adelorenzo
/
stable
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Merge pull request #3845 from freqtrade/feat/backtest_speedup_serialize 

Backtesting should not double-loop for sell signals
This commit is contained in:
Matthias 2020-10-19 07:52:33 +02:00 committed by GitHub
parent 340f25bd42 cf2ae788d7
commit 667f1b8b8c
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 135 additions and 124 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

231
freqtrade/optimize/backtesting.py
View File

@ -4,11 +4,11 @@
This module contains the backtesting logic
"""
import logging
from collections import defaultdict
from copy import deepcopy
from datetime import datetime, timedelta
from typing import Any, Dict, List, NamedTuple, Optional, Tuple
import arrow
from pandas import DataFrame
from freqtrade.configuration import TimeRange, remove_credentials, validate_config_consistency
@ -28,6 +28,15 @@ from freqtrade.strategy.interface import IStrategy, SellCheckTuple, SellType
logger = logging.getLogger(__name__)
# Indexes for backtest tuples
DATE_IDX = 0
BUY_IDX = 1
OPEN_IDX = 2
CLOSE_IDX = 3
SELL_IDX = 4
LOW_IDX = 5
HIGH_IDX = 6
class BacktestResult(NamedTuple):
"""
@ -115,7 +124,7 @@ class Backtesting:
"""
Load strategy into backtesting
"""
self.strategy = strategy
self.strategy: IStrategy = strategy
# Set stoploss_on_exchange to false for backtesting,
# since a "perfect" stoploss-sell is assumed anyway
# And the regular "stoploss" function would not apply to that case
@ -147,12 +156,14 @@ class Backtesting:
return data, timerange
def _get_ohlcv_as_lists(self, processed: Dict) -> Dict[str, DataFrame]:
def _get_ohlcv_as_lists(self, processed: Dict[str, DataFrame]) -> Dict[str, Tuple]:
"""
Helper function to convert a processed dataframes into lists for performance reasons.
Used by backtest() - so keep this optimized for performance.
"""
# Every change to this headers list must evaluate further usages of the resulting tuple
# and eventually change the constants for indexes at the top
headers = ['date', 'buy', 'open', 'close', 'sell', 'low', 'high']
data: Dict = {}
# Create dict with data
@ -172,10 +183,10 @@ class Backtesting:
# Convert from Pandas to list for performance reasons
# (Looping Pandas is slow.)
data[pair] = [x for x in df_analyzed.itertuples()]
data[pair] = [x for x in df_analyzed.itertuples(index=False, name=None)]
return data
def _get_close_rate(self, sell_row, trade: Trade, sell: SellCheckTuple,
def _get_close_rate(self, sell_row: Tuple, trade: Trade, sell: SellCheckTuple,
trade_dur: int) -> float:
"""
Get close rate for backtesting result
@ -186,12 +197,12 @@ class Backtesting:
return trade.stop_loss
elif sell.sell_type == (SellType.ROI):
roi_entry, roi = self.strategy.min_roi_reached_entry(trade_dur)
if roi is not None:
if roi is not None and roi_entry is not None:
if roi == -1 and roi_entry % self.timeframe_min == 0:
# When forceselling with ROI=-1, the roi time will always be equal to trade_dur.
# If that entry is a multiple of the timeframe (so on candle open)
# - we'll use open instead of close
return sell_row.open
return sell_row[OPEN_IDX]
# - (Expected abs profit + open_rate + open_fee) / (fee_close -1)
close_rate = - (trade.open_rate * roi + trade.open_rate *
@ -199,91 +210,79 @@ class Backtesting:
if (trade_dur > 0 and trade_dur == roi_entry
and roi_entry % self.timeframe_min == 0
and sell_row.open > close_rate):
and sell_row[OPEN_IDX] > close_rate):
# new ROI entry came into effect.
# use Open rate if open_rate > calculated sell rate
return sell_row.open
return sell_row[OPEN_IDX]
# Use the maximum between close_rate and low as we
# cannot sell outside of a candle.
# Applies when a new ROI setting comes in place and the whole candle is above that.
return max(close_rate, sell_row.low)
return max(close_rate, sell_row[LOW_IDX])
else:
# This should not be reached...
return sell_row.open
return sell_row[OPEN_IDX]
else:
return sell_row.open
return sell_row[OPEN_IDX]
def _get_sell_trade_entry(
self, pair: str, buy_row: DataFrame,
partial_ohlcv: List, trade_count_lock: Dict,
stake_amount: float, max_open_trades: int) -> Optional[BacktestResult]:
def _get_sell_trade_entry(self, trade: Trade, sell_row: Tuple) -> Optional[BacktestResult]:
trade = Trade(
pair=pair,
open_rate=buy_row.open,
open_date=buy_row.date,
stake_amount=stake_amount,
amount=round(stake_amount / buy_row.open, 8),
fee_open=self.fee,
fee_close=self.fee,
is_open=True,
)
logger.debug(f"{pair} - Backtesting emulates creation of new trade: {trade}.")
# calculate win/lose forwards from buy point
for sell_row in partial_ohlcv:
if max_open_trades > 0:
# Increase trade_count_lock for every iteration
trade_count_lock[sell_row.date] = trade_count_lock.get(sell_row.date, 0) + 1
sell = self.strategy.should_sell(trade, sell_row[OPEN_IDX], sell_row[DATE_IDX],
sell_row[BUY_IDX], sell_row[SELL_IDX],
low=sell_row[LOW_IDX], high=sell_row[HIGH_IDX])
if sell.sell_flag:
trade_dur = int((sell_row[DATE_IDX] - trade.open_date).total_seconds() // 60)
closerate = self._get_close_rate(sell_row, trade, sell, trade_dur)
sell = self.strategy.should_sell(trade, sell_row.open, sell_row.date, sell_row.buy,
sell_row.sell, low=sell_row.low, high=sell_row.high)
if sell.sell_flag:
trade_dur = int((sell_row.date - buy_row.date).total_seconds() // 60)
closerate = self._get_close_rate(sell_row, trade, sell, trade_dur)
return BacktestResult(pair=pair,
profit_percent=trade.calc_profit_ratio(rate=closerate),
profit_abs=trade.calc_profit(rate=closerate),
open_date=buy_row.date,
open_rate=buy_row.open,
open_fee=self.fee,
close_date=sell_row.date,
close_rate=closerate,
close_fee=self.fee,
amount=trade.amount,
trade_duration=trade_dur,
open_at_end=False,
sell_reason=sell.sell_type
)
if partial_ohlcv:
# no sell condition found - trade stil open at end of backtest period
sell_row = partial_ohlcv[-1]
bt_res = BacktestResult(pair=pair,
profit_percent=trade.calc_profit_ratio(rate=sell_row.open),
profit_abs=trade.calc_profit(rate=sell_row.open),
open_date=buy_row.date,
open_rate=buy_row.open,
open_fee=self.fee,
close_date=sell_row.date,
close_rate=sell_row.open,
close_fee=self.fee,
amount=trade.amount,
trade_duration=int((
sell_row.date - buy_row.date).total_seconds() // 60),
open_at_end=True,
sell_reason=SellType.FORCE_SELL
)
logger.debug(f"{pair} - Force selling still open trade, "
f"profit percent: {bt_res.profit_percent}, "
f"profit abs: {bt_res.profit_abs}")
return bt_res
return BacktestResult(pair=trade.pair,
profit_percent=trade.calc_profit_ratio(rate=closerate),
profit_abs=trade.calc_profit(rate=closerate),
open_date=trade.open_date,
open_rate=trade.open_rate,
open_fee=self.fee,
close_date=sell_row[DATE_IDX],
close_rate=closerate,
close_fee=self.fee,
amount=trade.amount,
trade_duration=trade_dur,
open_at_end=False,
sell_reason=sell.sell_type
)
return None
def handle_left_open(self, open_trades: Dict[str, List[Trade]],
data: Dict[str, List[Tuple]]) -> List[BacktestResult]:
"""
Handling of left open trades at the end of backtesting
"""
trades = []
for pair in open_trades.keys():
if len(open_trades[pair]) > 0:
for trade in open_trades[pair]:
sell_row = data[pair][-1]
trade_entry = BacktestResult(pair=trade.pair,
profit_percent=trade.calc_profit_ratio(
rate=sell_row[OPEN_IDX]),
profit_abs=trade.calc_profit(sell_row[OPEN_IDX]),
open_date=trade.open_date,
open_rate=trade.open_rate,
open_fee=self.fee,
close_date=sell_row[DATE_IDX],
close_rate=sell_row[OPEN_IDX],
close_fee=self.fee,
amount=trade.amount,
trade_duration=int((
sell_row[DATE_IDX] - trade.open_date
).total_seconds() // 60),
open_at_end=True,
sell_reason=SellType.FORCE_SELL
)
trades.append(trade_entry)
return trades
def backtest(self, processed: Dict, stake_amount: float,
start_date: arrow.Arrow, end_date: arrow.Arrow,
start_date: datetime, end_date: datetime,
max_open_trades: int = 0, position_stacking: bool = False) -> DataFrame:
"""
Implement backtesting functionality
@ -305,19 +304,21 @@ class Backtesting:
f"max_open_trades: {max_open_trades}, position_stacking: {position_stacking}"
)
trades = []
trade_count_lock: Dict = {}
# Use dict of lists with data for performance
# (looping lists is a lot faster than pandas DataFrames)
data: Dict = self._get_ohlcv_as_lists(processed)
lock_pair_until: Dict = {}
# Indexes per pair, so some pairs are allowed to have a missing start.
indexes: Dict = {}
tmp = start_date + timedelta(minutes=self.timeframe_min)
open_trades: Dict[str, List] = defaultdict(list)
open_trade_count = 0
# Loop timerange and get candle for each pair at that point in time
while tmp < end_date:
while tmp <= end_date:
open_trade_count_start = open_trade_count
for i, pair in enumerate(data):
if pair not in indexes:
@ -331,42 +332,52 @@ class Backtesting:
continue
# Waits until the time-counter reaches the start of the data for this pair.
if row.date > tmp.datetime:
if row[DATE_IDX] > tmp:
continue
indexes[pair] += 1
if row.buy == 0 or row.sell == 1:
continue # skip rows where no buy signal or that would immediately sell off
# without positionstacking, we can only have one open trade per pair.
# max_open_trades must be respected
# don't open on the last row
if ((position_stacking or len(open_trades[pair]) == 0)
and max_open_trades > 0 and open_trade_count_start < max_open_trades
and tmp != end_date
and row[BUY_IDX] == 1 and row[SELL_IDX] != 1):
# Enter trade
trade = Trade(
pair=pair,
open_rate=row[OPEN_IDX],
open_date=row[DATE_IDX],
stake_amount=stake_amount,
amount=round(stake_amount / row[OPEN_IDX], 8),
fee_open=self.fee,
fee_close=self.fee,
is_open=True,
)
# TODO: hacky workaround to avoid opening > max_open_trades
# This emulates previous behaviour - not sure if this is correct
# Prevents buying if the trade-slot was freed in this candle
open_trade_count_start += 1
open_trade_count += 1
# logger.debug(f"{pair} - Backtesting emulates creation of new trade: {trade}.")
open_trades[pair].append(trade)
if (not position_stacking and pair in lock_pair_until
and row.date <= lock_pair_until[pair]):
# without positionstacking, we can only have one open trade per pair.
continue
if max_open_trades > 0:
# Check if max_open_trades has already been reached for the given date
if not trade_count_lock.get(row.date, 0) < max_open_trades:
continue
trade_count_lock[row.date] = trade_count_lock.get(row.date, 0) + 1
# since indexes has been incremented before, we need to go one step back to
# also check the buying candle for sell conditions.
trade_entry = self._get_sell_trade_entry(pair, row, data[pair][indexes[pair]-1:],
trade_count_lock, stake_amount,
max_open_trades)
if trade_entry:
logger.debug(f"{pair} - Locking pair till "
f"close_date={trade_entry.close_date}")
lock_pair_until[pair] = trade_entry.close_date
trades.append(trade_entry)
else:
# Set lock_pair_until to end of testing period if trade could not be closed
lock_pair_until[pair] = end_date.datetime
for trade in open_trades[pair]:
# since indexes has been incremented before, we need to go one step back to
# also check the buying candle for sell conditions.
trade_entry = self._get_sell_trade_entry(trade, row)
# Sell occured
if trade_entry:
# logger.debug(f"{pair} - Backtesting sell {trade}")
open_trade_count -= 1
open_trades[pair].remove(trade)
trades.append(trade_entry)
# Move time one configured time_interval ahead.
tmp += timedelta(minutes=self.timeframe_min)
trades += self.handle_left_open(open_trades, data=data)
return DataFrame.from_records(trades, columns=BacktestResult._fields)
def start(self) -> None:
@ -412,8 +423,8 @@ class Backtesting:
results = self.backtest(
processed=preprocessed,
stake_amount=self.config['stake_amount'],
start_date=min_date,
end_date=max_date,
start_date=min_date.datetime,
end_date=max_date.datetime,
max_open_trades=max_open_trades,
position_stacking=position_stacking,
)

28
freqtrade/optimize/hyperopt.py
View File

@ -94,14 +94,14 @@ class Hyperopt:
# Populate functions here (hasattr is slow so should not be run during "regular" operations)
if hasattr(self.custom_hyperopt, 'populate_indicators'):
self.backtesting.strategy.advise_indicators = \
self.custom_hyperopt.populate_indicators # type: ignore
self.backtesting.strategy.advise_indicators = ( # type: ignore
self.custom_hyperopt.populate_indicators) # type: ignore
if hasattr(self.custom_hyperopt, 'populate_buy_trend'):
self.backtesting.strategy.advise_buy = \
self.custom_hyperopt.populate_buy_trend # type: ignore
self.backtesting.strategy.advise_buy = ( # type: ignore
self.custom_hyperopt.populate_buy_trend) # type: ignore
if hasattr(self.custom_hyperopt, 'populate_sell_trend'):
self.backtesting.strategy.advise_sell = \
self.custom_hyperopt.populate_sell_trend # type: ignore
self.backtesting.strategy.advise_sell = ( # type: ignore
self.custom_hyperopt.populate_sell_trend) # type: ignore
# Use max_open_trades for hyperopt as well, except --disable-max-market-positions is set
if self.config.get('use_max_market_positions', True):
@ -508,16 +508,16 @@ class Hyperopt:
params_details = self._get_params_details(params_dict)
if self.has_space('roi'):
self.backtesting.strategy.minimal_roi = \
self.custom_hyperopt.generate_roi_table(params_dict)
self.backtesting.strategy.minimal_roi = ( # type: ignore
self.custom_hyperopt.generate_roi_table(params_dict))
if self.has_space('buy'):
self.backtesting.strategy.advise_buy = \
self.custom_hyperopt.buy_strategy_generator(params_dict)
self.backtesting.strategy.advise_buy = ( # type: ignore
self.custom_hyperopt.buy_strategy_generator(params_dict))
if self.has_space('sell'):
self.backtesting.strategy.advise_sell = \
self.custom_hyperopt.sell_strategy_generator(params_dict)
self.backtesting.strategy.advise_sell = ( # type: ignore
self.custom_hyperopt.sell_strategy_generator(params_dict))
if self.has_space('stoploss'):
self.backtesting.strategy.stoploss = params_dict['stoploss']
@ -538,8 +538,8 @@ class Hyperopt:
backtesting_results = self.backtesting.backtest(
processed=processed,
stake_amount=self.config['stake_amount'],
start_date=min_date,
end_date=max_date,
start_date=min_date.datetime,
end_date=max_date.datetime,
max_open_trades=self.max_open_trades,
position_stacking=self.position_stacking,
)