import logging import shutil import threading import time from abc import ABC, abstractmethod from datetime import datetime, timezone from pathlib import Path from threading import Lock from typing import Any, Dict, List, Tuple import numpy as np import pandas as pd from numpy.typing import NDArray from pandas import DataFrame from freqtrade.configuration import TimeRange from freqtrade.constants import DATETIME_PRINT_FORMAT, Config from freqtrade.enums import RunMode from freqtrade.exceptions import OperationalException from freqtrade.exchange import timeframe_to_seconds from freqtrade.freqai.data_drawer import FreqaiDataDrawer from freqtrade.freqai.data_kitchen import FreqaiDataKitchen from freqtrade.strategy.interface import IStrategy pd.options.mode.chained_assignment = None logger = logging.getLogger(__name__) class IFreqaiModel(ABC): """ Class containing all tools for training and prediction in the strategy. Base*PredictionModels inherit from this class. 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) -> None: self.config = config self.assert_config(self.config) self.freqai_info: Dict[str, Any] = config["freqai"] self.data_split_parameters: Dict[str, Any] = config.get("freqai", {}).get( "data_split_parameters", {}) self.model_training_parameters: Dict[str, Any] = config.get("freqai", {}).get( "model_training_parameters", {}) self.retrain = False self.first = True self.set_full_path() self.follow_mode: bool = self.freqai_info.get("follow_mode", False) self.save_backtest_models: bool = self.freqai_info.get("save_backtest_models", False) if self.save_backtest_models: logger.info('Backtesting module configured to save all models.') self.dd = FreqaiDataDrawer(Path(self.full_path), self.config, self.follow_mode) self.identifier: str = self.freqai_info.get("identifier", "no_id_provided") self.scanning = False self.ft_params = self.freqai_info["feature_parameters"] self.keras: bool = self.freqai_info.get("keras", False) if self.keras and self.ft_params.get("DI_threshold", 0): self.ft_params["DI_threshold"] = 0 logger.warning("DI threshold is not configured for Keras models yet. Deactivating.") self.CONV_WIDTH = self.freqai_info.get("conv_width", 2) if self.ft_params.get("inlier_metric_window", 0): self.CONV_WIDTH = self.ft_params.get("inlier_metric_window", 0) * 2 self.pair_it = 0 self.pair_it_train = 0 self.total_pairs = len(self.config.get("exchange", {}).get("pair_whitelist")) self.last_trade_database_summary: DataFrame = {} self.current_trade_database_summary: DataFrame = {} self.analysis_lock = Lock() self.inference_time: float = 0 self.train_time: float = 0 self.begin_time: float = 0 self.begin_time_train: float = 0 self.base_tf_seconds = timeframe_to_seconds(self.config['timeframe']) self.continual_learning = self.freqai_info.get('continual_learning', False) self._threads: List[threading.Thread] = [] self._stop_event = threading.Event() def __getstate__(self): """ Return an empty state to be pickled in hyperopt """ return ({}) def assert_config(self, config: Config) -> None: if not config.get("freqai", {}): raise OperationalException("No freqai parameters found in configuration file.") def start(self, dataframe: DataFrame, metadata: dict, strategy: IStrategy) -> DataFrame: """ Entry point to the FreqaiModel from a specific pair, it will train a new model if necessary before making the prediction. :param dataframe: Full dataframe coming from strategy - it contains entire backtesting timerange + additional historical data necessary to train the model. :param metadata: pair metadata coming from strategy. :param strategy: Strategy to train on """ self.live = strategy.dp.runmode in (RunMode.DRY_RUN, RunMode.LIVE) self.dd.set_pair_dict_info(metadata) if self.live: self.inference_timer('start') self.dk = FreqaiDataKitchen(self.config, self.live, metadata["pair"]) dk = self.start_live(dataframe, metadata, strategy, self.dk) # For backtesting, each pair enters and then gets trained for each window along the # sliding window defined by "train_period_days" (training window) and "live_retrain_hours" # (backtest window, i.e. window immediately following the training window). # FreqAI slides the window and sequentially builds the backtesting results before returning # the concatenated results for the full backtesting period back to the strategy. elif not self.follow_mode: self.dk = FreqaiDataKitchen(self.config, self.live, metadata["pair"]) logger.info(f"Training {len(self.dk.training_timeranges)} timeranges") dataframe = self.dk.use_strategy_to_populate_indicators( strategy, prediction_dataframe=dataframe, pair=metadata["pair"] ) dk = self.start_backtesting(dataframe, metadata, self.dk) dataframe = dk.remove_features_from_df(dk.return_dataframe) self.clean_up() if self.live: self.inference_timer('stop') return dataframe def clean_up(self): """ Objects that should be handled by GC already between coins, but are explicitly shown here to help demonstrate the non-persistence of these objects. """ self.model = None self.dk = None def shutdown(self): """ Cleans up threads on Shutdown, set stop event. Join threads to wait for current training iteration. """ logger.info("Stopping FreqAI") self._stop_event.set() logger.info("Waiting on Training iteration") for _thread in self._threads: _thread.join() def start_scanning(self, *args, **kwargs) -> None: """ Start `self._start_scanning` in a separate thread """ _thread = threading.Thread(target=self._start_scanning, args=args, kwargs=kwargs) self._threads.append(_thread) _thread.start() def _start_scanning(self, strategy: IStrategy) -> None: """ Function designed to constantly scan pairs for retraining on a separate thread (intracandle) to improve model youth. This function is agnostic to data preparation/collection/storage, it simply trains on what ever data is available in the self.dd. :param strategy: IStrategy = The user defined strategy class """ while not self._stop_event.is_set(): time.sleep(1) for pair in self.config.get("exchange", {}).get("pair_whitelist"): (_, trained_timestamp, _) = self.dd.get_pair_dict_info(pair) if self.dd.pair_dict[pair]["priority"] != 1: continue dk = FreqaiDataKitchen(self.config, self.live, pair) dk.set_paths(pair, trained_timestamp) ( retrain, new_trained_timerange, data_load_timerange, ) = dk.check_if_new_training_required(trained_timestamp) dk.set_paths(pair, new_trained_timerange.stopts) if retrain: self.train_timer('start') self.extract_data_and_train_model( new_trained_timerange, pair, strategy, dk, data_load_timerange ) self.train_timer('stop') self.dd.save_historic_predictions_to_disk() def start_backtesting( self, dataframe: DataFrame, metadata: dict, dk: FreqaiDataKitchen ) -> FreqaiDataKitchen: """ The main broad execution for backtesting. For backtesting, each pair enters and then gets trained for each window along the sliding window defined by "train_period_days" (training window) and "backtest_period_days" (backtest window, i.e. window immediately following the training window). FreqAI slides the window and sequentially builds the backtesting results before returning the concatenated results for the full backtesting period back to the strategy. :param dataframe: DataFrame = strategy passed dataframe :param metadata: Dict = pair metadata :param dk: FreqaiDataKitchen = Data management/analysis tool associated to present pair only :return: FreqaiDataKitchen = Data management/analysis tool associated to present pair only """ self.pair_it += 1 train_it = 0 # Loop enforcing the sliding window training/backtesting paradigm # tr_train is the training time range e.g. 1 historical month # tr_backtest is the backtesting time range e.g. the week directly # following tr_train. Both of these windows slide through the # entire backtest for tr_train, tr_backtest in zip(dk.training_timeranges, dk.backtesting_timeranges): (_, _, _) = self.dd.get_pair_dict_info(metadata["pair"]) train_it += 1 total_trains = len(dk.backtesting_timeranges) self.training_timerange = tr_train dataframe_train = dk.slice_dataframe(tr_train, dataframe) dataframe_backtest = dk.slice_dataframe(tr_backtest, dataframe) trained_timestamp = tr_train tr_train_startts_str = datetime.fromtimestamp( tr_train.startts, tz=timezone.utc).strftime(DATETIME_PRINT_FORMAT) tr_train_stopts_str = datetime.fromtimestamp( tr_train.stopts, tz=timezone.utc).strftime(DATETIME_PRINT_FORMAT) logger.info( f"Training {metadata['pair']}, {self.pair_it}/{self.total_pairs} pairs" f" from {tr_train_startts_str} to {tr_train_stopts_str}, {train_it}/{total_trains} " "trains" ) trained_timestamp_int = int(trained_timestamp.stopts) dk.data_path = Path( dk.full_path / f"sub-train-{metadata['pair'].split('/')[0]}_{trained_timestamp_int}" ) dk.set_new_model_names(metadata["pair"], trained_timestamp) if dk.check_if_backtest_prediction_exists(): append_df = dk.get_backtesting_prediction() dk.append_predictions(append_df) else: if not self.model_exists( metadata["pair"], dk, trained_timestamp=trained_timestamp_int ): dk.find_features(dataframe_train) self.model = self.train(dataframe_train, metadata["pair"], dk) self.dd.pair_dict[metadata["pair"]]["trained_timestamp"] = int( trained_timestamp.stopts) if self.save_backtest_models: logger.info('Saving backtest model to disk.') self.dd.save_data(self.model, metadata["pair"], dk) else: self.model = self.dd.load_data(metadata["pair"], dk) self.check_if_feature_list_matches_strategy(dataframe_train, dk) pred_df, do_preds = self.predict(dataframe_backtest, dk) append_df = dk.get_predictions_to_append(pred_df, do_preds) dk.append_predictions(append_df) dk.save_backtesting_prediction(append_df) dk.fill_predictions(dataframe) return dk def start_live( self, dataframe: DataFrame, metadata: dict, strategy: IStrategy, dk: FreqaiDataKitchen ) -> FreqaiDataKitchen: """ The main broad execution for dry/live. This function will check if a retraining should be performed, and if so, retrain and reset the model. :param dataframe: DataFrame = strategy passed dataframe :param metadata: Dict = pair metadata :param strategy: IStrategy = currently employed strategy dk: FreqaiDataKitchen = Data management/analysis tool associated to present pair only :returns: dk: FreqaiDataKitchen = Data management/analysis tool associated to present pair only """ # update follower if self.follow_mode: self.dd.update_follower_metadata() # get the model metadata associated with the current pair (_, trained_timestamp, return_null_array) = self.dd.get_pair_dict_info(metadata["pair"]) # if the metadata doesn't exist, the follower returns null arrays to strategy if self.follow_mode and return_null_array: logger.info("Returning null array from follower to strategy") self.dd.return_null_values_to_strategy(dataframe, dk) return dk # append the historic data once per round if self.dd.historic_data: self.dd.update_historic_data(strategy, dk) logger.debug(f'Updating historic data on pair {metadata["pair"]}') if not self.follow_mode: (_, new_trained_timerange, data_load_timerange) = dk.check_if_new_training_required( trained_timestamp ) dk.set_paths(metadata["pair"], new_trained_timerange.stopts) # load candle history into memory if it is not yet. if not self.dd.historic_data: self.dd.load_all_pair_histories(data_load_timerange, dk) if not self.scanning: self.scanning = True self.start_scanning(strategy) elif self.follow_mode: dk.set_paths(metadata["pair"], trained_timestamp) logger.info( "FreqAI instance set to follow_mode, finding existing pair " f"using { self.identifier }" ) # load the model and associated data into the data kitchen self.model = self.dd.load_data(metadata["pair"], dk) with self.analysis_lock: dataframe = self.dk.use_strategy_to_populate_indicators( strategy, prediction_dataframe=dataframe, pair=metadata["pair"] ) if not self.model: logger.warning( f"No model ready for {metadata['pair']}, returning null values to strategy." ) self.dd.return_null_values_to_strategy(dataframe, dk) return dk # ensure user is feeding the correct indicators to the model self.check_if_feature_list_matches_strategy(dataframe, dk) self.build_strategy_return_arrays(dataframe, dk, metadata["pair"], trained_timestamp) return dk def build_strategy_return_arrays( self, dataframe: DataFrame, dk: FreqaiDataKitchen, pair: str, trained_timestamp: int ) -> None: # hold the historical predictions in memory so we are sending back # correct array to strategy if pair not in self.dd.model_return_values: # first predictions are made on entire historical candle set coming from strategy. This # allows FreqUI to show full return values. 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) self.dd.set_initial_return_values(pair, pred_df) dk.return_dataframe = self.dd.attach_return_values_to_return_dataframe(pair, dataframe) return elif self.dk.check_if_model_expired(trained_timestamp): pred_df = DataFrame(np.zeros((2, len(dk.label_list))), columns=dk.label_list) do_preds = np.ones(2, dtype=np.int_) * 2 dk.DI_values = np.zeros(2) logger.warning( f"Model expired for {pair}, returning null values to strategy. Strategy " "construction should take care to consider this event with " "prediction == 0 and do_predict == 2" ) else: # remaining predictions are made only on the most recent candles for performance and # historical accuracy reasons. pred_df, do_preds = self.predict(dataframe.iloc[-self.CONV_WIDTH:], dk, first=False) if self.freqai_info.get('fit_live_predictions_candles', 0) and self.live: self.fit_live_predictions(dk, pair) self.dd.append_model_predictions(pair, pred_df, do_preds, dk, len(dataframe)) dk.return_dataframe = self.dd.attach_return_values_to_return_dataframe(pair, dataframe) return def check_if_feature_list_matches_strategy( self, dataframe: DataFrame, dk: FreqaiDataKitchen ) -> None: """ Ensure user is passing the proper feature set if they are reusing an `identifier` pointing to a folder holding existing models. :param dataframe: DataFrame = strategy provided dataframe :param dk: FreqaiDataKitchen = non-persistent data container/analyzer for current coin/bot loop """ dk.find_features(dataframe) if "training_features_list_raw" in dk.data: feature_list = dk.data["training_features_list_raw"] else: feature_list = dk.training_features_list if dk.training_features_list != feature_list: raise OperationalException( "Trying to access pretrained model with `identifier` " "but found different features furnished by current strategy." "Change `identifier` to train from scratch, or ensure the" "strategy is furnishing the same features as the pretrained" "model" ) def data_cleaning_train(self, dk: FreqaiDataKitchen) -> None: """ Base data cleaning method for train. Functions here improve/modify the input data by identifying outliers, computing additional metrics, adding noise, reducing dimensionality etc. """ ft_params = self.freqai_info["feature_parameters"] if ft_params.get('inlier_metric_window', 0): dk.compute_inlier_metric(set_='train') if self.freqai_info["data_split_parameters"]["test_size"] > 0: dk.compute_inlier_metric(set_='test') if ft_params.get( "principal_component_analysis", False ): dk.principal_component_analysis() if ft_params.get("use_SVM_to_remove_outliers", False): dk.use_SVM_to_remove_outliers(predict=False) if ft_params.get("DI_threshold", 0): dk.data["avg_mean_dist"] = dk.compute_distances() if ft_params.get("use_DBSCAN_to_remove_outliers", False): if dk.pair in self.dd.old_DBSCAN_eps: eps = self.dd.old_DBSCAN_eps[dk.pair] else: eps = None dk.use_DBSCAN_to_remove_outliers(predict=False, eps=eps) self.dd.old_DBSCAN_eps[dk.pair] = dk.data['DBSCAN_eps'] if self.freqai_info["feature_parameters"].get('noise_standard_deviation', 0): dk.add_noise_to_training_features() def data_cleaning_predict(self, dk: FreqaiDataKitchen, dataframe: DataFrame) -> None: """ Base data cleaning method for predict. Functions here are complementary to the functions of data_cleaning_train. """ ft_params = self.freqai_info["feature_parameters"] if ft_params.get('inlier_metric_window', 0): dk.compute_inlier_metric(set_='predict') if ft_params.get( "principal_component_analysis", False ): dk.pca_transform(self.dk.data_dictionary['prediction_features']) if ft_params.get("use_SVM_to_remove_outliers", False): dk.use_SVM_to_remove_outliers(predict=True) if ft_params.get("DI_threshold", 0): dk.check_if_pred_in_training_spaces() if ft_params.get("use_DBSCAN_to_remove_outliers", False): dk.use_DBSCAN_to_remove_outliers(predict=True) def model_exists( self, pair: str, dk: FreqaiDataKitchen, trained_timestamp: int = None, model_filename: str = "", scanning: bool = False, ) -> bool: """ Given a pair and path, check if a model already exists :param pair: pair e.g. BTC/USD :param path: path to model :return: :boolean: whether the model file exists or not. """ path_to_modelfile = Path(dk.data_path / f"{model_filename}_model.joblib") file_exists = path_to_modelfile.is_file() if file_exists and not scanning: logger.info("Found model at %s", dk.data_path / dk.model_filename) elif not scanning: logger.info("Could not find model at %s", dk.data_path / dk.model_filename) return file_exists def set_full_path(self) -> None: self.full_path = Path( self.config["user_data_dir"] / "models" / f"{self.freqai_info['identifier']}" ) self.full_path.mkdir(parents=True, exist_ok=True) shutil.copy( self.config["config_files"][0], Path(self.full_path, Path(self.config["config_files"][0]).name), ) def extract_data_and_train_model( self, new_trained_timerange: TimeRange, pair: str, strategy: IStrategy, dk: FreqaiDataKitchen, data_load_timerange: TimeRange, ): """ Retrieve data and train model. :param new_trained_timerange: TimeRange = the timerange to train the model on :param metadata: dict = strategy provided metadata :param strategy: IStrategy = user defined strategy object :param dk: FreqaiDataKitchen = non-persistent data container for current coin/loop :param data_load_timerange: TimeRange = the amount of data to be loaded for populate_any_indicators (larger than new_trained_timerange so that new_trained_timerange does not contain any NaNs) """ corr_dataframes, base_dataframes = self.dd.get_base_and_corr_dataframes( data_load_timerange, pair, dk ) with self.analysis_lock: unfiltered_dataframe = dk.use_strategy_to_populate_indicators( strategy, corr_dataframes, base_dataframes, pair ) unfiltered_dataframe = dk.slice_dataframe(new_trained_timerange, unfiltered_dataframe) # find the features indicated by strategy and store in datakitchen dk.find_features(unfiltered_dataframe) model = self.train(unfiltered_dataframe, pair, dk) self.dd.pair_dict[pair]["trained_timestamp"] = new_trained_timerange.stopts dk.set_new_model_names(pair, new_trained_timerange) self.dd.pair_dict[pair]["first"] = False if self.dd.pair_dict[pair]["priority"] == 1 and self.scanning: self.dd.pair_to_end_of_training_queue(pair) self.dd.save_data(model, pair, dk) if self.freqai_info.get("purge_old_models", False): self.dd.purge_old_models() def set_initial_historic_predictions( self, pred_df: DataFrame, dk: FreqaiDataKitchen, pair: str ) -> None: """ This function is called only if the datadrawer failed to load an existing set of historic predictions. In this case, it builds the structure and sets fake predictions off the first training data. After that, FreqAI will append new real predictions to the set of historic predictions. These values are used to generate live statistics which can be used in the strategy for adaptive values. E.g. &*_mean/std are quantities that can computed based on live predictions from the set of historical predictions. Those values can be used in the user strategy to better assess prediction rarity, and thus wait for probabilistically favorable entries relative to the live historical predictions. 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: dk: FreqaiDataKitchen = object containing methods for data analysis :param: pair: str = current pair """ self.dd.historic_predictions[pair] = pred_df hist_preds_df = self.dd.historic_predictions[pair] for label in hist_preds_df.columns: if hist_preds_df[label].dtype == object: continue hist_preds_df[f'{label}_mean'] = 0 hist_preds_df[f'{label}_std'] = 0 hist_preds_df['do_predict'] = 0 if self.freqai_info['feature_parameters'].get('DI_threshold', 0) > 0: hist_preds_df['DI_values'] = 0 for return_str in dk.data['extra_returns_per_train']: hist_preds_df[return_str] = 0 # # for keras type models, the conv_window needs to be prepended so # # viewing is correct in frequi if self.freqai_info.get('keras', False) or self.ft_params.get('inlier_metric_window', 0): n_lost_points = self.freqai_info.get('conv_width', 2) zeros_df = DataFrame(np.zeros((n_lost_points, len(hist_preds_df.columns))), columns=hist_preds_df.columns) self.dd.historic_predictions[pair] = pd.concat( [zeros_df, hist_preds_df], axis=0, ignore_index=True) def fit_live_predictions(self, dk: FreqaiDataKitchen, pair: str) -> None: """ Fit the labels with a gaussian distribution """ import scipy as spy # add classes from classifier label types if used full_labels = dk.label_list + dk.unique_class_list num_candles = self.freqai_info.get("fit_live_predictions_candles", 100) dk.data["labels_mean"], dk.data["labels_std"] = {}, {} for label in full_labels: if self.dd.historic_predictions[dk.pair][label].dtype == object: continue f = spy.stats.norm.fit(self.dd.historic_predictions[dk.pair][label].tail(num_candles)) dk.data["labels_mean"][label], dk.data["labels_std"][label] = f[0], f[1] return def inference_timer(self, do='start'): """ Timer designed to track the cumulative time spent in FreqAI for one pass through the whitelist. This will check if the time spent is more than 1/4 the time of a single candle, and if so, it will warn the user of degraded performance """ if do == 'start': self.pair_it += 1 self.begin_time = time.time() elif do == 'stop': end = time.time() self.inference_time += (end - self.begin_time) if self.pair_it == self.total_pairs: logger.info( f'Total time spent inferencing pairlist {self.inference_time:.2f} seconds') if self.inference_time > 0.25 * self.base_tf_seconds: logger.warning("Inference took over 25% of the candle time. Reduce pairlist to" " avoid blinding open trades and degrading performance.") self.pair_it = 0 self.inference_time = 0 return def train_timer(self, do='start'): """ Timer designed to track the cumulative time spent training the full pairlist in FreqAI. """ if do == 'start': self.pair_it_train += 1 self.begin_time_train = time.time() elif do == 'stop': end = time.time() self.train_time += (end - self.begin_time_train) if self.pair_it_train == self.total_pairs: logger.info( f'Total time spent training pairlist {self.train_time:.2f} seconds') self.pair_it_train = 0 self.train_time = 0 return def get_init_model(self, pair: str) -> Any: if pair not in self.dd.model_dictionary or not self.continual_learning: init_model = None else: init_model = self.dd.model_dictionary[pair] return init_model # Following methods which are overridden by user made prediction models. # See freqai/prediction_models/CatboostPredictionModel.py for an example. @abstractmethod def train(self, unfiltered_df: DataFrame, pair: str, dk: FreqaiDataKitchen, **kwargs) -> Any: """ Filter the training data and train a model to it. Train makes heavy use of the datahandler for storing, saving, loading, and analyzing the data. :param unfiltered_df: Full dataframe for the current training period :param metadata: pair metadata from strategy. :return: Trained model which can be used to inference (self.predict) """ @abstractmethod def fit(self, data_dictionary: Dict[str, Any], dk: FreqaiDataKitchen, **kwargs) -> Any: """ Most regressors use the same function names and arguments e.g. user can drop in LGBMRegressor in place of CatBoostRegressor and all data management will be properly handled by Freqai. :param data_dictionary: Dict = the dictionary constructed by DataHandler to hold all the training and test data/labels. """ return @abstractmethod def predict( self, unfiltered_df: DataFrame, dk: FreqaiDataKitchen, **kwargs ) -> Tuple[DataFrame, NDArray[np.int_]]: """ Filter the prediction features data and predict with it. :param unfiltered_df: Full dataframe for the current backtest period. :param dk: FreqaiDataKitchen = Data management/analysis tool associated to present pair only :param first: boolean = whether this is the first prediction or not. :return: :predictions: np.array of predictions :do_predict: np.array of 1s and 0s to indicate places where freqai needed to remove data (NaNs) or felt uncertain about data (i.e. SVM and/or DI index) """