the full dataframe for the present training period :param training_feature_list: list, the training feature list constructed by self.build_feature_list() according to user specified parameters in the configuration file. :param labels: the labels for the dataset :param training_filter: boolean which lets the function know if it is training data or prediction data to be filtered. :returns: :filtered_df: dataframe cleaned of NaNs and only containing the user requested feature set. :labels: labels cleaned of NaNs. """ filtered_df = unfiltered_df.filter(training_feature_list, axis=1) filtered_df = filtered_df.replace([np.inf, -np.inf], np.nan) drop_index = pd.isnull(filtered_df).any(axis=1) # get the rows that have NaNs, drop_index = drop_index.replace(True, 1).replace(False, 0) # pep8 requirement. if (training_filter): const_cols = list((filtered_df.nunique() == 1).loc[lambda x: x].index) if const_cols: filtered_df = filtered_df.filter(filtered_df.columns.difference(const_cols)) self.data['constant_features_list'] = const_cols logger.warning(f"Removed features {const_cols} with constant values.") else: self.data['constant_features_list'] = [] # we don't care about total row number (total no. datapoints) in training, we only care # about removing any row with NaNs # if labels has multiple columns (user wants to train multiple modelEs), we detect here labels = unfiltered_df.filter(label_list, axis=1) drop_index_labels = pd.isnull(labels).any(axis=1) drop_index_labels = drop_index_labels.replace(True, 1).replace(False, 0) dates = unfiltered_df['date'] filtered_df = filtered_df[ (drop_index == 0) & (drop_index_labels == 0) ] # dropping values labels = labels[ (drop_index == 0) & (drop_index_labels == 0) ] # assuming the labels depend entirely on the dataframe here. self.train_dates = dates[ (drop_index == 0) & (drop_index_labels == 0) ] logger.info( f"dropped {len(unfiltered_df) - len(filtered_df)} training points" f" due to NaNs in populated dataset {len(unfiltered_df)}." ) if (1 - len(filtered_df) / len(unfiltered_df)) > 0.1 and self.live: worst_indicator = str(unfiltered_df.count().idxmin()) logger.warning( f" {(1 - len(filtered_df)/len(unfiltered_df)) * 100:.0f} percent " " of training data dropped due to NaNs, model may perform inconsistent " f"with expectations. Verify {worst_indicator}" ) self.data["filter_drop_index_training"] = drop_index else: if len(self.data['constant_features_list']): filtered_df = self.check_pred_labels(filtered_df) # we are backtesting so we need to preserve row number to send back to strategy, # so now we use do_predict to avoid any prediction based on a NaN drop_index = pd.isnull(filtered_df).any(axis=1) self.data["filter_drop_index_prediction"] = drop_index filtered_df.fillna(0, inplace=True) # replacing all NaNs with zeros to avoid issues in 'prediction', but any prediction # that was based on a single NaN is ultimately protected from buys with do_predict drop_index = ~drop_index self.do_predict = np.array(drop_index.replace(True, 1).replace(False, 0)) if (len(self.do_predict) - self.do_predict.sum()) > 0: logger.info( "dropped %s of %s prediction data points due to NaNs.", len(self.do_predict) - self.do_predict.sum(), len(filtered_df), ) labels = [] return filtered_df, labels def build_data_dictionary( self, train_df: DataFrame, test_df: DataFrame, train_labels: DataFrame, test_labels: DataFrame, train_weights: Any, test_weights: Any, ) -> Dict: self.data_dictionary = { "train_features": train_df, "test_features": test_df, "train_labels": train_labels, "test_labels": test_labels, "train_weights": train_weights, "test_weights": test_weights, "train_dates": self.train_dates } return self.data_dictionary def normalize_data(self, data_dictionary: Dict) -> Dict[Any, Any]: """ Normalize all data in the data_dictionary according to the training dataset :param data_dictionary: dictionary containing the cleaned and split training/test data/labels :returns: :data_dictionary: updated dictionary with standardized values. """ # standardize the data by training stats train_max = data_dictionary["train_features"].max() train_min = data_dictionary["train_features"].min() data_dictionary["train_features"] = ( 2 * (data_dictionary["train_features"] - train_min) / (train_max - train_min) - 1 ) data_dictionary["test_features"] = ( 2 * (data_dictionary["test_features"] - train_min) / (train_max - train_min) - 1 ) for item in train_max.keys(): self.data[item + "_max"] = train_max[item] self.data[item + "_min"] = train_min[item] for item in data_dictionary["train_labels"].keys(): if data_dictionary["train_labels"][item].dtype == object: continue train_labels_max = data_dictionary["train_labels"][item].max() train_labels_min = data_dictionary["train_labels"][item].min() data_dictionary["train_labels"][item] = ( 2 * (data_dictionary["train_labels"][item] - train_labels_min) / (train_labels_max - train_labels_min) - 1 ) if self.freqai_config.get('data_split_parameters', {}).get('test_size', 0.1) != 0: data_dictionary["test_labels"][item] = ( 2 * (data_dictionary["test_labels"][item] - train_labels_min) / (train_labels_max - train_labels_min) - 1 ) self.data[f"{item}_max"] = train_labels_max self.data[f"{item}_min"] = train_labels_min return data_dictionary def normalize_single_dataframe(self, df: DataFrame) -> DataFrame: train_max = df.max() train_min = df.min() df = ( 2 * (df - train_min) / (train_max - train_min) - 1 ) for item in train_max.keys(): self.data[item + "_max"] = train_max[item] self.data[item + "_min"] = train_min[item] return df def normalize_data_from_metadata(self, df: DataFrame) -> DataFrame: """ Normalize a set of data using the mean and standard deviation from the associated training data. :param df: Dataframe to be standardized """ train_max = [None] * len(df.keys()) train_min = [None] * len(df.keys()) for i, item in enumerate(df.keys()): train_max[i] = self.data[f"{item}_max"] train_min[i] = self.data[f"{item}_min"] train_max_series = pd.Series(train_max, index=df.keys()) train_min_series = pd.Series(train_min, index=df.keys()) df = ( 2 * (df - train_min_series) / (train_max_series - train_min_series) - 1 ) return df def denormalize_labels_from_metadata(self, df: DataFrame) -> DataFrame: """ Denormalize a set of data using the mean and standard deviation from the associated training data. :param df: Dataframe of predictions to be denormalized """ for label in df.columns: if df[label].dtype == object or label in self.unique_class_list: continue df[label] = ( (df[label] + 1) * (self.data[f"{label}_max"] - self.data[f"{label}_min"]) / 2 ) + self.data[f"{label}_min"] return df def split_timerange( self, tr: str, train_split: int = 28, bt_split: float = 7 ) -> Tuple[list, list]: """ Function which takes a single time range (tr) and splits it into sub timeranges to train and backtest on based on user input tr: str, full timerange to train on train_split: the period length for the each training (days). Specified in user configuration file bt_split: the backtesting length (days). Specified in user configuration file """ if not isinstance(train_split, int) or train_split < 1: raise OperationalException( f"train_period_days must be an integer greater than 0. Got {train_split}." ) train_period_days = train_split * SECONDS_IN_DAY bt_period = bt_split * SECONDS_IN_DAY full_timerange = TimeRange.parse_timerange(tr) config_timerange = TimeRange.parse_timerange(self.config["timerange"]) if config_timerange.stopts == 0: config_timerange.stopts = int( datetime.now(tz=timezone.utc).timestamp() ) timerange_train = copy.deepcopy(full_timerange) timerange_backtest = copy.deepcopy(full_timerange) tr_training_list = [] tr_backtesting_list = [] tr_training_list_timerange = [] tr_backtesting_list_timerange = [] first = True while True: if not first: timerange_train.startts = timerange_train.startts + int(bt_period) timerange_train.stopts = timerange_train.startts + train_period_days first = False start = datetime.fromtimestamp(timerange_train.startts, tz=timezone.utc) stop = datetime.fromtimestamp(timerange_train.stopts, tz=timezone.utc) tr_training_list.append(start.strftime("%Y%m%d") + "-" + stop.strftime("%Y%m%d")) tr_training_list_timerange.append(copy.deepcopy(timerange_train)) # associated backtest period timerange_backtest.startts = timerange_train.stopts timerange_backtest.stopts = timerange_backtest.startts + int(bt_period) if timerange_backtest.stopts > config_timerange.stopts: timerange_backtest.stopts = config_timerange.stopts start = datetime.fromtimestamp(timerange_backtest.startts, tz=timezone.utc) stop = datetime.fromtimestamp(timerange_backtest.stopts, tz=timezone.utc) tr_backtesting_list.append(start.strftime("%Y%m%d") + "-" + stop.strftime("%Y%m%d")) tr_backtesting_list_timerange.append(copy.deepcopy(timerange_backtest)) # ensure we are predicting on exactly same amount of data as requested by user defined # --timerange if timerange_backtest.stopts == config_timerange.stopts: break # print(tr_training_list, tr_backtesting_list) return tr_training_list_timerange, tr_backtesting_list_timerange def slice_dataframe(self, timerange: TimeRange, df: DataFrame) -> DataFrame: """ Given a full dataframe, extract the user desired window :param tr: timerange string that we wish to extract from df :param df: Dataframe containing all candles to run the entire backtest. Here it is sliced down to just the present training period. """ start = datetime.fromtimestamp(timerange.startts, tz=timezone.utc) stop = datetime.fromtimestamp(timerange.stopts, tz=timezone.utc) df = df.loc[df["date"] >= start, :] if not self.live: df = df.loc[df["date"] < stop, :] return df def check_pred_labels(self, df_predictions: DataFrame) -> DataFrame: """ Check that prediction feature labels match training feature labels. :param df_predictions: incoming predictions """ constant_labels = self.data['constant_features_list'] df_predictions = df_predictions.filter( df_predictions.columns.difference(constant_labels) ) logger.warning( f"Removed {len(constant_labels)} features from prediction features, " f"these were considered constant values during most recent training." ) return df_predictions def principal_component_analysis(self) -> None: """ Performs Principal Component Analysis on the data for dimensionality reduction and outlier detection (see self.remove_outliers()) No parameters or returns, it acts on the data_dictionary held by the DataHandler. """ from sklearn.decomposition import PCA # avoid importing if we dont need it pca = PCA(0.999) pca = pca.fit(self.data_dictionary["train_features"]) n_keep_components = pca.n_components_ self.data["n_kept_components"] = n_keep_components n_components = self.data_dictionary["train_features"].shape[1] logger.info("reduced feature dimension by %s", n_components - n_keep_components) logger.info("explained variance %f", np.sum(pca.explained_variance_ratio_)) train_components = pca.transform(self.data_dictionary["train_features"]) self.data_dictionary["train_features"] = pd.DataFrame( data=train_components, columns=["PC" + str(i) for i in range(0, n_keep_components)], index=self.data_dictionary["train_features"].index, ) # normalsing transformed training features self.data_dictionary["train_features"] = self.normalize_single_dataframe( self.data_dictionary["train_features"]) # keeping a copy of the non-transformed features so we can check for errors during # model load from disk self.data["training_features_list_raw"] = copy.deepcopy(self.training_features_list) self.training_features_list = self.data_dictionary["train_features"].columns if self.freqai_config.get('data_split_parameters', {}).get('test_size', 0.1) != 0: test_components = pca.transform(self.data_dictionary["test_features"]) self.data_dictionary["test_features"] = pd.DataFrame( data=test_components, columns=["PC" + str(i) for i in range(0, n_keep_components)], index=self.data_dictionary["test_features"].index, ) # normalise transformed test feature to transformed training features self.data_dictionary["test_features"] = self.normalize_data_from_metadata( self.data_dictionary["test_features"]) self.data["n_kept_components"] = n_keep_components self.pca = pca logger.info(f"PCA reduced total features from {n_components} to {n_keep_components}") if not self.data_path.is_dir(): self.data_path.mkdir(parents=True, exist_ok=True) return None def pca_transform(self, filtered_dataframe: DataFrame) -> None: """ Use an existing pca transform to transform data into components :param filtered_dataframe: DataFrame = the cleaned dataframe """ pca_components = self.pca.transform(filtered_dataframe) self.data_dictionary["prediction_features"] = pd.DataFrame( data=pca_components, columns=["PC" + str(i) for i in range(0, self.data["n_kept_components"])], index=filtered_dataframe.index, ) # normalise transformed predictions to transformed training features self.data_dictionary["prediction_features"] = self.normalize_data_from_metadata( self.data_dictionary["prediction_features"]) def compute_distances(self) -> float: """ Compute distances between each training point and every other training point. This metric defines the neighborhood of trained data and is used for prediction confidence in the Dissimilarity Index """ # logger.info("computing average mean distance for all training points") pairwise = pairwise_distances( self.data_dictionary["train_features"], n_jobs=self.thread_count) # remove the diagonal distances which are itself distances ~0 np.fill_diagonal(pairwise, np.NaN) pairwise = pairwise.reshape(-1, 1) avg_mean_dist = pairwise[~np.isnan(pairwise)].mean() 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: 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 in training data and prediction :param predict: bool = If true, inference an existing SVM model, else construct one """ if self.keras: logger.warning( "SVM outlier removal not currently supported for Keras based models. " "Skipping user requested function." ) if predict: self.do_predict = np.ones(len(self.data_dictionary["prediction_features"])) return if predict: if not self.svm_model: logger.warning("No svm model available for outlier removal") return y_pred = self.svm_model.predict(self.data_dictionary["prediction_features"]) do_predict = np.where(y_pred == -1, 0, y_pred) if (len(do_predict) - do_predict.sum()) > 0: logger.info(f"SVM tossed {len(do_predict) - do_predict.sum()} predictions.") self.do_predict += do_predict self.do_predict -= 1 else: # use SGDOneClassSVM to increase speed? svm_params = self.freqai_config["feature_parameters"].get( "svm_params", {"shuffle": False, "nu": 0.1}) self.svm_model = linear_model.SGDOneClassSVM(**svm_params).fit( self.data_dictionary["train_features"] ) y_pred = self.svm_model.predict(self.data_dictionary["train_features"]) 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." ) self.svm_model = None return self.data_dictionary["train_features"] = self.data_dictionary["train_features"][ (y_pred == 1) ] self.data_dictionary["train_labels"] = self.data_dictionary["train_labels"][ (y_pred == 1) ] self.data_dictionary["train_weights"] = self.data_dictionary["train_weights"][ (y_pred == 1) ] logger.info( f"SVM tossed {len(y_pred) - kept_points.sum()}" f" train points from {len(y_pred)} total points." ) # same for test data # TODO: This (and the part above) could be refactored into a separate function # 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"]) kept_points = np.where(y_pred == -1, 0, y_pred) self.data_dictionary["test_features"] = self.data_dictionary["test_features"][ (y_pred == 1) ] self.data_dictionary["test_labels"] = self.data_dictionary["test_labels"][( y_pred == 1)] self.data_dictionary["test_weights"] = self.data_dictionary["test_weights"][ (y_pred == 1) ] logger.info( f"SVM tossed {len(y_pred) - kept_points.sum()}" f" test points from {len(y_pred)} total points." ) return def use_DBSCAN_to_remove_outliers(self, predict: bool, eps=None) -> None: """ Use DBSCAN to cluster training data and remove "noisy" data (read outliers). User controls this via the config param `DBSCAN_outlier_pct` which indicates the pct of training data that they want to be considered outliers. :param predict: bool = If False (training), iterate to find the best hyper parameters to match user requested outlier percent target. If True (prediction), use the parameters determined from the previous training to estimate if the current prediction point is an outlier. """ if predict: if not self.data['DBSCAN_eps']: return train_ft_df = self.data_dictionary['train_features'] pred_ft_df = self.data_dictionary['prediction_features'] num_preds = len(pred_ft_df) df = pd.concat([train_ft_df, pred_ft_df], axis=0, ignore_index=True) clustering = DBSCAN(eps=self.data['DBSCAN_eps'], min_samples=self.data['DBSCAN_min_samples'], n_jobs=self.thread_count ).fit(df) do_predict = np.where(clustering.labels_[-num_preds:] == -1, 0, 1) if (len(do_predict) - do_predict.sum()) > 0: logger.info(f"DBSCAN tossed {len(do_predict) - do_predict.sum()} predictions") self.do_predict += do_predict self.do_predict -= 1 else: def normalise_distances(distances): normalised_distances = (distances - distances.min()) / \ (distances.max() - distances.min()) return normalised_distances def rotate_point(origin, point, angle): # rotate a point counterclockwise by a given angle (in radians) # around a given origin x = origin[0] + cos(angle) * (point[0] - origin[0]) - \ sin(angle) * (point[1] - origin[1]) y = origin[1] + sin(angle) * (point[0] - origin[0]) + \ cos(angle) * (point[1] - origin[1]) return (x, y) 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']) distances, _ = neighbors_fit.kneighbors(self.data_dictionary['train_features']) distances = np.sort(distances, axis=0).mean(axis=1) normalised_distances = normalise_distances(distances) x_range = np.linspace(0, 1, len(distances)) line = np.linspace(normalised_distances[0], normalised_distances[-1], len(normalised_distances)) deflection = np.abs(normalised_distances - line) max_deflection_loc = np.where(deflection == deflection.max())[0][0] origin = x_range[max_deflection_loc], line[max_deflection_loc] point = x_range[max_deflection_loc], normalised_distances[max_deflection_loc] rot_angle = np.pi / 4 elbow_loc = rotate_point(origin, point, rot_angle) epsilon = elbow_loc[1] * (distances[-1] - distances[0]) + distances[0] clustering = DBSCAN(eps=epsilon, min_samples=MinPts, n_jobs=int(self.thread_count)).fit( self.data_dictionary['train_features'] ) logger.info(f'DBSCAN found eps of {epsilon:.2f}.') self.data['DBSCAN_eps'] = epsilon 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." ) self.data['DBSCAN_eps'] = 0 return self.data_dictionary['train_features'] = self.data_dictionary['train_features'][ (clustering.labels_ != -1) ] self.data_dictionary["train_labels"] = self.data_dictionary["train_labels"][ (clustering.labels_ != -1) ] self.data_dictionary["train_weights"] = self.data_dictionary["train_weights"][ (clustering.labels_ != -1) ] logger.info( f"DBSCAN tossed {dropped_points.sum()}" f" train points from {len(clustering.labels_)}" ) return def compute_inlier_metric(self, set_='train') -> None: """ Compute inlier metric from backwards distance distributions. This metric defines how well features from a timepoint fit into previous timepoints. """ def normalise(dataframe: DataFrame, key: str) -> DataFrame: if set_ == 'train': min_value = dataframe.min() max_value = dataframe.max() self.data[f'{key}_min'] = min_value self.data[f'{key}_max'] = max_value else: min_value = self.data[f'{key}_min'] max_value = self.data[f'{key}_max'] return (dataframe - min_value) / (max_value - min_value) no_prev_pts = self.freqai_config["feature_parameters"]["inlier_metric_window"] if set_ == 'train': compute_df = copy.deepcopy(self.data_dictionary['train_features']) elif set_ == 'test': compute_df = copy.deepcopy(self.data_dictionary['test_features']) else: compute_df = copy.deepcopy(self.data_dictionary['prediction_features']) compute_df_reindexed = compute_df.reindex( index=np.flip(compute_df.index) ) pairwise = pd.DataFrame( np.triu( pairwise_distances(compute_df_reindexed, n_jobs=self.thread_count) ), columns=compute_df_reindexed.index, index=compute_df_reindexed.index ) pairwise = pairwise.round(5) column_labels = [ '{}{}'.format('d', i) for i in range(1, no_prev_pts + 1) ] distances = pd.DataFrame( columns=column_labels, index=compute_df.index ) for index in compute_df.index[no_prev_pts:]: current_row = pairwise.loc[[index]] current_row_no_zeros = current_row.loc[ :, (current_row != 0).any(axis=0) ] distances.loc[[index]] = current_row_no_zeros.iloc[ :, :no_prev_pts ] distances = distances.replace([np.inf, -np.inf], np.nan) drop_index = pd.isnull(distances).any(axis=1) distances = distances[drop_index == 0] inliers = pd.DataFrame(index=distances.index) for key in distances.keys(): current_distances = distances[key].dropna() current_distances = normalise(current_distances, key) if set_ == 'train': fit_params = stats.weibull_min.fit(current_distances) self.data[f'{key}_fit_params'] = fit_params else: fit_params = self.data[f'{key}_fit_params'] quantiles = stats.weibull_min.cdf(current_distances, *fit_params) df_inlier = pd.DataFrame( {key: quantiles}, index=distances.index ) inliers = pd.concat( [inliers, df_inlier], axis=1 ) inlier_metric = pd.DataFrame( data=inliers.sum(axis=1) / no_prev_pts, columns=['%-inlier_metric'], index=compute_df.index ) inlier_metric = (2 * (inlier_metric - inlier_metric.min()) / (inlier_metric.max() - inlier_metric.min()) - 1) if set_ in ('train', 'test'): inlier_metric = inlier_metric.iloc[no_prev_pts:] compute_df = compute_df.iloc[no_prev_pts:] self.remove_beginning_points_from_data_dict(set_, no_prev_pts) self.data_dictionary[f'{set_}_features'] = pd.concat( [compute_df, inlier_metric], axis=1) else: self.data_dictionary['prediction_features'] = pd.concat( [compute_df, inlier_metric], axis=1) self.data_dictionary['prediction_features'].fillna(0, inplace=True) logger.info('Inlier metric computed and added to features.') return None def remove_beginning_points_from_data_dict(self, set_='train', no_prev_pts: int = 10): features = self.data_dictionary[f'{set_}_features'] weights = self.data_dictionary[f'{set_}_weights'] labels = self.data_dictionary[f'{set_}_labels'] self.data_dictionary[f'{set_}_weights'] = weights[no_prev_pts:] self.data_dictionary[f'{set_}_features'] = features.iloc[no_prev_pts:] self.data_dictionary[f'{set_}_labels'] = labels.iloc[no_prev_pts:] def add_noise_to_training_features(self) -> None: """ Add noise to train features to reduce the risk of overfitting. """ mu = 0 # no shift sigma = self.freqai_config["feature_parameters"]["noise_standard_deviation"] compute_df = self.data_dictionary['train_features'] noise = np.random.normal(mu, sigma, [compute_df.shape[0], compute_df.shape[1]]) self.data_dictionary['train_features'] += noise return def find_features(self, dataframe: DataFrame) -> None: """ Find features in the strategy provided dataframe :param dataframe: DataFrame = strategy provided dataframe :return: features: list = the features to be used for training/prediction """ column_names = dataframe.columns features = [c for c in column_names if "%" in c] if not features: raise OperationalException("Could not find any features!") self.training_features_list = features def find_labels(self, dataframe: DataFrame) -> None: column_names = dataframe.columns labels = [c for c in column_names if "&" in c] self.label_list = labels def check_if_pred_in_training_spaces(self) -> None: """ Compares the distance from each prediction point to each training data point. It uses this information to estimate a Dissimilarity Index (DI) and avoid making predictions on any points that are too far away from the training data set. """ distance = pairwise_distances( self.data_dictionary["train_features"], self.data_dictionary["prediction_features"], n_jobs=self.thread_count, ) self.DI_values = distance.min(axis=0) / self.data["avg_mean_dist"] do_predict = np.where( self.DI_values < self.freqai_config["feature_parameters"]["DI_threshold"], 1, 0, ) if (len(do_predict) - do_predict.sum()) > 0: logger.info( f"DI tossed {len(do_predict) - do_predict.sum()} predictions for " "being too far from training data." ) self.do_predict += do_predict self.do_predict -= 1 def set_weights_higher_recent(self, num_weights: int) -> npt.ArrayLike: """ Set weights so that recent data is more heavily weighted during training than older data. """ wfactor = self.config["freqai"]["feature_parameters"]["weight_factor"] weights = np.exp(-np.arange(num_weights) / (wfactor * num_weights))[::-1] return weights def get_predictions_to_append(self, predictions: DataFrame, do_predict: npt.ArrayLike) -> DataFrame: """ Get backtest prediction from current backtest period """ append_df = DataFrame() for label in predictions.columns: append_df[label] = predictions[label] if append_df[label].dtype == object: continue append_df[f"{label}_mean"] = self.data["labels_mean"][label] append_df[f"{label}_std"] = self.data["labels_std"][label] for extra_col in self.data["extra_returns_per_train"]: append_df[f"{extra_col}"] = self.data["extra_returns_per_train"][extra_col] append_df["do_predict"] = do_predict if self.freqai_config["feature_parameters"].get("DI_threshold", 0) > 0: append_df["DI_values"] = self.DI_values return append_df def append_predictions(self, append_df: DataFrame) -> None: """ Append backtest prediction from current backtest period to all previous periods """ if self.full_df.empty: self.full_df = append_df else: self.full_df = pd.concat([self.full_df, append_df], axis=0) def fill_predictions(self, dataframe): """ Back fill values to before the backtesting range so that the dataframe matches size when it goes back to the strategy. These rows are not included in the backtest. """ len_filler = len(dataframe) - len(self.full_df.index) # startup_candle_count filler_df = pd.DataFrame( np.zeros((len_filler, len(self.full_df.columns))), columns=self.full_df.columns ) self.full_df = pd.concat([filler_df, self.full_df], axis=0, ignore_index=True) to_keep = [col for col in dataframe.columns if not col.startswith("&")] self.return_dataframe = pd.concat([dataframe[to_keep], self.full_df], axis=1) self.full_df = DataFrame() return def create_fulltimerange(self, backtest_tr: str, backtest_period_days: int) -> str: if not isinstance(backtest_period_days, int): raise OperationalException("backtest_period_days must be an integer") if backtest_period_days < 0: raise OperationalException("backtest_period_days must be positive") backtest_timerange = TimeRange.parse_timerange(backtest_tr) if backtest_timerange.stopts == 0: # typically open ended time ranges do work, however, there are some edge cases where # it does not. accommodating these kinds of edge cases just to allow open-ended # timerange is not high enough priority to warrant the effort. It is safer for now # to simply ask user to add their end date raise OperationalException("FreqAI backtesting does not allow open ended timeranges. " "Please indicate the end date of your desired backtesting. " "timerange.") # backtest_timerange.stopts = int( # datetime.now(tz=timezone.utc).timestamp() # ) backtest_timerange.startts = ( backtest_timerange.startts - backtest_period_days * SECONDS_IN_DAY ) start = datetime.fromtimestamp(backtest_timerange.startts, tz=timezone.utc) stop = datetime.fromtimestamp(backtest_timerange.stopts, tz=timezone.utc) full_timerange = start.strftime("%Y%m%d") + "-" + stop.strftime("%Y%m%d") self.full_path = Path( self.config["user_data_dir"] / "models" / f"{self.freqai_config['identifier']}" ) config_path = Path(self.config["config_files"][0]) if not self.full_path.is_dir(): self.full_path.mkdir(parents=True, exist_ok=True) shutil.copy( config_path.resolve(), Path(self.full_path / config_path.parts[-1]), ) return full_timerange def check_if_model_expired(self, trained_timestamp: int) -> bool: """ A model age checker to determine if the model is trustworthy based on user defined `expiration_hours` in the configuration file. :param trained_timestamp: int = The time of training for the most recent model. :return: bool = If the model is expired or not. """ time = datetime.now(tz=timezone.utc).timestamp() elapsed_time = (time - trained_timestamp) / 3600 # hours max_time = self.freqai_config.get("expiration_hours", 0) if max_time > 0: return elapsed_time > max_time else: return False def check_if_new_training_required( self, trained_timestamp: int ) -> Tuple[bool, TimeRange, TimeRange]: time = datetime.now(tz=timezone.utc).timestamp() trained_timerange = TimeRange() data_load_timerange = TimeRange() timeframes = self.freqai_config["feature_parameters"].get("include_timeframes") max_tf_seconds = 0 for tf in timeframes: secs = timeframe_to_seconds(tf) if secs > max_tf_seconds: max_tf_seconds = secs # We notice that users like to use exotic indicators where # they do not know the required timeperiod. Here we include a factor # of safety by multiplying the user considered "max" by 2. max_period = self.config.get('startup_candle_count', 20) * 2 additional_seconds = max_period * max_tf_seconds if trained_timestamp != 0: elapsed_time = (time - trained_timestamp) / SECONDS_IN_HOUR retrain = elapsed_time > self.freqai_config.get("live_retrain_hours", 0) if retrain: trained_timerange.startts = int( time - self.freqai_config.get("train_period_days", 0) * SECONDS_IN_DAY ) trained_timerange.stopts = int(time) # we want to load/populate indicators on more data than we plan to train on so # because most of the indicators have a rolling timeperiod, and are thus NaNs # unless they have data further back in time before the start of the train period data_load_timerange.startts = int( time - self.freqai_config.get("train_period_days", 0) * SECONDS_IN_DAY - additional_seconds ) data_load_timerange.stopts = int(time) else: # user passed no live_trained_timerange in config trained_timerange.startts = int( time - self.freqai_config.get("train_period_days", 0) * SECONDS_IN_DAY ) trained_timerange.stopts = int(time) data_load_timerange.startts = int( time - self.freqai_config.get("train_period_days", 0) * SECONDS_IN_DAY - additional_seconds ) data_load_timerange.stopts = int(time) retrain = True return retrain, trained_timerange, data_load_timerange def set_new_model_names(self, pair: str, trained_timerange: TimeRange): coin, _ = pair.split("/") self.data_path = Path( self.full_path / f"sub-train-{pair.split('/')[0]}_{int(trained_timerange.stopts)}" ) self.model_filename = f"cb_{coin.lower()}_{int(trained_timerange.stopts)}" def set_all_pairs(self) -> None: self.all_pairs = copy.deepcopy( self.freqai_config["feature_parameters"].get("include_corr_pairlist", []) ) for pair in self.config.get("exchange", "").get("pair_whitelist"): if pair not in self.all_pairs: self.all_pairs.append(pair) def extract_corr_pair_columns_from_populated_indicators( self, dataframe: DataFrame ) -> Dict[str, DataFrame]: """ Find the columns of the dataframe corresponding to the corr_pairlist, save them in a dictionary to be reused and attached to other pairs. :param dataframe: fully populated dataframe (current pair + corr_pairs) :return: corr_dataframes, dictionary of dataframes to be attached to other pairs in same candle. """ corr_dataframes: Dict[str, DataFrame] = {} pairs = self.freqai_config["feature_parameters"].get("include_corr_pairlist", []) for pair in pairs: valid_strs = [f"%-{pair}", f"%{pair}", f"%_{pair}"] pair_cols = [col for col in dataframe.columns if any(substr in col for substr in valid_strs)] pair_cols.insert(0, 'date') corr_dataframes[pair] = dataframe.filter(pair_cols, axis=1) return corr_dataframes def attach_corr_pair_columns(self, dataframe: DataFrame, corr_dataframes: Dict[str, DataFrame], current_pair: str) -> DataFrame: """ Attach the existing corr_pair dataframes to the current pair dataframe before training :param dataframe: current pair strategy dataframe, indicators populated already :param corr_dataframes: dictionary of saved dataframes from earlier in the same candle :param current_pair: current pair to which we will attach corr pair dataframe :return: :dataframe: current pair dataframe of populated indicators, concatenated with corr_pairs ready for training """ pairs = self.freqai_config["feature_parameters"].get("include_corr_pairlist", []) for pair in pairs: if current_pair != pair: dataframe = dataframe.merge(corr_dataframes[pair], how='left', on='date') return dataframe def use_strategy_to_populate_indicators( self, strategy: IStrategy, corr_dataframes: dict = {}, base_dataframes: dict = {}, pair: str = "", prediction_dataframe: DataFrame = pd.DataFrame(), do_corr_pairs: bool = True, ) -> DataFrame: """ Use the user defined strategy for populating indicators during retrain :param strategy: IStrategy = user defined strategy object :param corr_dataframes: dict = dict containing the informative pair dataframes (for user defined timeframes) :param base_dataframes: dict = dict containing the current pair dataframes (for user defined timeframes) :param metadata: dict = strategy furnished pair metadata :return: dataframe: DataFrame = dataframe containing populated indicators """ # for prediction dataframe creation, we let dataprovider handle everything in the strategy # so we create empty dictionaries, which allows us to pass None to # `populate_any_indicators()`. Signaling we want the dp to give us the live dataframe. tfs: List[str] = self.freqai_config["feature_parameters"].get("include_timeframes") pairs: List[str] = self.freqai_config["feature_parameters"].get("include_corr_pairlist", []) if not prediction_dataframe.empty: dataframe = prediction_dataframe.copy() for tf in tfs: base_dataframes[tf] = None for p in pairs: if p not in corr_dataframes: corr_dataframes[p] = {} corr_dataframes[p][tf] = None else: dataframe = base_dataframes[self.config["timeframe"]].copy() sgi = False for tf in tfs: if tf == tfs[-1]: sgi = True # doing this last allows user to use all tf raw prices in labels dataframe = strategy.populate_any_indicators( pair, dataframe.copy(), tf, informative=base_dataframes[tf], set_generalized_indicators=sgi ) # ensure corr pairs are always last for corr_pair in pairs: if pair == corr_pair: continue # dont repeat anything from whitelist for tf in tfs: if pairs and do_corr_pairs: dataframe = strategy.populate_any_indicators( corr_pair, dataframe.copy(), tf, informative=corr_dataframes[corr_pair][tf] ) self.get_unique_classes_from_labels(dataframe) return dataframe def fit_labels(self) -> None: """ Fit the labels with a gaussian distribution """ import scipy as spy self.data["labels_mean"], self.data["labels_std"] = {}, {} for label in self.data_dictionary["train_labels"].columns: if self.data_dictionary["train_labels"][label].dtype == object: continue f = spy.stats.norm.fit(self.data_dictionary["train_labels"][label]) self.data["labels_mean"][label], self.data["labels_std"][label] = f[0], f[1] # incase targets are classifications for label in self.unique_class_list: self.data["labels_mean"][label], self.data["labels_std"][label] = 0, 0 return def remove_features_from_df(self, dataframe: DataFrame) -> DataFrame: """ Remove the features from the dataframe before returning it to strategy. This keeps it compact for Frequi purposes. """ to_keep = [ col for col in dataframe.columns if not col.startswith("%") or col.startswith("%%") ] return dataframe[to_keep] def get_unique_classes_from_labels(self, dataframe: DataFrame) -> None: # self.find_features(dataframe) self.find_labels(dataframe) for key in self.label_list: if dataframe[key].dtype == object: self.unique_classes[key] = dataframe[key].dropna().unique() if self.unique_classes: for label in self.unique_classes: self.unique_class_list += list(self.unique_classes[label]) def save_backtesting_prediction( self, append_df: DataFrame ) -> None: """ Save prediction dataframe from backtesting to h5 file format :param append_df: dataframe for backtesting period """ full_predictions_folder = Path(self.full_path / self.backtest_predictions_folder) if not full_predictions_folder.is_dir(): full_predictions_folder.mkdir(parents=True, exist_ok=True) append_df.to_hdf(self.backtesting_results_path, key='append_df', mode='w') def get_backtesting_prediction( self ) -> DataFrame: """ Get prediction dataframe from h5 file format """ append_df = pd.read_hdf(self.backtesting_results_path) return append_df def check_if_backtest_prediction_is_valid( self, length_backtesting_dataframe: int ) -> bool: """ Check if a backtesting prediction already exists and if the predictions to append has the same size of backtesting dataframe slice :param length_backtesting_dataframe: Length of backtesting dataframe slice :return: :boolean: whether the prediction file is valid. """ path_to_predictionfile = Path(self.full_path / self.backtest_predictions_folder / f"{self.model_filename}_prediction.h5") self.backtesting_results_path = path_to_predictionfile file_exists = path_to_predictionfile.is_file() if file_exists: append_df = self.get_backtesting_prediction() if len(append_df) == length_backtesting_dataframe: logger.info(f"Found backtesting prediction file at {path_to_predictionfile}") return True else: logger.info("A new backtesting prediction file is required. " "(Number of predictions is different from dataframe length).") return False else: logger.info( f"Could not find backtesting prediction file at {path_to_predictionfile}" ) return False