Merge branch 'develop' into precise_calcs

freqtrade/commands/arguments.py

@@ -12,7 +12,8 @@ from freqtrade.constants import DEFAULT_CONFIG
 
 ARGS_COMMON = ["verbosity", "logfile", "version", "config", "datadir", "user_data_dir"]
 
-ARGS_STRATEGY = ["strategy", "strategy_path", "recursive_strategy_search"]
+ARGS_STRATEGY = ["strategy", "strategy_path", "recursive_strategy_search", "freqaimodel",
+                 "freqaimodel_path"]
 
 ARGS_TRADE = ["db_url", "sd_notify", "dry_run", "dry_run_wallet", "fee"]
 

freqtrade/commands/cli_options.py

@@ -647,4 +647,14 @@ AVAILABLE_CLI_OPTIONS = {
         nargs='+',
         default=[],
     ),
+    "freqaimodel": Arg(
+        '--freqaimodel',
+        help='Specify a custom freqaimodels.',
+        metavar='NAME',
+    ),
+    "freqaimodel_path": Arg(
+        '--freqaimodel-path',
+        help='Specify additional lookup path for freqaimodels.',
+        metavar='PATH',
+    ),
 }

freqtrade/commands/data_commands.py

@@ -12,7 +12,7 @@ from freqtrade.enums import CandleType, RunMode, TradingMode
 from freqtrade.exceptions import OperationalException
 from freqtrade.exchange import timeframe_to_minutes
 from freqtrade.exchange.exchange import market_is_active
-from freqtrade.plugins.pairlist.pairlist_helpers import expand_pairlist
+from freqtrade.plugins.pairlist.pairlist_helpers import dynamic_expand_pairlist, expand_pairlist
 from freqtrade.resolvers import ExchangeResolver
 
 
@@ -50,7 +50,8 @@ def start_download_data(args: Dict[str, Any]) -> None:
     exchange = ExchangeResolver.load_exchange(config['exchange']['name'], config, validate=False)
     markets = [p for p, m in exchange.markets.items() if market_is_active(m)
                or config.get('include_inactive')]
-    expanded_pairs = expand_pairlist(config['pairs'], markets)
+
+    expanded_pairs = dynamic_expand_pairlist(config, markets)
 
     # Manual validations of relevant settings
     if not config['exchange'].get('skip_pair_validation', False):

freqtrade/configuration/configuration.py

@@ -97,6 +97,8 @@ class Configuration:
 
         self._process_analyze_options(config)
 
+        self._process_freqai_options(config)
+
         # Check if the exchange set by the user is supported
         check_exchange(config, config.get('experimental', {}).get('block_bad_exchanges', True))
 
@@ -461,6 +463,16 @@ class Configuration:
 
         config.update({'runmode': self.runmode})
 
+    def _process_freqai_options(self, config: Dict[str, Any]) -> None:
+
+        self._args_to_config(config, argname='freqaimodel',
+                             logstring='Using freqaimodel class name: {}')
+
+        self._args_to_config(config, argname='freqaimodel_path',
+                             logstring='Using freqaimodel path: {}')
+
+        return
+
     def _args_to_config(self, config: Dict[str, Any], argname: str,
                         logstring: str, logfun: Optional[Callable] = None,
                         deprecated_msg: Optional[str] = None) -> None:

freqtrade/constants.py

@@ -55,6 +55,7 @@ FTHYPT_FILEVERSION = 'fthypt_fileversion'
 USERPATH_HYPEROPTS = 'hyperopts'
 USERPATH_STRATEGIES = 'strategies'
 USERPATH_NOTEBOOKS = 'notebooks'
+USERPATH_FREQAIMODELS = 'freqaimodels'
 
 TELEGRAM_SETTING_OPTIONS = ['on', 'off', 'silent']
 WEBHOOK_FORMAT_OPTIONS = ['form', 'json', 'raw']
@@ -240,6 +241,7 @@ CONF_SCHEMA = {
         },
         'exchange': {'$ref': '#/definitions/exchange'},
         'edge': {'$ref': '#/definitions/edge'},
+        'freqai': {'$ref': '#/definitions/freqai'},
         'experimental': {
             'type': 'object',
             'properties': {
@@ -480,7 +482,60 @@ CONF_SCHEMA = {
                 'remove_pumps': {'type': 'boolean'}
             },
             'required': ['process_throttle_secs', 'allowed_risk']
-        }
+        },
+        "freqai": {
+            "type": "object",
+            "properties": {
+                "enabled": {"type": "boolean", "default": False},
+                "keras": {"type": "boolean", "default": False},
+                "conv_width": {"type": "integer", "default": 2},
+                "train_period_days": {"type": "integer", "default": 0},
+                "backtest_period_days": {"type": "number", "default": 7},
+                "identifier": {"type": "string", "default": "example"},
+                "feature_parameters": {
+                    "type": "object",
+                    "properties": {
+                        "include_corr_pairlist": {"type": "array"},
+                        "include_timeframes": {"type": "array"},
+                        "label_period_candles": {"type": "integer"},
+                        "include_shifted_candles": {"type": "integer", "default": 0},
+                        "DI_threshold": {"type": "number", "default": 0},
+                        "weight_factor": {"type": "number", "default": 0},
+                        "principal_component_analysis": {"type": "boolean", "default": False},
+                        "use_SVM_to_remove_outliers": {"type": "boolean", "default": False},
+                        "svm_params": {"type": "object",
+                                       "properties": {
+                                           "shuffle": {"type": "boolean", "default": False},
+                                           "nu": {"type": "number", "default": 0.1}
+                                           },
+                                       }
+                    },
+                    "required": ["include_timeframes", "include_corr_pairlist", ]
+                },
+                "data_split_parameters": {
+                    "type": "object",
+                    "properties": {
+                        "test_size": {"type": "number"},
+                        "random_state": {"type": "integer"},
+                    },
+                },
+                "model_training_parameters": {
+                    "type": "object",
+                    "properties": {
+                        "n_estimators": {"type": "integer", "default": 1000}
+                    },
+                },
+            },
+            "required": [
+                "enabled",
+                "train_period_days",
+                "backtest_period_days",
+                "identifier",
+                "feature_parameters",
+                "data_split_parameters",
+                "model_training_parameters"
+                ]
+        },
     },
 }
 

freqtrade/exchange/exchange.py

@@ -1009,7 +1009,8 @@ class Exchange:
         time_in_force: str = 'gtc',
     ) -> Dict:
         if self._config['dry_run']:
-            dry_order = self.create_dry_run_order(pair, ordertype, side, amount, rate, leverage)
+            dry_order = self.create_dry_run_order(
+                pair, ordertype, side, amount, self.price_to_precision(pair, rate), leverage)
             return dry_order
 
         params = self._get_params(side, ordertype, leverage, reduceOnly, time_in_force)

freqtrade/freqai/data_drawer.py

@@ -0,0 +1,609 @@
+import collections
+import json
+import logging
+import re
+import shutil
+import threading
+from pathlib import Path
+from typing import Any, Dict, Tuple, TypedDict
+
+import numpy as np
+import pandas as pd
+import rapidjson
+from joblib import dump, load
+from joblib.externals import cloudpickle
+from numpy.typing import NDArray
+from pandas import DataFrame
+
+from freqtrade.configuration import TimeRange
+from freqtrade.data.history import load_pair_history
+from freqtrade.exceptions import OperationalException
+from freqtrade.freqai.data_kitchen import FreqaiDataKitchen
+from freqtrade.strategy.interface import IStrategy
+
+
+logger = logging.getLogger(__name__)
+
+
+class pair_info(TypedDict):
+    model_filename: str
+    first: bool
+    trained_timestamp: int
+    priority: int
+    data_path: str
+    extras: dict
+
+
+class FreqaiDataDrawer:
+    """
+    Class aimed at holding all pair models/info in memory for better inferencing/retrainig/saving
+    /loading to/from disk.
+    This object remains persistent throughout live/dry.
+
+    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, full_path: Path, config: dict, follow_mode: bool = False):
+
+        self.config = config
+        self.freqai_info = config.get("freqai", {})
+        # dictionary holding all pair metadata necessary to load in from disk
+        self.pair_dict: Dict[str, pair_info] = {}
+        # dictionary holding all actively inferenced models in memory given a model filename
+        self.model_dictionary: Dict[str, Any] = {}
+        self.model_return_values: Dict[str, DataFrame] = {}
+        self.historic_data: Dict[str, Dict[str, DataFrame]] = {}
+        self.historic_predictions: Dict[str, DataFrame] = {}
+        self.follower_dict: Dict[str, pair_info] = {}
+        self.full_path = full_path
+        self.follower_name: str = self.config.get("bot_name", "follower1")
+        self.follower_dict_path = Path(
+            self.full_path / f"follower_dictionary-{self.follower_name}.json"
+        )
+        self.historic_predictions_path = Path(self.full_path / "historic_predictions.pkl")
+        self.pair_dictionary_path = Path(self.full_path / "pair_dictionary.json")
+        self.follow_mode = follow_mode
+        if follow_mode:
+            self.create_follower_dict()
+        self.load_drawer_from_disk()
+        self.load_historic_predictions_from_disk()
+        self.training_queue: Dict[str, int] = {}
+        self.history_lock = threading.Lock()
+        self.save_lock = threading.Lock()
+        self.pair_dict_lock = threading.Lock()
+        self.old_DBSCAN_eps: Dict[str, float] = {}
+        self.empty_pair_dict: pair_info = {
+                "model_filename": "", "trained_timestamp": 0,
+                "priority": 1, "first": True, "data_path": "", "extras": {}}
+
+    def load_drawer_from_disk(self):
+        """
+        Locate and load a previously saved data drawer full of all pair model metadata in
+        present model folder.
+        :return: bool - whether or not the drawer was located
+        """
+        exists = self.pair_dictionary_path.is_file()
+        if exists:
+            with open(self.pair_dictionary_path, "r") as fp:
+                self.pair_dict = json.load(fp)
+        elif not self.follow_mode:
+            logger.info("Could not find existing datadrawer, starting from scratch")
+        else:
+            logger.warning(
+                f"Follower could not find pair_dictionary at {self.full_path} "
+                "sending null values back to strategy"
+            )
+
+        return exists
+
+    def load_historic_predictions_from_disk(self):
+        """
+        Locate and load a previously saved historic predictions.
+        :return: bool - whether or not the drawer was located
+        """
+        exists = self.historic_predictions_path.is_file()
+        if exists:
+            with open(self.historic_predictions_path, "rb") as fp:
+                self.historic_predictions = cloudpickle.load(fp)
+            logger.info(
+                f"Found existing historic predictions at {self.full_path}, but beware "
+                "that statistics may be inaccurate if the bot has been offline for "
+                "an extended period of time."
+            )
+        elif not self.follow_mode:
+            logger.info("Could not find existing historic_predictions, starting from scratch")
+        else:
+            logger.warning(
+                f"Follower could not find historic predictions at {self.full_path} "
+                "sending null values back to strategy"
+            )
+
+        return exists
+
+    def save_historic_predictions_to_disk(self):
+        """
+        Save data drawer full of all pair model metadata in present model folder.
+        """
+        with open(self.historic_predictions_path, "wb") as fp:
+            cloudpickle.dump(self.historic_predictions, fp, protocol=cloudpickle.DEFAULT_PROTOCOL)
+
+    def save_drawer_to_disk(self):
+        """
+        Save data drawer full of all pair model metadata in present model folder.
+        """
+        with self.save_lock:
+            with open(self.pair_dictionary_path, 'w') as fp:
+                rapidjson.dump(self.pair_dict, fp, default=self.np_encoder,
+                               number_mode=rapidjson.NM_NATIVE)
+
+    def save_follower_dict_to_disk(self):
+        """
+        Save follower dictionary to disk (used by strategy for persistent prediction targets)
+        """
+        with open(self.follower_dict_path, "w") as fp:
+            rapidjson.dump(self.follower_dict, fp, default=self.np_encoder,
+                           number_mode=rapidjson.NM_NATIVE)
+
+    def create_follower_dict(self):
+        """
+        Create or dictionary for each follower to maintain unique persistent prediction targets
+        """
+
+        whitelist_pairs = self.config.get("exchange", {}).get("pair_whitelist")
+
+        exists = self.follower_dict_path.is_file()
+
+        if exists:
+            logger.info("Found an existing follower dictionary")
+
+        for pair in whitelist_pairs:
+            self.follower_dict[pair] = {}
+
+        self.save_follower_dict_to_disk()
+
+    def np_encoder(self, object):
+        if isinstance(object, np.generic):
+            return object.item()
+
+    def get_pair_dict_info(self, pair: str) -> Tuple[str, int, bool]:
+        """
+        Locate and load existing model metadata from persistent storage. If not located,
+        create a new one and append the current pair to it and prepare it for its first
+        training
+        :param pair: str: pair to lookup
+        :return:
+            model_filename: str = unique filename used for loading persistent objects from disk
+            trained_timestamp: int = the last time the coin was trained
+            return_null_array: bool = Follower could not find pair metadata
+        """
+
+        pair_dict = self.pair_dict.get(pair)
+        data_path_set = self.pair_dict.get(pair, self.empty_pair_dict).get("data_path", "")
+        return_null_array = False
+
+        if pair_dict:
+            model_filename = pair_dict["model_filename"]
+            trained_timestamp = pair_dict["trained_timestamp"]
+        elif not self.follow_mode:
+            self.pair_dict[pair] = self.empty_pair_dict.copy()
+            model_filename = ""
+            trained_timestamp = 0
+            self.pair_dict[pair]["priority"] = len(self.pair_dict)
+
+        if not data_path_set and self.follow_mode:
+            logger.warning(
+                f"Follower could not find current pair {pair} in "
+                f"pair_dictionary at path {self.full_path}, sending null values "
+                "back to strategy."
+            )
+            trained_timestamp = 0
+            model_filename = ''
+            return_null_array = True
+
+        return model_filename, trained_timestamp, return_null_array
+
+    def set_pair_dict_info(self, metadata: dict) -> None:
+        pair_in_dict = self.pair_dict.get(metadata["pair"])
+        if pair_in_dict:
+            return
+        else:
+            self.pair_dict[metadata["pair"]] = self.empty_pair_dict.copy()
+            self.pair_dict[metadata["pair"]]["priority"] = len(self.pair_dict)
+
+            return
+
+    def pair_to_end_of_training_queue(self, pair: str) -> None:
+        # march all pairs up in the queue
+        with self.pair_dict_lock:
+            for p in self.pair_dict:
+                self.pair_dict[p]["priority"] -= 1
+            # send pair to end of queue
+            self.pair_dict[pair]["priority"] = len(self.pair_dict)
+
+    def set_initial_return_values(self, pair: str, pred_df: 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)
+        """
+
+        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())
+        else:
+            df_concat = hist_df[pair].tail(len(pred_df.index)).reset_index(drop=True)
+        df_concat = df_concat.fillna(0)
+        self.model_return_values[pair] = df_concat
+
+    def append_model_predictions(self, pair: str, predictions: DataFrame,
+                                 do_preds: NDArray[np.int_],
+                                 dk: FreqaiDataKitchen, len_df: int) -> None:
+        """
+        Append model predictions to historic predictions dataframe, then set the
+        strategy return dataframe to the tail of the historic predictions. The length of
+        the tail is equivalent to the length of the dataframe that entered FreqAI from
+        the strategy originally. Doing this allows FreqUI to always display the correct
+        historic predictions.
+        """
+
+        index = self.historic_predictions[pair].index[-1:]
+        columns = self.historic_predictions[pair].columns
+
+        nan_df = pd.DataFrame(np.nan, index=index, columns=columns)
+        self.historic_predictions[pair] = pd.concat(
+            [self.historic_predictions[pair], nan_df], ignore_index=True, axis=0)
+        df = self.historic_predictions[pair]
+
+        # model outputs and associated statistics
+        for label in predictions.columns:
+            df[label].iloc[-1] = predictions[label].iloc[-1]
+            if df[label].dtype == object:
+                continue
+            df[f"{label}_mean"].iloc[-1] = dk.data["labels_mean"][label]
+            df[f"{label}_std"].iloc[-1] = dk.data["labels_std"][label]
+
+        # outlier indicators
+        df["do_predict"].iloc[-1] = do_preds[-1]
+        if self.freqai_info["feature_parameters"].get("DI_threshold", 0) > 0:
+            df["DI_values"].iloc[-1] = dk.DI_values[-1]
+
+        # extra values the user added within custom prediction model
+        if dk.data['extra_returns_per_train']:
+            rets = dk.data['extra_returns_per_train']
+            for return_str in rets:
+                df[return_str].iloc[-1] = rets[return_str]
+
+        self.model_return_values[pair] = df.tail(len_df).reset_index(drop=True)
+
+    def attach_return_values_to_return_dataframe(
+            self, pair: str, dataframe: DataFrame) -> DataFrame:
+        """
+        Attach the return values to the strat dataframe
+        :param dataframe: DataFrame = strategy dataframe
+        :return: DataFrame = strat dataframe with return values attached
+        """
+        df = self.model_return_values[pair]
+        to_keep = [col for col in dataframe.columns if not col.startswith("&")]
+        dataframe = pd.concat([dataframe[to_keep], df], axis=1)
+        return dataframe
+
+    def return_null_values_to_strategy(self, dataframe: DataFrame, dk: FreqaiDataKitchen) -> None:
+        """
+        Build 0 filled dataframe to return to strategy
+        """
+
+        dk.find_features(dataframe)
+
+        full_labels = dk.label_list + dk.unique_class_list
+
+        for label in full_labels:
+            dataframe[label] = 0
+            dataframe[f"{label}_mean"] = 0
+            dataframe[f"{label}_std"] = 0
+
+        dataframe["do_predict"] = 0
+
+        if self.freqai_info["feature_parameters"].get("DI_threshold", 0) > 0:
+            dataframe["DI_values"] = 0
+
+        if dk.data['extra_returns_per_train']:
+            rets = dk.data['extra_returns_per_train']
+            for return_str in rets:
+                dataframe[return_str] = 0
+
+        dk.return_dataframe = dataframe
+
+    def purge_old_models(self) -> None:
+
+        model_folders = [x for x in self.full_path.iterdir() if x.is_dir()]
+
+        pattern = re.compile(r"sub-train-(\w+)_(\d{10})")
+
+        delete_dict: Dict[str, Any] = {}
+
+        for dir in model_folders:
+            result = pattern.match(str(dir.name))
+            if result is None:
+                break
+            coin = result.group(1)
+            timestamp = result.group(2)
+
+            if coin not in delete_dict:
+                delete_dict[coin] = {}
+                delete_dict[coin]["num_folders"] = 1
+                delete_dict[coin]["timestamps"] = {int(timestamp): dir}
+            else:
+                delete_dict[coin]["num_folders"] += 1
+                delete_dict[coin]["timestamps"][int(timestamp)] = dir
+
+        for coin in delete_dict:
+            if delete_dict[coin]["num_folders"] > 2:
+                sorted_dict = collections.OrderedDict(
+                    sorted(delete_dict[coin]["timestamps"].items())
+                )
+                num_delete = len(sorted_dict) - 2
+                deleted = 0
+                for k, v in sorted_dict.items():
+                    if deleted >= num_delete:
+                        break
+                    logger.info(f"Freqai purging old model file {v}")
+                    shutil.rmtree(v)
+                    deleted += 1
+
+    def update_follower_metadata(self):
+        # follower needs to load from disk to get any changes made by leader to pair_dict
+        self.load_drawer_from_disk()
+        if self.config.get("freqai", {}).get("purge_old_models", False):
+            self.purge_old_models()
+
+    # Functions pulled back from FreqaiDataKitchen because they relied on DataDrawer
+
+    def save_data(self, model: Any, coin: str, dk: FreqaiDataKitchen) -> None:
+        """
+        Saves all data associated with a model for a single sub-train time range
+        :params:
+        :model: User trained model which can be reused for inferencing to generate
+        predictions
+        """
+
+        if not dk.data_path.is_dir():
+            dk.data_path.mkdir(parents=True, exist_ok=True)
+
+        save_path = Path(dk.data_path)
+
+        # Save the trained model
+        if not dk.keras:
+            dump(model, save_path / f"{dk.model_filename}_model.joblib")
+        else:
+            model.save(save_path / f"{dk.model_filename}_model.h5")
+
+        if dk.svm_model is not None:
+            dump(dk.svm_model, save_path / f"{dk.model_filename}_svm_model.joblib")
+
+        dk.data["data_path"] = str(dk.data_path)
+        dk.data["model_filename"] = str(dk.model_filename)
+        dk.data["training_features_list"] = list(dk.data_dictionary["train_features"].columns)
+        dk.data["label_list"] = dk.label_list
+        # store the metadata
+        with open(save_path / f"{dk.model_filename}_metadata.json", "w") as fp:
+            rapidjson.dump(dk.data, fp, default=self.np_encoder, number_mode=rapidjson.NM_NATIVE)
+
+        # save the train data to file so we can check preds for area of applicability later
+        dk.data_dictionary["train_features"].to_pickle(
+            save_path / f"{dk.model_filename}_trained_df.pkl"
+        )
+
+        dk.data_dictionary["train_dates"].to_pickle(
+            save_path / f"{dk.model_filename}_trained_dates_df.pkl"
+        )
+
+        if self.freqai_info["feature_parameters"].get("principal_component_analysis"):
+            cloudpickle.dump(
+                dk.pca, open(dk.data_path / f"{dk.model_filename}_pca_object.pkl", "wb")
+            )
+
+        # if self.live:
+        self.model_dictionary[dk.model_filename] = model
+        self.pair_dict[coin]["model_filename"] = dk.model_filename
+        self.pair_dict[coin]["data_path"] = str(dk.data_path)
+        self.save_drawer_to_disk()
+
+        return
+
+    def load_data(self, coin: str, dk: FreqaiDataKitchen) -> Any:
+        """
+        loads all data required to make a prediction on a sub-train time range
+        :returns:
+        :model: User trained model which can be inferenced for new predictions
+        """
+
+        if not self.pair_dict[coin]["model_filename"]:
+            return None
+
+        if dk.live:
+            dk.model_filename = self.pair_dict[coin]["model_filename"]
+            dk.data_path = Path(self.pair_dict[coin]["data_path"])
+            if self.freqai_info.get("follow_mode", False):
+                # follower can be on a different system which is rsynced from the leader:
+                dk.data_path = Path(
+                    self.config["user_data_dir"]
+                    / "models"
+                    / dk.data_path.parts[-2]
+                    / dk.data_path.parts[-1]
+                )
+
+        with open(dk.data_path / f"{dk.model_filename}_metadata.json", "r") as fp:
+            dk.data = json.load(fp)
+            dk.training_features_list = dk.data["training_features_list"]
+            dk.label_list = dk.data["label_list"]
+
+        dk.data_dictionary["train_features"] = pd.read_pickle(
+            dk.data_path / f"{dk.model_filename}_trained_df.pkl"
+        )
+
+        # try to access model in memory instead of loading object from disk to save time
+        if dk.live and dk.model_filename in self.model_dictionary:
+            model = self.model_dictionary[dk.model_filename]
+        elif not dk.keras:
+            model = load(dk.data_path / f"{dk.model_filename}_model.joblib")
+        else:
+            from tensorflow import keras
+
+            model = keras.models.load_model(dk.data_path / f"{dk.model_filename}_model.h5")
+
+        if Path(dk.data_path / f"{dk.model_filename}_svm_model.joblib").is_file():
+            dk.svm_model = load(dk.data_path / f"{dk.model_filename}_svm_model.joblib")
+
+        if not model:
+            raise OperationalException(
+                f"Unable to load model, ensure model exists at " f"{dk.data_path} "
+            )
+
+        if self.config["freqai"]["feature_parameters"]["principal_component_analysis"]:
+            dk.pca = cloudpickle.load(
+                open(dk.data_path / f"{dk.model_filename}_pca_object.pkl", "rb")
+            )
+
+        return model
+
+    def update_historic_data(self, strategy: IStrategy, dk: FreqaiDataKitchen) -> None:
+        """
+        Append new candles to our stores historic data (in memory) so that
+        we do not need to load candle history from disk and we dont need to
+        pinging exchange multiple times for the same candle.
+        :params:
+        dataframe: DataFrame = strategy provided dataframe
+        """
+        feat_params = self.freqai_info["feature_parameters"]
+        with self.history_lock:
+            history_data = self.historic_data
+
+            for pair in dk.all_pairs:
+                for tf in feat_params.get("include_timeframes"):
+
+                    # check if newest candle is already appended
+                    df_dp = strategy.dp.get_pair_dataframe(pair, tf)
+                    if len(df_dp.index) == 0:
+                        continue
+                    if str(history_data[pair][tf].iloc[-1]["date"]) == str(
+                        df_dp.iloc[-1:]["date"].iloc[-1]
+                    ):
+                        continue
+
+                    try:
+                        index = (
+                            df_dp.loc[
+                                df_dp["date"] == history_data[pair][tf].iloc[-1]["date"]
+                            ].index[0]
+                            + 1
+                        )
+                    except IndexError:
+                        logger.warning(
+                            f"Unable to update pair history for {pair}. "
+                            "If this does not resolve itself after 1 additional candle, "
+                            "please report the error to #freqai discord channel"
+                        )
+                        return
+
+                    history_data[pair][tf] = pd.concat(
+                        [
+                            history_data[pair][tf],
+                            df_dp.iloc[index:],
+                        ],
+                        ignore_index=True,
+                        axis=0,
+                    )
+
+    def load_all_pair_histories(self, timerange: TimeRange, dk: FreqaiDataKitchen) -> None:
+        """
+        Load pair histories for all whitelist and corr_pairlist pairs.
+        Only called once upon startup of bot.
+        :params:
+        timerange: TimeRange = full timerange required to populate all indicators
+        for training according to user defined train_period_days
+        """
+        history_data = self.historic_data
+
+        for pair in dk.all_pairs:
+            if pair not in history_data:
+                history_data[pair] = {}
+            for tf in self.freqai_info["feature_parameters"].get("include_timeframes"):
+                history_data[pair][tf] = load_pair_history(
+                    datadir=self.config["datadir"],
+                    timeframe=tf,
+                    pair=pair,
+                    timerange=timerange,
+                    data_format=self.config.get("dataformat_ohlcv", "json"),
+                    candle_type=self.config.get("trading_mode", "spot"),
+                )
+
+    def get_base_and_corr_dataframes(
+        self, timerange: TimeRange, pair: str, dk: FreqaiDataKitchen
+    ) -> Tuple[Dict[Any, Any], Dict[Any, Any]]:
+        """
+        Searches through our historic_data in memory and returns the dataframes relevant
+        to the present pair.
+        :params:
+        timerange: TimeRange = full timerange required to populate all indicators
+        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] = {}
+            historic_data = self.historic_data
+            pairs = self.freqai_info["feature_parameters"].get(
+                "include_corr_pairlist", []
+            )
+
+            for tf in self.freqai_info["feature_parameters"].get("include_timeframes"):
+                base_dataframes[tf] = dk.slice_dataframe(timerange, historic_data[pair][tf])
+                if pairs:
+                    for p in pairs:
+                        if pair in p:
+                            continue  # dont repeat anything from whitelist
+                        if p not in corr_dataframes:
+                            corr_dataframes[p] = {}
+                        corr_dataframes[p][tf] = dk.slice_dataframe(
+                            timerange, historic_data[p][tf]
+                        )
+
+        return corr_dataframes, base_dataframes
+
+    # to be used if we want to send predictions directly to the follower instead of forcing
+    # follower to load models and inference
+    # def save_model_return_values_to_disk(self) -> None:
+    #     with open(self.full_path / str('model_return_values.json'), "w") as fp:
+    #         json.dump(self.model_return_values, fp, default=self.np_encoder)
+
+    # def load_model_return_values_from_disk(self, dk: FreqaiDataKitchen) -> FreqaiDataKitchen:
+    #     exists = Path(self.full_path / str('model_return_values.json')).resolve().exists()
+    #     if exists:
+    #         with open(self.full_path / str('model_return_values.json'), "r") as fp:
+    #             self.model_return_values = json.load(fp)
+    #     elif not self.follow_mode:
+    #         logger.info("Could not find existing datadrawer, starting from scratch")
+    #     else:
+    #         logger.warning(f'Follower could not find pair_dictionary at {self.full_path} '
+    #                        'sending null values back to strategy')
+
+    #     return exists, dk

freqtrade/freqai/freqai_interface.py

@@ -0,0 +1,653 @@
+# import contextlib
+import datetime
+import logging
+import shutil
+import threading
+import time
+from abc import ABC, abstractmethod
+from pathlib import Path
+from threading import Lock
+from typing import Any, Dict, 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.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__)
+
+
+def threaded(fn):
+    def wrapper(*args, **kwargs):
+        threading.Thread(target=fn, args=args, kwargs=kwargs).start()
+
+    return wrapper
+
+
+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: Dict[str, Any]) -> 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.feature_parameters = config.get("freqai", {}).get("feature_parameters")
+        self.retrain = False
+        self.first = True
+        self.set_full_path()
+        self.follow_mode: bool = self.freqai_info.get("follow_mode", False)
+        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.keras: bool = self.freqai_info.get("keras", False)
+        if self.keras and self.freqai_info.get("feature_parameters", {}).get("DI_threshold", 0):
+            self.freqai_info["feature_parameters"]["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)
+        self.pair_it = 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.begin_time: float = 0
+        self.base_tf_seconds = timeframe_to_seconds(self.config['timeframe'])
+
+    def assert_config(self, config: Dict[str, Any]) -> 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")
+            with self.analysis_lock:
+                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)
+        del dk
+        if self.live:
+            self.inference_timer('stop')
+        return dataframe
+
+    @threaded
+    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 1:
+            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_model_in_series(
+                        new_trained_timerange, pair, strategy, dk, data_load_timerange
+                    )
+
+            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.datetime.utcfromtimestamp(tr_train.startts).strftime(
+                "%Y-%m-%d %H:%M:%S"
+            )
+            tr_train_stopts_str = datetime.datetime.utcfromtimestamp(tr_train.stopts).strftime(
+                "%Y-%m-%d %H:%M:%S"
+            )
+            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"
+            )
+
+            dk.data_path = Path(
+                dk.full_path
+                /
+                f"sub-train-{metadata['pair'].split('/')[0]}_{int(trained_timestamp.stopts)}"
+                )
+            if not self.model_exists(
+                metadata["pair"], dk, trained_timestamp=int(trained_timestamp.stopts)
+            ):
+                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)
+                dk.set_new_model_names(metadata["pair"], trained_timestamp)
+                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)
+
+            dk.append_predictions(pred_df, do_preds)
+
+        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)
+
+            # download candle history if it is not already in memory
+            if not self.dd.historic_data:
+                logger.info(
+                    "Downloading all training data for all pairs in whitelist and "
+                    "corr_pairlist, this may take a while if you do not have the "
+                    "data saved"
+                )
+                dk.download_all_data_for_training(data_load_timerange, strategy.dp)
+                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
+        Any function inside this method should drop training data points from the filtered_dataframe
+        based on user decided logic. See FreqaiDataKitchen::use_SVM_to_remove_outliers() for an
+        example of how outlier data points are dropped from the dataframe used for training.
+        """
+
+        if self.freqai_info["feature_parameters"].get(
+            "principal_component_analysis", False
+        ):
+            dk.principal_component_analysis()
+
+        if self.freqai_info["feature_parameters"].get("use_SVM_to_remove_outliers", False):
+            dk.use_SVM_to_remove_outliers(predict=False)
+
+        if self.freqai_info["feature_parameters"].get("DI_threshold", 0):
+            dk.data["avg_mean_dist"] = dk.compute_distances()
+
+        if self.freqai_info["feature_parameters"].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']
+
+    def data_cleaning_predict(self, dk: FreqaiDataKitchen, dataframe: DataFrame) -> None:
+        """
+        Base data cleaning method for predict.
+        These functions each modify dk.do_predict, which is a dataframe with equal length
+        to the number of candles coming from and returning to the strategy. Inside do_predict,
+         1 allows prediction and < 0 signals to the strategy that the model is not confident in
+         the prediction.
+         See FreqaiDataKitchen::remove_outliers() for an example
+        of how the do_predict vector is modified. do_predict is ultimately passed back to strategy
+        for buy signals.
+        """
+        if self.freqai_info["feature_parameters"].get(
+            "principal_component_analysis", False
+        ):
+            dk.pca_transform(dataframe)
+
+        if self.freqai_info["feature_parameters"].get("use_SVM_to_remove_outliers", False):
+            dk.use_SVM_to_remove_outliers(predict=True)
+
+        if self.freqai_info["feature_parameters"].get("DI_threshold", 0):
+            dk.check_if_pred_in_training_spaces()
+
+        if self.freqai_info["feature_parameters"].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.
+        """
+        coin, _ = pair.split("/")
+
+        if not self.live:
+            dk.model_filename = model_filename = f"cb_{coin.lower()}_{trained_timestamp}"
+
+        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 train_model_in_series(
+        self,
+        new_trained_timerange: TimeRange,
+        pair: str,
+        strategy: IStrategy,
+        dk: FreqaiDataKitchen,
+        data_load_timerange: TimeRange,
+    ):
+        """
+        Retrieve data and train model in single threaded mode (only used if model directory is empty
+        upon startup for dry/live )
+        :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):
+            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
+
+    # Following methods which are overridden by user made prediction models.
+    # See freqai/prediction_models/CatboostPredictionModel.py for an example.
+
+    @abstractmethod
+    def train(self, unfiltered_dataframe: DataFrame, pair: str, dk: FreqaiDataKitchen) -> 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_dataframe: 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]) -> 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, dataframe: DataFrame, dk: FreqaiDataKitchen, first: bool = True
+    ) -> Tuple[DataFrame, NDArray[np.int_]]:
+        """
+        Filter the prediction features data and predict with it.
+        :param unfiltered_dataframe: 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)
+        """

freqtrade/freqai/prediction_models/BaseClassifierModel.py

@@ -0,0 +1,99 @@
+import logging
+from typing import Any, Tuple
+
+import numpy as np
+import numpy.typing as npt
+import pandas as pd
+from pandas import DataFrame
+
+from freqtrade.freqai.data_kitchen import FreqaiDataKitchen
+from freqtrade.freqai.freqai_interface import IFreqaiModel
+
+
+logger = logging.getLogger(__name__)
+
+
+class BaseClassifierModel(IFreqaiModel):
+    """
+    Base class for regression type models (e.g. Catboost, LightGBM, XGboost etc.).
+    User *must* inherit from this class and set fit() and predict(). See example scripts
+    such as prediction_models/CatboostPredictionModel.py for guidance.
+    """
+
+    def train(
+        self, unfiltered_dataframe: DataFrame, pair: str, dk: FreqaiDataKitchen
+    ) -> Any:
+        """
+        Filter the training data and train a model to it. Train makes heavy use of the datakitchen
+        for storing, saving, loading, and analyzing the data.
+        :param unfiltered_dataframe: Full dataframe for the current training period
+        :param metadata: pair metadata from strategy.
+        :return:
+        :model: Trained model which can be used to inference (self.predict)
+        """
+
+        logger.info("-------------------- Starting training " f"{pair} --------------------")
+
+        # filter the features requested by user in the configuration file and elegantly handle NaNs
+        features_filtered, labels_filtered = dk.filter_features(
+            unfiltered_dataframe,
+            dk.training_features_list,
+            dk.label_list,
+            training_filter=True,
+        )
+
+        start_date = unfiltered_dataframe["date"].iloc[0].strftime("%Y-%m-%d")
+        end_date = unfiltered_dataframe["date"].iloc[-1].strftime("%Y-%m-%d")
+        logger.info(f"-------------------- Training on data from {start_date} to "
+                    f"{end_date}--------------------")
+        # split data into train/test data.
+        data_dictionary = dk.make_train_test_datasets(features_filtered, labels_filtered)
+        if not self.freqai_info.get('fit_live_predictions', 0) or not self.live:
+            dk.fit_labels()
+        # normalize all data based on train_dataset only
+        data_dictionary = dk.normalize_data(data_dictionary)
+
+        # optional additional data cleaning/analysis
+        self.data_cleaning_train(dk)
+
+        logger.info(
+            f'Training model on {len(dk.data_dictionary["train_features"].columns)}' " features"
+        )
+        logger.info(f'Training model on {len(data_dictionary["train_features"])} data points')
+
+        model = self.fit(data_dictionary)
+
+        logger.info(f"--------------------done training {pair}--------------------")
+
+        return model
+
+    def predict(
+        self, unfiltered_dataframe: DataFrame, dk: FreqaiDataKitchen, first: bool = False
+    ) -> Tuple[DataFrame, npt.NDArray[np.int_]]:
+        """
+        Filter the prediction features data and predict with it.
+        :param: unfiltered_dataframe: Full dataframe for the current backtest period.
+        :return:
+        :pred_df: dataframe containing the predictions
+        :do_predict: np.array of 1s and 0s to indicate places where freqai needed to remove
+        data (NaNs) or felt uncertain about data (PCA and DI index)
+        """
+
+        dk.find_features(unfiltered_dataframe)
+        filtered_dataframe, _ = dk.filter_features(
+            unfiltered_dataframe, dk.training_features_list, training_filter=False
+        )
+        filtered_dataframe = dk.normalize_data_from_metadata(filtered_dataframe)
+        dk.data_dictionary["prediction_features"] = filtered_dataframe
+
+        self.data_cleaning_predict(dk, filtered_dataframe)
+
+        predictions = self.model.predict(dk.data_dictionary["prediction_features"])
+        pred_df = DataFrame(predictions, columns=dk.label_list)
+
+        predictions_prob = self.model.predict_proba(dk.data_dictionary["prediction_features"])
+        pred_df_prob = DataFrame(predictions_prob, columns=self.model.classes_)
+
+        pred_df = pd.concat([pred_df, pred_df_prob], axis=1)
+
+        return (pred_df, dk.do_predict)

freqtrade/freqai/prediction_models/BaseRegressionModel.py

@@ -0,0 +1,96 @@
+import logging
+from typing import Any, Tuple
+
+import numpy as np
+import numpy.typing as npt
+from pandas import DataFrame
+
+from freqtrade.freqai.data_kitchen import FreqaiDataKitchen
+from freqtrade.freqai.freqai_interface import IFreqaiModel
+
+
+logger = logging.getLogger(__name__)
+
+
+class BaseRegressionModel(IFreqaiModel):
+    """
+    Base class for regression type models (e.g. Catboost, LightGBM, XGboost etc.).
+    User *must* inherit from this class and set fit() and predict(). See example scripts
+    such as prediction_models/CatboostPredictionModel.py for guidance.
+    """
+
+    def train(
+        self, unfiltered_dataframe: DataFrame, pair: str, dk: FreqaiDataKitchen
+    ) -> Any:
+        """
+        Filter the training data and train a model to it. Train makes heavy use of the datakitchen
+        for storing, saving, loading, and analyzing the data.
+        :param unfiltered_dataframe: Full dataframe for the current training period
+        :param metadata: pair metadata from strategy.
+        :return:
+        :model: Trained model which can be used to inference (self.predict)
+        """
+
+        logger.info("-------------------- Starting training " f"{pair} --------------------")
+
+        # filter the features requested by user in the configuration file and elegantly handle NaNs
+        features_filtered, labels_filtered = dk.filter_features(
+            unfiltered_dataframe,
+            dk.training_features_list,
+            dk.label_list,
+            training_filter=True,
+        )
+
+        start_date = unfiltered_dataframe["date"].iloc[0].strftime("%Y-%m-%d")
+        end_date = unfiltered_dataframe["date"].iloc[-1].strftime("%Y-%m-%d")
+        logger.info(f"-------------------- Training on data from {start_date} to "
+                    f"{end_date}--------------------")
+        # split data into train/test data.
+        data_dictionary = dk.make_train_test_datasets(features_filtered, labels_filtered)
+        if not self.freqai_info.get('fit_live_predictions', 0) or not self.live:
+            dk.fit_labels()
+        # normalize all data based on train_dataset only
+        data_dictionary = dk.normalize_data(data_dictionary)
+
+        # optional additional data cleaning/analysis
+        self.data_cleaning_train(dk)
+
+        logger.info(
+            f'Training model on {len(dk.data_dictionary["train_features"].columns)}' " features"
+        )
+        logger.info(f'Training model on {len(data_dictionary["train_features"])} data points')
+
+        model = self.fit(data_dictionary)
+
+        logger.info(f"--------------------done training {pair}--------------------")
+
+        return model
+
+    def predict(
+        self, unfiltered_dataframe: DataFrame, dk: FreqaiDataKitchen, first: bool = False
+    ) -> Tuple[DataFrame, npt.NDArray[np.int_]]:
+        """
+        Filter the prediction features data and predict with it.
+        :param: unfiltered_dataframe: Full dataframe for the current backtest period.
+        :return:
+        :pred_df: dataframe containing the predictions
+        :do_predict: np.array of 1s and 0s to indicate places where freqai needed to remove
+        data (NaNs) or felt uncertain about data (PCA and DI index)
+        """
+
+        dk.find_features(unfiltered_dataframe)
+        filtered_dataframe, _ = dk.filter_features(
+            unfiltered_dataframe, dk.training_features_list, training_filter=False
+        )
+        filtered_dataframe = dk.normalize_data_from_metadata(filtered_dataframe)
+        dk.data_dictionary["prediction_features"] = filtered_dataframe
+
+        # optional additional data cleaning/analysis
+        self.data_cleaning_predict(dk, filtered_dataframe)
+
+        predictions = self.model.predict(dk.data_dictionary["prediction_features"])
+        pred_df = DataFrame(predictions, columns=dk.label_list)
+
+        pred_df = dk.denormalize_labels_from_metadata(pred_df)
+
+        return (pred_df, dk.do_predict)

freqtrade/freqai/prediction_models/BaseTensorFlowModel.py

@@ -0,0 +1,64 @@
+import logging
+from typing import Any
+
+from pandas import DataFrame
+
+from freqtrade.freqai.data_kitchen import FreqaiDataKitchen
+from freqtrade.freqai.freqai_interface import IFreqaiModel
+
+
+logger = logging.getLogger(__name__)
+
+
+class BaseTensorFlowModel(IFreqaiModel):
+    """
+    Base class for TensorFlow type models.
+    User *must* inherit from this class and set fit() and predict().
+    """
+
+    def train(
+        self, unfiltered_dataframe: DataFrame, pair: str, dk: FreqaiDataKitchen
+    ) -> Any:
+        """
+        Filter the training data and train a model to it. Train makes heavy use of the datakitchen
+        for storing, saving, loading, and analyzing the data.
+        :param unfiltered_dataframe: Full dataframe for the current training period
+        :param metadata: pair metadata from strategy.
+        :return:
+        :model: Trained model which can be used to inference (self.predict)
+        """
+
+        logger.info("-------------------- Starting training " f"{pair} --------------------")
+
+        # filter the features requested by user in the configuration file and elegantly handle NaNs
+        features_filtered, labels_filtered = dk.filter_features(
+            unfiltered_dataframe,
+            dk.training_features_list,
+            dk.label_list,
+            training_filter=True,
+        )
+
+        start_date = unfiltered_dataframe["date"].iloc[0].strftime("%Y-%m-%d")
+        end_date = unfiltered_dataframe["date"].iloc[-1].strftime("%Y-%m-%d")
+        logger.info(f"-------------------- Training on data from {start_date} to "
+                    f"{end_date}--------------------")
+        # split data into train/test data.
+        data_dictionary = dk.make_train_test_datasets(features_filtered, labels_filtered)
+        if not self.freqai_info.get('fit_live_predictions', 0) or not self.live:
+            dk.fit_labels()
+        # normalize all data based on train_dataset only
+        data_dictionary = dk.normalize_data(data_dictionary)
+
+        # optional additional data cleaning/analysis
+        self.data_cleaning_train(dk)
+
+        logger.info(
+            f'Training model on {len(dk.data_dictionary["train_features"].columns)}' " features"
+        )
+        logger.info(f'Training model on {len(data_dictionary["train_features"])} data points')
+
+        model = self.fit(data_dictionary)
+
+        logger.info(f"--------------------done training {pair}--------------------")
+
+        return model

freqtrade/freqai/prediction_models/CatboostClassifier.py

@@ -0,0 +1,41 @@
+import logging
+from typing import Any, Dict
+
+from catboost import CatBoostClassifier, Pool
+
+from freqtrade.freqai.prediction_models.BaseClassifierModel import BaseClassifierModel
+
+
+logger = logging.getLogger(__name__)
+
+
+class CatboostClassifier(BaseClassifierModel):
+    """
+    User created prediction model. The class needs to override three necessary
+    functions, predict(), train(), fit(). The class inherits ModelHandler which
+    has its own DataHandler where data is held, saved, loaded, and managed.
+    """
+
+    def fit(self, data_dictionary: Dict) -> Any:
+        """
+        User sets up the training and test data to fit their desired model here
+        :params:
+        :data_dictionary: the dictionary constructed by DataHandler to hold
+        all the training and test data/labels.
+        """
+
+        train_data = Pool(
+            data=data_dictionary["train_features"],
+            label=data_dictionary["train_labels"],
+            weight=data_dictionary["train_weights"],
+        )
+
+        cbr = CatBoostClassifier(
+            allow_writing_files=False,
+            loss_function='MultiClass',
+            **self.model_training_parameters,
+        )
+
+        cbr.fit(train_data)
+
+        return cbr

freqtrade/freqai/prediction_models/CatboostRegressor.py

@@ -0,0 +1,53 @@
+import gc
+import logging
+from typing import Any, Dict
+
+from catboost import CatBoostRegressor, Pool
+
+from freqtrade.freqai.prediction_models.BaseRegressionModel import BaseRegressionModel
+
+
+logger = logging.getLogger(__name__)
+
+
+class CatboostRegressor(BaseRegressionModel):
+    """
+    User created prediction model. The class needs to override three necessary
+    functions, predict(), train(), fit(). The class inherits ModelHandler which
+    has its own DataHandler where data is held, saved, loaded, and managed.
+    """
+
+    def fit(self, data_dictionary: Dict) -> Any:
+        """
+        User sets up the training and test data to fit their desired model here
+        :param data_dictionary: the dictionary constructed by DataHandler to hold
+                                all the training and test data/labels.
+        """
+
+        train_data = Pool(
+            data=data_dictionary["train_features"],
+            label=data_dictionary["train_labels"],
+            weight=data_dictionary["train_weights"],
+        )
+        if self.freqai_info.get('data_split_parameters', {}).get('test_size', 0.1) == 0:
+            test_data = None
+        else:
+            test_data = Pool(
+                data=data_dictionary["test_features"],
+                label=data_dictionary["test_labels"],
+                weight=data_dictionary["test_weights"],
+            )
+
+        model = CatBoostRegressor(
+            allow_writing_files=False,
+            **self.model_training_parameters,
+        )
+
+        model.fit(X=train_data, eval_set=test_data)
+
+        # some evidence that catboost pools have memory leaks:
+        # https://github.com/catboost/catboost/issues/1835
+        del train_data, test_data
+        gc.collect()
+
+        return model

freqtrade/freqai/prediction_models/CatboostRegressorMultiTarget.py

@@ -0,0 +1,44 @@
+import logging
+from typing import Any, Dict
+
+from catboost import CatBoostRegressor  # , Pool
+from sklearn.multioutput import MultiOutputRegressor
+
+from freqtrade.freqai.prediction_models.BaseRegressionModel import BaseRegressionModel
+
+
+logger = logging.getLogger(__name__)
+
+
+class CatboostRegressorMultiTarget(BaseRegressionModel):
+    """
+    User created prediction model. The class needs to override three necessary
+    functions, predict(), train(), fit(). The class inherits ModelHandler which
+    has its own DataHandler where data is held, saved, loaded, and managed.
+    """
+
+    def fit(self, data_dictionary: Dict) -> Any:
+        """
+        User sets up the training and test data to fit their desired model here
+        :param data_dictionary: the dictionary constructed by DataHandler to hold
+                                all the training and test data/labels.
+        """
+
+        cbr = CatBoostRegressor(
+            allow_writing_files=False,
+            **self.model_training_parameters,
+        )
+
+        X = data_dictionary["train_features"]
+        y = data_dictionary["train_labels"]
+        eval_set = (data_dictionary["test_features"], data_dictionary["test_labels"])
+        sample_weight = data_dictionary["train_weights"]
+
+        model = MultiOutputRegressor(estimator=cbr)
+        model.fit(X=X, y=y, sample_weight=sample_weight)  # , eval_set=eval_set)
+
+        if self.freqai_info.get('data_split_parameters', {}).get('test_size', 0.1) != 0:
+            train_score = model.score(X, y)
+            test_score = model.score(*eval_set)
+            logger.info(f"Train score {train_score}, Test score {test_score}")
+        return model

freqtrade/freqai/prediction_models/LightGBMClassifier.py

@@ -0,0 +1,38 @@
+import logging
+from typing import Any, Dict
+
+from lightgbm import LGBMClassifier
+
+from freqtrade.freqai.prediction_models.BaseClassifierModel import BaseClassifierModel
+
+
+logger = logging.getLogger(__name__)
+
+
+class LightGBMClassifier(BaseClassifierModel):
+    """
+    User created prediction model. The class needs to override three necessary
+    functions, predict(), train(), fit(). The class inherits ModelHandler which
+    has its own DataHandler where data is held, saved, loaded, and managed.
+    """
+
+    def fit(self, data_dictionary: Dict) -> Any:
+        """
+        User sets up the training and test data to fit their desired model here
+        :params:
+        :data_dictionary: the dictionary constructed by DataHandler to hold
+        all the training and test data/labels.
+        """
+
+        if self.freqai_info.get('data_split_parameters', {}).get('test_size', 0.1) == 0:
+            eval_set = None
+        else:
+            eval_set = (data_dictionary["test_features"], data_dictionary["test_labels"])
+        X = data_dictionary["train_features"]
+        y = data_dictionary["train_labels"]
+
+        model = LGBMClassifier(**self.model_training_parameters)
+
+        model.fit(X=X, y=y, eval_set=eval_set)
+
+        return model

freqtrade/freqai/prediction_models/LightGBMRegressor.py

@@ -0,0 +1,39 @@
+import logging
+from typing import Any, Dict
+
+from lightgbm import LGBMRegressor
+
+from freqtrade.freqai.prediction_models.BaseRegressionModel import BaseRegressionModel
+
+
+logger = logging.getLogger(__name__)
+
+
+class LightGBMRegressor(BaseRegressionModel):
+    """
+    User created prediction model. The class needs to override three necessary
+    functions, predict(), train(), fit(). The class inherits ModelHandler which
+    has its own DataHandler where data is held, saved, loaded, and managed.
+    """
+
+    def fit(self, data_dictionary: Dict) -> 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: the dictionary constructed by DataHandler to hold
+                                all the training and test data/labels.
+        """
+
+        if self.freqai_info.get('data_split_parameters', {}).get('test_size', 0.1) == 0:
+            eval_set = None
+        else:
+            eval_set = (data_dictionary["test_features"], data_dictionary["test_labels"])
+        X = data_dictionary["train_features"]
+        y = data_dictionary["train_labels"]
+
+        model = LGBMRegressor(**self.model_training_parameters)
+
+        model.fit(X=X, y=y, eval_set=eval_set)
+
+        return model

freqtrade/freqai/prediction_models/LightGBMRegressorMultiTarget.py

@@ -0,0 +1,39 @@
+import logging
+from typing import Any, Dict
+
+from lightgbm import LGBMRegressor
+from sklearn.multioutput import MultiOutputRegressor
+
+from freqtrade.freqai.prediction_models.BaseRegressionModel import BaseRegressionModel
+
+
+logger = logging.getLogger(__name__)
+
+
+class LightGBMRegressorMultiTarget(BaseRegressionModel):
+    """
+    User created prediction model. The class needs to override three necessary
+    functions, predict(), train(), fit(). The class inherits ModelHandler which
+    has its own DataHandler where data is held, saved, loaded, and managed.
+    """
+
+    def fit(self, data_dictionary: Dict) -> Any:
+        """
+        User sets up the training and test data to fit their desired model here
+        :param data_dictionary: the dictionary constructed by DataHandler to hold
+                                all the training and test data/labels.
+        """
+
+        lgb = LGBMRegressor(**self.model_training_parameters)
+
+        X = data_dictionary["train_features"]
+        y = data_dictionary["train_labels"]
+        eval_set = (data_dictionary["test_features"], data_dictionary["test_labels"])
+        sample_weight = data_dictionary["train_weights"]
+
+        model = MultiOutputRegressor(estimator=lgb)
+        model.fit(X=X, y=y, sample_weight=sample_weight)  # , eval_set=eval_set)
+        train_score = model.score(X, y)
+        test_score = model.score(*eval_set)
+        logger.info(f"Train score {train_score}, Test score {test_score}")
+        return model

freqtrade/freqtradebot.py

@@ -1489,12 +1489,6 @@ class FreqtradeBot(LoggingMixin):
                 ExitType.STOP_LOSS, ExitType.TRAILING_STOP_LOSS, ExitType.LIQUIDATION):
             exit_type = 'stoploss'
 
-        # if stoploss is on exchange and we are on dry_run mode,
-        # we consider the sell price stop price
-        if (self.config['dry_run'] and exit_type == 'stoploss'
-                and self.strategy.order_types['stoploss_on_exchange']):
-            limit = trade.stoploss_or_liquidation
-
         # set custom_exit_price if available
         proposed_limit_rate = limit
         current_profit = trade.calc_profit_ratio(limit)

freqtrade/optimize/backtesting.py

@@ -89,6 +89,9 @@ class Backtesting:
         self.dataprovider = DataProvider(self.config, self.exchange)
 
         if self.config.get('strategy_list'):
+            if self.config.get('freqai', {}).get('enabled', False):
+                raise OperationalException(
+                    "You can't use strategy_list and freqai at the same time.")
             for strat in list(self.config['strategy_list']):
                 stratconf = deepcopy(self.config)
                 stratconf['strategy'] = strat
@@ -207,6 +210,15 @@ class Backtesting:
         """
         self.progress.init_step(BacktestState.DATALOAD, 1)
 
+        if self.config.get('freqai', {}).get('enabled', False):
+            startup_candles = int(self.config.get('freqai', {}).get('startup_candles', 0))
+            if not startup_candles:
+                raise OperationalException('FreqAI backtesting module requires user set '
+                                           'startup_candles in config.')
+            self.required_startup += int(self.config.get('freqai', {}).get('startup_candles', 0))
+            logger.info(f'Increasing startup_candle_count for freqai to {self.required_startup}')
+            self.config['startup_candle_count'] = self.required_startup
+
         data = history.load_data(
             datadir=self.config['datadir'],
             pairs=self.pairlists.whitelist,

freqtrade/plugins/pairlist/pairlist_helpers.py

@@ -1,5 +1,5 @@
 import re
-from typing import List
+from typing import Any, Dict, List
 
 
 def expand_pairlist(wildcardpl: List[str], available_pairs: List[str],
@@ -40,3 +40,13 @@ def expand_pairlist(wildcardpl: List[str], available_pairs: List[str],
             except re.error as err:
                 raise ValueError(f"Wildcard error in {pair_wc}, {err}")
     return result
+
+
+def dynamic_expand_pairlist(config: Dict[str, Any], markets: List[str]) -> List[str]:
+    expanded_pairs = expand_pairlist(config['pairs'], markets)
+    if config.get('freqai', {}).get('enabled', False):
+        corr_pairlist = config['freqai']['feature_parameters']['include_corr_pairlist']
+        expanded_pairs += [pair for pair in corr_pairlist
+                           if pair not in config['pairs']]
+
+    return expanded_pairs

freqtrade/resolvers/freqaimodel_resolver.py

@@ -0,0 +1,57 @@
+# pragma pylint: disable=attribute-defined-outside-init
+
+"""
+This module load a custom model for freqai
+"""
+import logging
+from pathlib import Path
+from typing import Dict
+
+from freqtrade.constants import USERPATH_FREQAIMODELS
+from freqtrade.exceptions import OperationalException
+from freqtrade.freqai.freqai_interface import IFreqaiModel
+from freqtrade.resolvers import IResolver
+
+
+logger = logging.getLogger(__name__)
+
+
+class FreqaiModelResolver(IResolver):
+    """
+    This class contains all the logic to load custom hyperopt loss class
+    """
+
+    object_type = IFreqaiModel
+    object_type_str = "FreqaiModel"
+    user_subdir = USERPATH_FREQAIMODELS
+    initial_search_path = (
+        Path(__file__).parent.parent.joinpath("freqai/prediction_models").resolve()
+    )
+
+    @staticmethod
+    def load_freqaimodel(config: Dict) -> IFreqaiModel:
+        """
+        Load the custom class from config parameter
+        :param config: configuration dictionary
+        """
+        disallowed_models = ["BaseRegressionModel", "BaseTensorFlowModel"]
+
+        freqaimodel_name = config.get("freqaimodel")
+        if not freqaimodel_name:
+            raise OperationalException(
+                "No freqaimodel set. Please use `--freqaimodel` to "
+                "specify the FreqaiModel class to use.\n"
+            )
+        if freqaimodel_name in disallowed_models:
+            raise OperationalException(
+                f"{freqaimodel_name} is a baseclass and cannot be used directly. Please choose "
+                "an existing child class or inherit from this baseclass.\n"
+            )
+        freqaimodel = FreqaiModelResolver.load_object(
+            freqaimodel_name,
+            config,
+            kwargs={"config": config},
+            extra_dir=config.get("freqaimodel_path"),
+        )
+
+        return freqaimodel

freqtrade/rpc/telegram.py

@@ -120,7 +120,8 @@ class Telegram(RPCHandler):
                                  r'/daily$', r'/daily \d+$', r'/profit$', r'/profit \d+',
                                  r'/stats$', r'/count$', r'/locks$', r'/balance$',
                                  r'/stopbuy$', r'/reload_config$', r'/show_config$',
-                                 r'/logs$', r'/whitelist$', r'/blacklist$', r'/bl_delete$',
+                                 r'/logs$', r'/whitelist$', r'/whitelist(\ssorted|\sbaseonly)+$',
+                                 r'/blacklist$', r'/bl_delete$',
                                  r'/weekly$', r'/weekly \d+$', r'/monthly$', r'/monthly \d+$',
                                  r'/forcebuy$', r'/forcelong$', r'/forceshort$',
                                  r'/forcesell$', r'/forceexit$',
@@ -1368,6 +1369,12 @@ class Telegram(RPCHandler):
         try:
             whitelist = self._rpc._rpc_whitelist()
 
+            if context.args:
+                if "sorted" in context.args:
+                    whitelist['whitelist'] = sorted(whitelist['whitelist'])
+                if "baseonly" in context.args:
+                    whitelist['whitelist'] = [pair.split("/")[0] for pair in whitelist['whitelist']]
+
             message = f"Using whitelist `{whitelist['method']}` with {whitelist['length']} pairs\n"
             message += f"`{', '.join(whitelist['whitelist'])}`"
 
@@ -1487,7 +1494,8 @@ class Telegram(RPCHandler):
             "*/fx <trade_id>|all:* `Alias to /forceexit`\n"
             f"{force_enter_text if self._config.get('force_entry_enable', False) else ''}"
             "*/delete <trade_id>:* `Instantly delete the given trade in the database`\n"
-            "*/whitelist:* `Show current whitelist` \n"
+            "*/whitelist [sorted] [baseonly]:* `Show current whitelist. Optionally in "
+            "order and/or only displaying the base currency of each pairing.`\n"
             "*/blacklist [pair]:* `Show current blacklist, or adds one or more pairs "
             "to the blacklist.` \n"
             "*/blacklist_delete [pairs]| /bl_delete [pairs]:* "
@@ -1524,7 +1532,7 @@ class Telegram(RPCHandler):
             "*/weekly <n>:* `Shows statistics per week, over the last n weeks`\n"
             "*/monthly <n>:* `Shows statistics per month, over the last n months`\n"
             "*/stats:* `Shows Wins / losses by Sell reason as well as "
-            "Avg. holding durationsfor buys and sells.`\n"
+            "Avg. holding durations for buys and sells.`\n"
             "*/help:* `This help message`\n"
             "*/version:* `Show version`"
             )

freqtrade/strategy/interface.py

@@ -145,11 +145,29 @@ class IStrategy(ABC, HyperStrategyMixin):
                     informative_data.candle_type = config['candle_type_def']
                 self._ft_informative.append((informative_data, cls_method))
 
+    def load_freqAI_model(self) -> None:
+        if self.config.get('freqai', {}).get('enabled', False):
+            # Import here to avoid importing this if freqAI is disabled
+            from freqtrade.resolvers.freqaimodel_resolver import FreqaiModelResolver
+
+            self.freqai = FreqaiModelResolver.load_freqaimodel(self.config)
+            self.freqai_info = self.config["freqai"]
+        else:
+            # Gracious failures if freqAI is disabled but "start" is called.
+            class DummyClass():
+                def start(self, *args, **kwargs):
+                    raise OperationalException(
+                        'freqAI is not enabled. '
+                        'Please enable it in your config to use this strategy.')
+            self.freqai = DummyClass()  # type: ignore
+
     def ft_bot_start(self, **kwargs) -> None:
         """
         Strategy init - runs after dataprovider has been added.
         Must call bot_start()
         """
+        self.load_freqAI_model()
+
         strategy_safe_wrapper(self.bot_start)()
 
         self.ft_load_hyper_params(self.config.get('runmode') == RunMode.HYPEROPT)
@@ -557,6 +575,22 @@ class IStrategy(ABC, HyperStrategyMixin):
         """
         return None
 
+    def populate_any_indicators(self, pair: str, df: DataFrame, tf: str,
+                                informative: DataFrame = None,
+                                set_generalized_indicators: bool = False) -> DataFrame:
+        """
+        Function designed to automatically generate, name and merge features
+        from user indicated timeframes in the configuration file. User can add
+        additional features here, but must follow the naming convention.
+        This method is *only* used in FreqaiDataKitchen class and therefore
+        it is only called if FreqAI is active.
+        :param pair: pair to be used as informative
+        :param df: strategy dataframe which will receive merges from informatives
+        :param tf: timeframe of the dataframe which will modify the feature names
+        :param informative: the dataframe associated with the informative pair
+        """
+        return df
+
 ###
 # END - Intended to be overridden by strategy
 ###

freqtrade/templates/FreqaiExampleStrategy.py

@@ -0,0 +1,328 @@
+import logging
+from functools import reduce
+
+import pandas as pd
+import talib.abstract as ta
+from pandas import DataFrame
+from technical import qtpylib
+
+from freqtrade.exchange import timeframe_to_prev_date
+from freqtrade.persistence import Trade
+from freqtrade.strategy import DecimalParameter, IntParameter, IStrategy, merge_informative_pair
+
+
+logger = logging.getLogger(__name__)
+
+
+class FreqaiExampleStrategy(IStrategy):
+    """
+    Example strategy showing how the user connects their own
+    IFreqaiModel to the strategy. Namely, the user uses:
+    self.freqai.start(dataframe, metadata)
+
+    to make predictions on their data. populate_any_indicators() automatically
+    generates the variety of features indicated by the user in the
+    canonical freqtrade configuration file under config['freqai'].
+    """
+
+    minimal_roi = {"0": 0.1, "240": -1}
+
+    plot_config = {
+        "main_plot": {},
+        "subplots": {
+            "prediction": {"prediction": {"color": "blue"}},
+            "target_roi": {
+                "target_roi": {"color": "brown"},
+            },
+            "do_predict": {
+                "do_predict": {"color": "brown"},
+            },
+        },
+    }
+
+    process_only_new_candles = True
+    stoploss = -0.05
+    use_exit_signal = True
+    startup_candle_count: int = 300
+    can_short = False
+
+    linear_roi_offset = DecimalParameter(
+        0.00, 0.02, default=0.005, space="sell", optimize=False, load=True
+    )
+    max_roi_time_long = IntParameter(0, 800, default=400, space="sell", optimize=False, load=True)
+
+    def informative_pairs(self):
+        whitelist_pairs = self.dp.current_whitelist()
+        corr_pairs = self.config["freqai"]["feature_parameters"]["include_corr_pairlist"]
+        informative_pairs = []
+        for tf in self.config["freqai"]["feature_parameters"]["include_timeframes"]:
+            for pair in whitelist_pairs:
+                informative_pairs.append((pair, tf))
+            for pair in corr_pairs:
+                if pair in whitelist_pairs:
+                    continue  # avoid duplication
+                informative_pairs.append((pair, tf))
+        return informative_pairs
+
+    def populate_any_indicators(
+        self, pair, df, tf, informative=None, set_generalized_indicators=False
+    ):
+        """
+        Function designed to automatically generate, name and merge features
+        from user indicated timeframes in the configuration file. User controls the indicators
+        passed to the training/prediction by prepending indicators with `'%-' + coin `
+        (see convention below). I.e. user should not prepend any supporting metrics
+        (e.g. bb_lowerband below) with % unless they explicitly want to pass that metric to the
+        model.
+        :param pair: pair to be used as informative
+        :param df: strategy dataframe which will receive merges from informatives
+        :param tf: timeframe of the dataframe which will modify the feature names
+        :param informative: the dataframe associated with the informative pair
+        """
+
+        coin = pair.split('/')[0]
+
+        if informative is None:
+            informative = self.dp.get_pair_dataframe(pair, tf)
+
+        # first loop is automatically duplicating indicators for time periods
+        for t in self.freqai_info["feature_parameters"]["indicator_periods_candles"]:
+
+            t = int(t)
+            informative[f"%-{coin}rsi-period_{t}"] = ta.RSI(informative, timeperiod=t)
+            informative[f"%-{coin}mfi-period_{t}"] = ta.MFI(informative, timeperiod=t)
+            informative[f"%-{coin}adx-period_{t}"] = ta.ADX(informative, window=t)
+            informative[f"%-{coin}sma-period_{t}"] = ta.SMA(informative, timeperiod=t)
+            informative[f"%-{coin}ema-period_{t}"] = ta.EMA(informative, timeperiod=t)
+
+            informative[f"%-{coin}mfi-period_{t}"] = ta.MFI(informative, timeperiod=t)
+
+            bollinger = qtpylib.bollinger_bands(
+                qtpylib.typical_price(informative), window=t, stds=2.2
+            )
+            informative[f"{coin}bb_lowerband-period_{t}"] = bollinger["lower"]
+            informative[f"{coin}bb_middleband-period_{t}"] = bollinger["mid"]
+            informative[f"{coin}bb_upperband-period_{t}"] = bollinger["upper"]
+
+            informative[f"%-{coin}bb_width-period_{t}"] = (
+                informative[f"{coin}bb_upperband-period_{t}"]
+                - informative[f"{coin}bb_lowerband-period_{t}"]
+            ) / informative[f"{coin}bb_middleband-period_{t}"]
+            informative[f"%-{coin}close-bb_lower-period_{t}"] = (
+                informative["close"] / informative[f"{coin}bb_lowerband-period_{t}"]
+            )
+
+            informative[f"%-{coin}roc-period_{t}"] = ta.ROC(informative, timeperiod=t)
+
+            informative[f"%-{coin}relative_volume-period_{t}"] = (
+                informative["volume"] / informative["volume"].rolling(t).mean()
+            )
+
+        informative[f"%-{coin}pct-change"] = informative["close"].pct_change()
+        informative[f"%-{coin}raw_volume"] = informative["volume"]
+        informative[f"%-{coin}raw_price"] = informative["close"]
+
+        indicators = [col for col in informative if col.startswith("%")]
+        # This loop duplicates and shifts all indicators to add a sense of recency to data
+        for n in range(self.freqai_info["feature_parameters"]["include_shifted_candles"] + 1):
+            if n == 0:
+                continue
+            informative_shift = informative[indicators].shift(n)
+            informative_shift = informative_shift.add_suffix("_shift-" + str(n))
+            informative = pd.concat((informative, informative_shift), axis=1)
+
+        df = merge_informative_pair(df, informative, self.config["timeframe"], tf, ffill=True)
+        skip_columns = [
+            (s + "_" + tf) for s in ["date", "open", "high", "low", "close", "volume"]
+        ]
+        df = df.drop(columns=skip_columns)
+
+        # Add generalized indicators here (because in live, it will call this
+        # function to populate indicators during training). Notice how we ensure not to
+        # add them multiple times
+        if set_generalized_indicators:
+            df["%-day_of_week"] = (df["date"].dt.dayofweek + 1) / 7
+            df["%-hour_of_day"] = (df["date"].dt.hour + 1) / 25
+
+            # user adds targets here by prepending them with &- (see convention below)
+            df["&-s_close"] = (
+                df["close"]
+                .shift(-self.freqai_info["feature_parameters"]["label_period_candles"])
+                .rolling(self.freqai_info["feature_parameters"]["label_period_candles"])
+                .mean()
+                / df["close"]
+                - 1
+            )
+
+            # Classifiers are typically set up with strings as targets:
+            # df['&s-up_or_down'] = np.where( df["close"].shift(-100) >
+            #                                 df["close"], 'up', 'down')
+
+            # If user wishes to use multiple targets, they can add more by
+            # appending more columns with '&'. User should keep in mind that multi targets
+            # requires a multioutput prediction model such as
+            # templates/CatboostPredictionMultiModel.py,
+
+            # df["&-s_range"] = (
+            #     df["close"]
+            #     .shift(-self.freqai_info["feature_parameters"]["label_period_candles"])
+            #     .rolling(self.freqai_info["feature_parameters"]["label_period_candles"])
+            #     .max()
+            #     -
+            #     df["close"]
+            #     .shift(-self.freqai_info["feature_parameters"]["label_period_candles"])
+            #     .rolling(self.freqai_info["feature_parameters"]["label_period_candles"])
+            #     .min()
+            # )
+
+        return df
+
+    def populate_indicators(self, dataframe: DataFrame, metadata: dict) -> DataFrame:
+
+        # All indicators must be populated by populate_any_indicators() for live functionality
+        # to work correctly.
+
+        # the model will return all labels created by user in `populate_any_indicators`
+        # (& appended targets), an indication of whether or not the prediction should be accepted,
+        # the target mean/std values for each of the labels created by user in
+        # `populate_any_indicators()` for each training period.
+
+        dataframe = self.freqai.start(dataframe, metadata, self)
+
+        dataframe["target_roi"] = dataframe["&-s_close_mean"] + dataframe["&-s_close_std"] * 1.25
+        dataframe["sell_roi"] = dataframe["&-s_close_mean"] - dataframe["&-s_close_std"] * 1.25
+        return dataframe
+
+    def populate_entry_trend(self, df: DataFrame, metadata: dict) -> DataFrame:
+
+        enter_long_conditions = [df["do_predict"] == 1, df["&-s_close"] > df["target_roi"]]
+
+        if enter_long_conditions:
+            df.loc[
+                reduce(lambda x, y: x & y, enter_long_conditions), ["enter_long", "enter_tag"]
+            ] = (1, "long")
+
+        enter_short_conditions = [df["do_predict"] == 1, df["&-s_close"] < df["sell_roi"]]
+
+        if enter_short_conditions:
+            df.loc[
+                reduce(lambda x, y: x & y, enter_short_conditions), ["enter_short", "enter_tag"]
+            ] = (1, "short")
+
+        return df
+
+    def populate_exit_trend(self, df: DataFrame, metadata: dict) -> DataFrame:
+        exit_long_conditions = [df["do_predict"] == 1, df["&-s_close"] < df["sell_roi"] * 0.25]
+        if exit_long_conditions:
+            df.loc[reduce(lambda x, y: x & y, exit_long_conditions), "exit_long"] = 1
+
+        exit_short_conditions = [df["do_predict"] == 1, df["&-s_close"] > df["target_roi"] * 0.25]
+        if exit_short_conditions:
+            df.loc[reduce(lambda x, y: x & y, exit_short_conditions), "exit_short"] = 1
+
+        return df
+
+    def get_ticker_indicator(self):
+        return int(self.config["timeframe"][:-1])
+
+    def custom_exit(
+        self, pair: str, trade: Trade, current_time, current_rate, current_profit, **kwargs
+    ):
+
+        dataframe, _ = self.dp.get_analyzed_dataframe(pair=pair, timeframe=self.timeframe)
+
+        trade_date = timeframe_to_prev_date(self.config["timeframe"], trade.open_date_utc)
+        trade_candle = dataframe.loc[(dataframe["date"] == trade_date)]
+
+        if trade_candle.empty:
+            return None
+        trade_candle = trade_candle.squeeze()
+
+        follow_mode = self.config.get("freqai", {}).get("follow_mode", False)
+
+        if not follow_mode:
+            pair_dict = self.freqai.dd.pair_dict
+        else:
+            pair_dict = self.freqai.dd.follower_dict
+
+        entry_tag = trade.enter_tag
+
+        if (
+            "prediction" + entry_tag not in pair_dict[pair]
+            or pair_dict[pair]['extras']["prediction" + entry_tag] == 0
+        ):
+            pair_dict[pair]['extras']["prediction" + entry_tag] = abs(trade_candle["&-s_close"])
+            if not follow_mode:
+                self.freqai.dd.save_drawer_to_disk()
+            else:
+                self.freqai.dd.save_follower_dict_to_disk()
+
+        roi_price = pair_dict[pair]['extras']["prediction" + entry_tag]
+        roi_time = self.max_roi_time_long.value
+
+        roi_decay = roi_price * (
+            1 - ((current_time - trade.open_date_utc).seconds) / (roi_time * 60)
+        )
+        if roi_decay < 0:
+            roi_decay = self.linear_roi_offset.value
+        else:
+            roi_decay += self.linear_roi_offset.value
+
+        if current_profit > roi_decay:
+            return "roi_custom_win"
+
+        if current_profit < -roi_decay:
+            return "roi_custom_loss"
+
+    def confirm_trade_exit(
+        self,
+        pair: str,
+        trade: Trade,
+        order_type: str,
+        amount: float,
+        rate: float,
+        time_in_force: str,
+        exit_reason: str,
+        current_time,
+        **kwargs,
+    ) -> bool:
+
+        entry_tag = trade.enter_tag
+        follow_mode = self.config.get("freqai", {}).get("follow_mode", False)
+        if not follow_mode:
+            pair_dict = self.freqai.dd.pair_dict
+        else:
+            pair_dict = self.freqai.dd.follower_dict
+
+        pair_dict[pair]['extras']["prediction" + entry_tag] = 0
+        if not follow_mode:
+            self.freqai.dd.save_drawer_to_disk()
+        else:
+            self.freqai.dd.save_follower_dict_to_disk()
+
+        return True
+
+    def confirm_trade_entry(
+        self,
+        pair: str,
+        order_type: str,
+        amount: float,
+        rate: float,
+        time_in_force: str,
+        current_time,
+        entry_tag,
+        side: str,
+        **kwargs,
+    ) -> bool:
+
+        df, _ = self.dp.get_analyzed_dataframe(pair, self.timeframe)
+        last_candle = df.iloc[-1].squeeze()
+
+        if side == "long":
+            if rate > (last_candle["close"] * (1 + 0.0025)):
+                return False
+        else:
+            if rate < (last_candle["close"] * (1 - 0.0025)):
+                return False
+
+        return True