stable/freqtrade/edge/__init__.py

445 lines
18 KiB
Python
Raw Normal View History

2018-09-24 17:22:30 +00:00
# pragma pylint: disable=W0603
""" Edge positioning package """
2018-09-21 15:41:31 +00:00
import logging
2018-11-21 23:04:20 +00:00
from typing import Any, Dict, NamedTuple
2018-09-21 15:41:31 +00:00
import arrow
2018-09-23 02:51:53 +00:00
2018-09-27 10:23:46 +00:00
import numpy as np
import utils_find_1st as utf1st
2018-09-24 17:22:30 +00:00
from pandas import DataFrame
2018-09-21 15:41:31 +00:00
import freqtrade.optimize as optimize
2018-11-26 20:06:32 +00:00
from freqtrade import constants, OperationalException
2018-09-21 15:41:31 +00:00
from freqtrade.arguments import Arguments
from freqtrade.arguments import TimeRange
2018-09-21 15:41:31 +00:00
from freqtrade.strategy.interface import SellType
2018-09-23 02:51:53 +00:00
2018-09-21 15:41:31 +00:00
logger = logging.getLogger(__name__)
2018-09-24 17:22:30 +00:00
2018-11-21 23:04:20 +00:00
class PairInfo(NamedTuple):
stoploss: float
winrate: float
risk_reward_ratio: float
required_risk_reward: float
expectancy: float
nb_trades: int
avg_trade_duration: float
2018-09-21 15:41:31 +00:00
class Edge():
"""
Calculates Win Rate, Risk Reward Ratio, Expectancy
against historical data for a give set of markets and a strategy
it then adjusts stoploss and position size accordingly
and force it into the strategy
Author: https://github.com/mishaker
"""
2018-09-21 15:41:31 +00:00
config: Dict = {}
_cached_pairs: Dict[str, Any] = {} # Keeps a list of pairs
2018-09-21 15:41:31 +00:00
2018-11-14 11:53:20 +00:00
def __init__(self, config: Dict[str, Any], exchange, strategy) -> None:
2018-09-21 15:41:31 +00:00
self.config = config
self.exchange = exchange
self.strategy = strategy
2018-09-21 15:41:31 +00:00
self.ticker_interval = self.strategy.ticker_interval
self.tickerdata_to_dataframe = self.strategy.tickerdata_to_dataframe
self.get_timeframe = optimize.get_timeframe
2018-09-26 15:03:10 +00:00
self.advise_sell = self.strategy.advise_sell
self.advise_buy = self.strategy.advise_buy
2018-09-24 17:22:30 +00:00
self.edge_config = self.config.get('edge', {})
self._cached_pairs: Dict[str, Any] = {} # Keeps a list of pairs
self._final: list = []
2018-11-26 20:06:32 +00:00
# checking max_open_trades. it should be -1 as with Edge
# the number of trades is determined by position size
if self.config['max_open_trades'] != -1:
logger.critical('max_open_trades should be -1 in config !')
if self.config['stake_amount'] != constants.UNLIMITED_STAKE_AMOUNT:
raise OperationalException('Edge works only with unlimited stake amount')
self._capital_percentage: float = self.edge_config.get('capital_available_percentage')
self._allowed_risk: float = self.edge_config.get('allowed_risk')
self._since_number_of_days: int = self.edge_config.get('calculate_since_number_of_days', 14)
self._last_updated: int = 0 # Timestamp of pairs last updated time
2018-11-14 11:53:20 +00:00
self._refresh_pairs = True
self._stoploss_range_min = float(self.edge_config.get('stoploss_range_min', -0.01))
self._stoploss_range_max = float(self.edge_config.get('stoploss_range_max', -0.05))
self._stoploss_range_step = float(self.edge_config.get('stoploss_range_step', -0.001))
# calculating stoploss range
self._stoploss_range = np.arange(
self._stoploss_range_min,
self._stoploss_range_max,
self._stoploss_range_step
)
self._timerange: TimeRange = Arguments.parse_timerange("%s-" % arrow.now().shift(
days=-1 * self._since_number_of_days).format('YYYYMMDD'))
2018-09-21 15:41:31 +00:00
self.fee = self.exchange.get_fee()
def calculate(self) -> bool:
pairs = self.config['exchange']['pair_whitelist']
heartbeat = self.edge_config.get('process_throttle_secs')
2018-09-21 15:41:31 +00:00
if (self._last_updated > 0) and (
2018-09-27 10:23:46 +00:00
self._last_updated + heartbeat > arrow.utcnow().timestamp):
2018-09-21 15:41:31 +00:00
return False
2018-09-26 14:03:51 +00:00
data: Dict[str, Any] = {}
2018-09-21 15:41:31 +00:00
logger.info('Using stake_currency: %s ...', self.config['stake_currency'])
logger.info('Using local backtesting data (using whitelist in given config) ...')
data = optimize.load_data(
self.config['datadir'],
pairs=pairs,
ticker_interval=self.ticker_interval,
2018-11-14 11:53:20 +00:00
refresh_pairs=self._refresh_pairs,
2018-09-21 15:41:31 +00:00
exchange=self.exchange,
timerange=self._timerange
2018-09-21 15:41:31 +00:00
)
if not data:
# Reinitializing cached pairs
self._cached_pairs = {}
2018-09-21 15:41:31 +00:00
logger.critical("No data found. Edge is stopped ...")
2018-09-26 14:03:51 +00:00
return False
2018-09-24 17:22:30 +00:00
2018-09-21 15:41:31 +00:00
preprocessed = self.tickerdata_to_dataframe(data)
# Print timeframe
min_date, max_date = self.get_timeframe(preprocessed)
logger.info(
'Measuring data from %s up to %s (%s days) ...',
min_date.isoformat(),
max_date.isoformat(),
(max_date - min_date).days
)
headers = ['date', 'buy', 'open', 'close', 'sell', 'high', 'low']
2018-09-26 13:20:53 +00:00
trades: list = []
2018-09-21 15:41:31 +00:00
for pair, pair_data in preprocessed.items():
2018-09-24 17:22:30 +00:00
# Sorting dataframe by date and reset index
2018-09-21 15:41:31 +00:00
pair_data = pair_data.sort_values(by=['date'])
pair_data = pair_data.reset_index(drop=True)
2018-09-26 15:09:20 +00:00
2018-09-26 15:03:10 +00:00
ticker_data = self.advise_sell(
self.advise_buy(pair_data, {'pair': pair}), {'pair': pair})[headers].copy()
2018-09-21 15:41:31 +00:00
trades += self._find_trades_for_stoploss_range(ticker_data, pair, self._stoploss_range)
2018-09-21 15:41:31 +00:00
2018-10-05 15:06:17 +00:00
# If no trade found then exit
if len(trades) == 0:
return False
2018-09-24 17:22:30 +00:00
2018-10-05 15:06:17 +00:00
# Fill missing, calculable columns, profit, duration , abs etc.
trades_df = self._fill_calculable_fields(DataFrame(trades))
2018-09-26 13:20:53 +00:00
self._cached_pairs = self._process_expectancy(trades_df)
2018-09-21 15:41:31 +00:00
self._last_updated = arrow.utcnow().timestamp
# Not a nice hack but probably simplest solution:
# When backtest load data it loads the delta between disk and exchange
2018-09-28 12:28:05 +00:00
# The problem is that exchange consider that recent.
# it is but it is incomplete (c.f. _async_get_candle_history)
# So it causes get_signal to exit cause incomplete ticker_hist
2018-09-28 12:28:05 +00:00
# A patch to that would be update _pairs_last_refresh_time of exchange
# so it will download again all pairs
# Another solution is to add new data to klines instead of reassigning it:
# self.klines[pair].update(data) instead of self.klines[pair] = data in exchange package.
# But that means indexing timestamp and having a verification so that
# there is no empty range between two timestaps (recently added and last
# one)
self.exchange._pairs_last_refresh_time = {}
2018-09-21 15:41:31 +00:00
return True
2018-09-24 17:22:30 +00:00
2018-11-28 14:36:32 +00:00
def stake_amount(self, pair: str, free_capital: float, total_capital: float) -> float:
if pair not in self._cached_pairs:
logger.warning("cannot find %s in calculated pairs, "
"stake_amount of strategy is used instead.", pair)
2018-11-30 16:59:51 +00:00
return self.config['stake_amount']
stoploss = self._cached_pairs[pair].stoploss
2018-11-28 14:36:32 +00:00
available_capital = total_capital * self._capital_percentage
2018-11-27 13:02:34 +00:00
allowed_capital_at_risk = round(available_capital * self._allowed_risk, 15)
2018-11-28 14:36:32 +00:00
max_position_size = abs(round((allowed_capital_at_risk / stoploss), 15))
position_size = min(max_position_size, free_capital)
logger.info(
2018-11-29 11:24:04 +00:00
'winrate: %s, expectancy: %s, position size: %s, pair: %s,'
' stoploss: %s, available capital: %s.',
self._cached_pairs[pair].winrate,
self._cached_pairs[pair].expectancy,
2018-11-28 14:36:32 +00:00
position_size, pair, stoploss, available_capital
)
2018-09-26 13:20:53 +00:00
return position_size
def stoploss(self, pair: str) -> float:
if pair in self._cached_pairs:
return self._cached_pairs[pair].stoploss
else:
logger.warning('tried to access stoploss of a non-existing pair, '
'strategy stoploss is returned instead.')
return self.strategy.stoploss
2018-09-26 13:20:53 +00:00
def adjust(self, pairs) -> list:
"""
Filters out and sorts "pairs" according to Edge calculated pairs
"""
final = []
for pair, info in self._cached_pairs.items():
if info.expectancy > float(self.edge_config.get('minimum_expectancy', 0.2)) and \
info.winrate > float(self.edge_config.get('minimum_winrate', 0.60)) and \
pair in pairs:
2018-11-04 17:43:57 +00:00
final.append(pair)
if self._final != final:
self._final = final
if self._final:
logger.info('Edge validated only %s', self._final)
else:
logger.info('Edge removed all pairs as no pair with minimum expectancy was found !')
return self._final
2018-09-24 17:22:30 +00:00
2018-10-02 09:55:14 +00:00
def _fill_calculable_fields(self, result: DataFrame) -> DataFrame:
2018-09-21 15:41:31 +00:00
"""
2018-09-26 13:20:53 +00:00
The result frame contains a number of columns that are calculable
2018-10-02 09:53:59 +00:00
from other columns. These are left blank till all rows are added,
2018-09-21 15:41:31 +00:00
to be populated in single vector calls.
Columns to be populated are:
- Profit
- trade duration
- profit abs
2018-09-26 13:20:53 +00:00
:param result Dataframe
:return: result Dataframe
2018-09-21 15:41:31 +00:00
"""
# stake and fees
# stake = 0.015
# 0.05% is 0.0005
# fee = 0.001
2018-11-27 13:02:34 +00:00
stake = 0.015
2018-09-21 15:41:31 +00:00
fee = self.fee
2018-10-05 15:06:17 +00:00
2018-09-21 15:41:31 +00:00
open_fee = fee / 2
close_fee = fee / 2
2018-09-26 13:20:53 +00:00
result['trade_duration'] = result['close_time'] - result['open_time']
2018-09-21 15:41:31 +00:00
2018-09-27 10:23:46 +00:00
result['trade_duration'] = result['trade_duration'].map(
lambda x: int(x.total_seconds() / 60))
2018-09-26 14:50:17 +00:00
# Spends, Takes, Profit, Absolute Profit
2018-09-26 13:20:53 +00:00
2018-09-21 15:41:31 +00:00
# Buy Price
2018-09-26 13:20:53 +00:00
result['buy_vol'] = stake / result['open_rate'] # How many target are we buying
result['buy_fee'] = stake * open_fee
result['buy_spend'] = stake + result['buy_fee'] # How much we're spending
2018-09-21 15:41:31 +00:00
# Sell price
2018-09-26 13:20:53 +00:00
result['sell_sum'] = result['buy_vol'] * result['close_rate']
result['sell_fee'] = result['sell_sum'] * close_fee
result['sell_take'] = result['sell_sum'] - result['sell_fee']
2018-09-21 15:41:31 +00:00
# profit_percent
2018-09-27 10:23:46 +00:00
result['profit_percent'] = (result['sell_take'] - result['buy_spend']) / result['buy_spend']
2018-09-26 14:50:17 +00:00
2018-09-21 15:41:31 +00:00
# Absolute profit
2018-09-26 13:20:53 +00:00
result['profit_abs'] = result['sell_take'] - result['buy_spend']
2018-09-21 15:41:31 +00:00
2018-09-26 13:20:53 +00:00
return result
2018-09-21 15:41:31 +00:00
def _process_expectancy(self, results: DataFrame) -> Dict[str, Any]:
2018-09-21 15:41:31 +00:00
"""
This calculates WinRate, Required Risk Reward, Risk Reward and Expectancy of all pairs
2018-09-24 17:22:30 +00:00
The calulation will be done per pair and per strategy.
2018-09-21 15:41:31 +00:00
"""
# Removing pairs having less than min_trades_number
min_trades_number = self.edge_config.get('min_trade_number', 10)
results = results.groupby(['pair', 'stoploss']).filter(lambda x: len(x) > min_trades_number)
2018-09-21 15:41:31 +00:00
###################################
2018-09-24 17:22:30 +00:00
# Removing outliers (Only Pumps) from the dataset
2018-09-21 15:41:31 +00:00
# The method to detect outliers is to calculate standard deviation
# Then every value more than (standard deviation + 2*average) is out (pump)
#
# Removing Pumps
if self.edge_config.get('remove_pumps', False):
results = results.groupby(['pair', 'stoploss']).apply(
lambda x: x[x['profit_abs'] < 2 * x['profit_abs'].std() + x['profit_abs'].mean()])
2018-09-21 15:41:31 +00:00
##########################################################################
# Removing trades having a duration more than X minutes (set in config)
max_trade_duration = self.edge_config.get('max_trade_duration_minute', 1440)
2018-09-21 15:41:31 +00:00
results = results[results.trade_duration < max_trade_duration]
#######################################################################
2018-10-18 09:09:10 +00:00
if results.empty:
return {}
2018-10-18 09:09:10 +00:00
groupby_aggregator = {
'profit_abs': [
('nb_trades', 'count'), # number of all trades
('profit_sum', lambda x: x[x > 0].sum()), # cumulative profit of all winning trades
('loss_sum', lambda x: abs(x[x < 0].sum())), # cumulative loss of all losing trades
('nb_win_trades', lambda x: x[x > 0].count()) # number of winning trades
],
'trade_duration': [('avg_trade_duration', 'mean')]
}
# Group by (pair and stoploss) by applying above aggregator
df = results.groupby(['pair', 'stoploss'])['profit_abs', 'trade_duration'].agg(
groupby_aggregator).reset_index(col_level=1)
# Dropping level 0 as we don't need it
df.columns = df.columns.droplevel(0)
# Calculating number of losing trades, average win and average loss
df['nb_loss_trades'] = df['nb_trades'] - df['nb_win_trades']
df['average_win'] = df['profit_sum'] / df['nb_win_trades']
df['average_loss'] = df['loss_sum'] / df['nb_loss_trades']
# Win rate = number of profitable trades / number of trades
df['winrate'] = df['nb_win_trades'] / df['nb_trades']
# risk_reward_ratio = average win / average loss
df['risk_reward_ratio'] = df['average_win'] / df['average_loss']
# required_risk_reward = (1 / winrate) - 1
df['required_risk_reward'] = (1 / df['winrate']) - 1
# expectancy = (risk_reward_ratio * winrate) - (lossrate)
df['expectancy'] = (df['risk_reward_ratio'] * df['winrate']) - (1 - df['winrate'])
# sort by expectancy and stoploss
df = df.sort_values(by=['expectancy', 'stoploss'], ascending=False).groupby(
2018-10-18 09:09:10 +00:00
'pair').first().sort_values(by=['expectancy'], ascending=False).reset_index()
final = {}
for x in df.itertuples():
2018-11-21 23:04:20 +00:00
final[x.pair] = PairInfo(
x.stoploss,
x.winrate,
x.risk_reward_ratio,
x.required_risk_reward,
x.expectancy,
x.nb_trades,
x.avg_trade_duration
)
2018-09-24 17:22:30 +00:00
# Returning a list of pairs in order of "expectancy"
return final
2018-09-21 15:41:31 +00:00
2018-09-26 13:20:53 +00:00
def _find_trades_for_stoploss_range(self, ticker_data, pair, stoploss_range):
buy_column = ticker_data['buy'].values
sell_column = ticker_data['sell'].values
date_column = ticker_data['date'].values
ohlc_columns = ticker_data[['open', 'high', 'low', 'close']].values
2018-09-26 14:50:17 +00:00
2018-09-26 13:20:53 +00:00
result: list = []
for stoploss in stoploss_range:
result += self._detect_next_stop_or_sell_point(
2018-09-26 14:50:17 +00:00
buy_column, sell_column, date_column, ohlc_columns, round(stoploss, 6), pair
2018-09-27 10:23:46 +00:00
)
2018-09-26 13:20:53 +00:00
return result
def _detect_next_stop_or_sell_point(self, buy_column, sell_column, date_column,
ohlc_columns, stoploss, pair, start_point=0):
"""
Iterate through ohlc_columns recursively in order to find the next trade
Next trade opens from the first buy signal noticed to
The sell or stoploss signal after it.
It then calls itself cutting OHLC, buy_column, sell_colum and date_column
Cut from (the exit trade index) + 1
Author: https://github.com/mishaker
"""
2018-09-26 14:50:17 +00:00
2018-09-26 13:20:53 +00:00
result: list = []
open_trade_index = utf1st.find_1st(buy_column, 1, utf1st.cmp_equal)
# return empty if we don't find trade entry (i.e. buy==1) or
# we find a buy but at the of array
if open_trade_index == -1 or open_trade_index == len(buy_column) - 1:
2018-09-26 13:20:53 +00:00
return []
else:
open_trade_index += 1 # when a buy signal is seen,
# trade opens in reality on the next candle
2018-09-26 13:20:53 +00:00
stop_price_percentage = stoploss + 1
open_price = ohlc_columns[open_trade_index, 0]
2018-09-26 13:20:53 +00:00
stop_price = (open_price * stop_price_percentage)
# Searching for the index where stoploss is hit
2018-09-27 10:23:46 +00:00
stop_index = utf1st.find_1st(
ohlc_columns[open_trade_index:, 2], stop_price, utf1st.cmp_smaller)
2018-09-26 13:20:53 +00:00
# If we don't find it then we assume stop_index will be far in future (infinite number)
if stop_index == -1:
stop_index = float('inf')
# Searching for the index where sell is hit
sell_index = utf1st.find_1st(sell_column[open_trade_index:], 1, utf1st.cmp_equal)
2018-09-26 13:20:53 +00:00
# If we don't find it then we assume sell_index will be far in future (infinite number)
if sell_index == -1:
sell_index = float('inf')
# Check if we don't find any stop or sell point (in that case trade remains open)
# It is not interesting for Edge to consider it so we simply ignore the trade
# And stop iterating there is no more entry
2018-09-26 13:20:53 +00:00
if stop_index == sell_index == float('inf'):
return []
if stop_index <= sell_index:
exit_index = open_trade_index + stop_index
2018-09-26 13:20:53 +00:00
exit_type = SellType.STOP_LOSS
exit_price = stop_price
elif stop_index > sell_index:
# if exit is SELL then we exit at the next candle
2018-09-26 13:20:53 +00:00
exit_index = open_trade_index + sell_index + 1
# check if we have the next candle
if len(ohlc_columns) - 1 < exit_index:
return []
2018-09-26 13:20:53 +00:00
exit_type = SellType.SELL_SIGNAL
exit_price = ohlc_columns[exit_index, 0]
2018-09-26 13:20:53 +00:00
2018-09-26 14:03:51 +00:00
trade = {'pair': pair,
'stoploss': stoploss,
'profit_percent': '',
'profit_abs': '',
'open_time': date_column[open_trade_index],
2018-09-26 14:03:51 +00:00
'close_time': date_column[exit_index],
'open_index': start_point + open_trade_index,
2018-09-26 14:03:51 +00:00
'close_index': start_point + exit_index,
'trade_duration': '',
'open_rate': round(open_price, 15),
'close_rate': round(exit_price, 15),
'exit_type': exit_type
}
2018-09-26 13:20:53 +00:00
result.append(trade)
# Calling again the same function recursively but giving
# it a view of exit_index till the end of array
return result + self._detect_next_stop_or_sell_point(
2018-09-26 13:20:53 +00:00
buy_column[exit_index:],
sell_column[exit_index:],
date_column[exit_index:],
ohlc_columns[exit_index:],
stoploss,
pair,
(start_point + exit_index)
)