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

Merge pull request #1113 from berlinguyinca/backslap_develop

Browse Source 
Backslap develop
This commit is contained in:
Matthias 2018-08-02 10:30:36 +02:00 committed by GitHub
commit 4f5b530dcb
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
6 changed files with 1000 additions and 60 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
freqtrade/arguments.py
View File

@ -161,6 +161,14 @@ class Arguments(object):
dest='exportfilename', dest='exportfilename',
metavar='PATH', metavar='PATH',
) )
parser.add_argument(
'--backslap',
help="Utilize the Backslapping approach instead of the default Backtesting. This should provide more "
"accurate results, unless you are utilizing Min/Max function in your strategy.",
required=False,
dest='backslap',
action='store_true'
)
@staticmethod @staticmethod
def optimizer_shared_options(parser: argparse.ArgumentParser) -> None: def optimizer_shared_options(parser: argparse.ArgumentParser) -> None:

22
freqtrade/optimize/__init__.py
View File

@ -10,9 +10,16 @@ import arrow
from freqtrade import misc, constants, OperationalException from freqtrade import misc, constants, OperationalException
from freqtrade.exchange import Exchange from freqtrade.exchange import Exchange
from freqtrade.arguments import TimeRange from freqtrade.arguments import TimeRange
import importlib
ujson_found = importlib.util.find_spec("ujson")
if ujson_found is not None:
import ujson
logger = logging.getLogger(__name__) logger = logging.getLogger(__name__)
if ujson_found is not None:
logger.debug('Loaded UltraJson ujson in optimize.py')
def trim_tickerlist(tickerlist: List[Dict], timerange: TimeRange) -> List[Dict]: def trim_tickerlist(tickerlist: List[Dict], timerange: TimeRange) -> List[Dict]:
if not tickerlist: if not tickerlist:
@ -63,11 +70,17 @@ def load_tickerdata_file(
if os.path.isfile(gzipfile): if os.path.isfile(gzipfile):
logger.debug('Loading ticker data from file %s', gzipfile) logger.debug('Loading ticker data from file %s', gzipfile)
with gzip.open(gzipfile) as tickerdata: with gzip.open(gzipfile) as tickerdata:
pairdata = json.load(tickerdata) if ujson_found is not None:
pairdata = ujson.load(tickerdata, precise_float=True)
else:
pairdata = json.load(tickerdata)
elif os.path.isfile(file): elif os.path.isfile(file):
logger.debug('Loading ticker data from file %s', file) logger.debug('Loading ticker data from file %s', file)
with open(file) as tickerdata: with open(file) as tickerdata:
pairdata = json.load(tickerdata) if ujson_found is not None:
pairdata = ujson.load(tickerdata, precise_float=True)
else:
pairdata = json.load(tickerdata)
else: else:
return None return None
@ -163,7 +176,10 @@ def load_cached_data_for_updating(filename: str,
# read the cached file # read the cached file
if os.path.isfile(filename): if os.path.isfile(filename):
with open(filename, "rt") as file: with open(filename, "rt") as file:
data = json.load(file) if ujson_found is not None:
data = ujson.load(file, precise_float=True)
else:
data = json.load(file)
# remove the last item, because we are not sure if it is correct # remove the last item, because we are not sure if it is correct
# it could be fetched when the candle was incompleted # it could be fetched when the candle was incompleted
if data: if data:

789
freqtrade/optimize/backslapping.py Normal file
View File

@ -0,0 +1,789 @@
import timeit
from typing import Dict, Any
from pandas import DataFrame
from freqtrade.exchange import Exchange
from freqtrade.strategy import IStrategy
from freqtrade.strategy.interface import SellType
from freqtrade.strategy.resolver import StrategyResolver
class Backslapping:
"""
provides a quick way to evaluate strategies over a longer term of time
"""
def __init__(self, config: Dict[str, Any], exchange = None) -> None:
"""
constructor
"""
self.config = config
self.strategy: IStrategy = StrategyResolver(self.config).strategy
self.ticker_interval = self.strategy.ticker_interval
self.tickerdata_to_dataframe = self.strategy.tickerdata_to_dataframe
self.populate_buy_trend = self.strategy.populate_buy_trend
self.populate_sell_trend = self.strategy.populate_sell_trend
###
#
###
if exchange is None:
self.config['exchange']['secret'] = ''
self.config['exchange']['password'] = ''
self.config['exchange']['uid'] = ''
self.config['dry_run'] = True
self.exchange = Exchange(self.config)
else:
self.exchange = exchange
self.fee = self.exchange.get_fee()
self.stop_loss_value = self.strategy.stoploss
#### backslap config
'''
Numpy arrays are used for 100x speed up
We requires setting Int values for
buy stop triggers and stop calculated on
# buy 0 - open 1 - close 2 - sell 3 - high 4 - low 5 - stop 6
'''
self.np_buy: int = 0
self.np_open: int = 1
self.np_close: int = 2
self.np_sell: int = 3
self.np_high: int = 4
self.np_low: int = 5
self.np_stop: int = 6
self.np_bto: int = self.np_close # buys_triggered_on - should be close
self.np_bco: int = self.np_open # buys calculated on - open of the next candle.
self.np_sto: int = self.np_low # stops_triggered_on - Should be low, FT uses close
self.np_sco: int = self.np_stop # stops_calculated_on - Should be stop, FT uses close
# self.np_sto: int = self.np_close # stops_triggered_on - Should be low, FT uses close
# self.np_sco: int = self.np_close # stops_calculated_on - Should be stop, FT uses close
self.debug = False # Main debug enable, very print heavy, enable 2 loops recommended
self.debug_timing = False # Stages within Backslap
self.debug_2loops = False # Limit each pair to two loops, useful when debugging
self.debug_vector = False # Debug vector calcs
self.debug_timing_main_loop = False # print overall timing per pair - works in Backtest and Backslap
self.backslap_show_trades = False # prints trades in addition to summary report
self.backslap_save_trades = True # saves trades as a pretty table to backslap.txt
self.stop_stops: int = 9999 # stop back testing any pair with this many stops, set to 999999 to not hit
def s(self):
st = timeit.default_timer()
return st
def f(self, st):
return (timeit.default_timer() - st)
def run(self,args):
headers = ['date', 'buy', 'open', 'close', 'sell', 'high', 'low']
processed = args['processed']
max_open_trades = args.get('max_open_trades', 0)
realistic = args.get('realistic', False)
trades = []
trade_count_lock: Dict = {}
########################### Call out BSlap Loop instead of Original BT code
bslap_results: list = []
for pair, pair_data in processed.items():
if self.debug_timing: # Start timer
fl = self.s()
ticker_data = self.populate_sell_trend(
self.populate_buy_trend(pair_data))[headers].copy()
if self.debug_timing: # print time taken
flt = self.f(fl)
# print("populate_buy_trend:", pair, round(flt, 10))
st = self.s()
# #dump same DFs to disk for offline testing in scratch
# f_pair:str = pair
# csv = f_pair.replace("/", "_")
# csv="/Users/creslin/PycharmProjects/freqtrade_new/frames/" + csv
# ticker_data.to_csv(csv, sep='\t', encoding='utf-8')
# call bslap - results are a list of dicts
bslap_pair_results = self.backslap_pair(ticker_data, pair)
last_bslap_results = bslap_results
bslap_results = last_bslap_results + bslap_pair_results
if self.debug_timing: # print time taken
tt = self.f(st)
print("Time to BackSlap :", pair, round(tt, 10))
print("-----------------------")
# Switch List of Trade Dicts (bslap_results) to Dataframe
# Fill missing, calculable columns, profit, duration , abs etc.
bslap_results_df = DataFrame(bslap_results)
if len(bslap_results_df) > 0: # Only post process a frame if it has a record
# bslap_results_df['open_time'] = to_datetime(bslap_results_df['open_time'])
# bslap_results_df['close_time'] = to_datetime(bslap_results_df['close_time'])
# if debug:
# print("open_time and close_time converted to datetime columns")
bslap_results_df = self.vector_fill_results_table(bslap_results_df, pair)
else:
from freqtrade.optimize.backtesting import BacktestResult
bslap_results_df = []
bslap_results_df = DataFrame.from_records(bslap_results_df, columns=BacktestResult._fields)
return bslap_results_df
def vector_fill_results_table(self, bslap_results_df: DataFrame, pair: str):
"""
The Results frame contains a number of columns that are calculable
from othe columns. These are left blank till all rows are added,
to be populated in single vector calls.
Columns to be populated are:
- Profit
- trade duration
- profit abs
:param bslap_results Dataframe
:return: bslap_results Dataframe
"""
import pandas as pd
import numpy as np
debug = self.debug_vector
# stake and fees
# stake = 0.015
# 0.05% is 0.0005
# fee = 0.001
stake = self.config.get('stake_amount')
fee = self.fee
open_fee = fee / 2
close_fee = fee / 2
if debug:
print("Stake is,", stake, "the sum of currency to spend per trade")
print("The open fee is", open_fee, "The close fee is", close_fee)
if debug:
from pandas import set_option
set_option('display.max_rows', 5000)
set_option('display.max_columns', 20)
pd.set_option('display.width', 1000)
pd.set_option('max_colwidth', 40)
pd.set_option('precision', 12)
# # Get before
# csv = "cryptosher_before_debug"
# bslap_results_df.to_csv(csv, sep='\t', encoding='utf-8')
# bslap_results_df.to_csv(csv, sep='\t', encoding='utf-8')
bslap_results_df['trade_duration'] = bslap_results_df['close_time'] - bslap_results_df['open_time']
bslap_results_df['trade_duration'] = bslap_results_df['trade_duration'].map(lambda x: int(x.total_seconds() / 60))
## Spends, Takes, Profit, Absolute Profit
# print(bslap_results_df)
# Buy Price
bslap_results_df['buy_vol'] = stake / bslap_results_df['open_rate'] # How many target are we buying
bslap_results_df['buy_fee'] = stake * open_fee
bslap_results_df['buy_spend'] = stake + bslap_results_df['buy_fee'] # How much we're spending
# Sell price
bslap_results_df['sell_sum'] = bslap_results_df['buy_vol'] * bslap_results_df['close_rate']
bslap_results_df['sell_fee'] = bslap_results_df['sell_sum'] * close_fee
bslap_results_df['sell_take'] = bslap_results_df['sell_sum'] - bslap_results_df['sell_fee']
# profit_percent
bslap_results_df['profit_percent'] = (bslap_results_df['sell_take'] - bslap_results_df['buy_spend']) \
/ bslap_results_df['buy_spend']
# Absolute profit
bslap_results_df['profit_abs'] = bslap_results_df['sell_take'] - bslap_results_df['buy_spend']
# # Get After
# csv="cryptosher_after_debug"
# bslap_results_df.to_csv(csv, sep='\t', encoding='utf-8')
if debug:
print("\n")
print(bslap_results_df[
['buy_vol', 'buy_fee', 'buy_spend', 'sell_sum', 'sell_fee', 'sell_take', 'profit_percent',
'profit_abs', 'exit_type']])
return bslap_results_df
def np_get_t_open_ind(self, np_buy_arr, t_exit_ind: int, np_buy_arr_len: int, stop_stops: int,
stop_stops_count: int):
import utils_find_1st as utf1st
"""
The purpose of this def is to return the next "buy" = 1
after t_exit_ind.
This function will also check is the stop limit for the pair has been reached.
if stop_stops is the limit and stop_stops_count it the number of times the stop has been hit.
t_exit_ind is the index the last trade exited on
or 0 if first time around this loop.
stop_stops i
"""
debug = self.debug
# Timers, to be called if in debug
def s():
st = timeit.default_timer()
return st
def f(st):
return (timeit.default_timer() - st)
st = s()
t_open_ind: int
"""
Create a view on our buy index starting after last trade exit
Search for next buy
"""
np_buy_arr_v = np_buy_arr[t_exit_ind:]
t_open_ind = utf1st.find_1st(np_buy_arr_v, 1, utf1st.cmp_equal)
'''
If -1 is returned no buy has been found, preserve the value
'''
if t_open_ind != -1: # send back the -1 if no buys found. otherwise update index
t_open_ind = t_open_ind + t_exit_ind # Align numpy index
if t_open_ind == np_buy_arr_len - 1: # If buy found on last candle ignore, there is no OPEN in next to use
t_open_ind = -1 # -1 ends the loop
if stop_stops_count >= stop_stops: # if maximum number of stops allowed in a pair is hit, exit loop
t_open_ind = -1 # -1 ends the loop
if debug:
print("Max stop limit ", stop_stops, "reached. Moving to next pair")
return t_open_ind
def backslap_pair(self, ticker_data, pair):
import pandas as pd
import numpy as np
import timeit
import utils_find_1st as utf1st
from datetime import datetime
### backslap debug wrap
# debug_2loops = False # only loop twice, for faster debug
# debug_timing = False # print timing for each step
# debug = False # print values, to check accuracy
debug_2loops = self.debug_2loops # only loop twice, for faster debug
debug_timing = self.debug_timing # print timing for each step
debug = self.debug # print values, to check accuracy
# Read Stop Loss Values and Stake
stop = self.stop_loss_value
p_stop = (stop + 1) # What stop really means, e.g 0.01 is 0.99 of price
if debug:
print("Stop is ", stop, "value from stragey file")
print("p_stop is", p_stop, "value used to multiply to entry price")
if debug:
from pandas import set_option
set_option('display.max_rows', 5000)
set_option('display.max_columns', 8)
pd.set_option('display.width', 1000)
pd.set_option('max_colwidth', 40)
pd.set_option('precision', 12)
def s():
st = timeit.default_timer()
return st
def f(st):
return (timeit.default_timer() - st)
#### backslap config
'''
Numpy arrays are used for 100x speed up
We requires setting Int values for
buy stop triggers and stop calculated on
# buy 0 - open 1 - close 2 - sell 3 - high 4 - low 5 - stop 6
'''
#######
# Use vars set at top of backtest
np_buy: int = self.np_buy
np_open: int = self.np_open
np_close: int = self.np_close
np_sell: int = self.np_sell
np_high: int = self.np_high
np_low: int = self.np_low
np_stop: int = self.np_stop
np_bto: int = self.np_bto # buys_triggered_on - should be close
np_bco: int = self.np_bco # buys calculated on - open of the next candle.
np_sto: int = self.np_sto # stops_triggered_on - Should be low, FT uses close
np_sco: int = self.np_sco # stops_calculated_on - Should be stop, FT uses close
### End Config
pair: str = pair
# ticker_data: DataFrame = ticker_dfs[t_file]
bslap: DataFrame = ticker_data
# Build a single dimension numpy array from "buy" index for faster search
# (500x faster than pandas)
np_buy_arr = bslap['buy'].values
np_buy_arr_len: int = len(np_buy_arr)
# use numpy array for faster searches in loop, 20x faster than pandas
# buy 0 - open 1 - close 2 - sell 3 - high 4 - low 5
np_bslap = np.array(bslap[['buy', 'open', 'close', 'sell', 'high', 'low']])
# Build a numpy list of date-times.
# We use these when building the trade
# The rationale is to address a value from a pandas cell is thousands of
# times more expensive. Processing time went X25 when trying to use any data from pandas
np_bslap_dates = bslap['date'].values
loop: int = 0 # how many time around the loop
t_exit_ind = 0 # Start loop from first index
t_exit_last = 0 # To test for exit
stop_stops = self.stop_stops # Int of stops within a pair to stop trading a pair at
stop_stops_count = 0 # stop counter per pair
st = s() # Start timer for processing dataframe
if debug:
print('Processing:', pair)
# Results will be stored in a list of dicts
bslap_pair_results: list = []
bslap_result: dict = {}
while t_exit_ind < np_buy_arr_len:
loop = loop + 1
if debug or debug_timing:
print("-- T_exit_Ind - Numpy Index is", t_exit_ind, " ----------------------- Loop", loop, pair)
if debug_2loops:
if loop == 3:
print(
"++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++Loop debug max met - breaking")
break
'''
Dev phases
Phase 1
1) Manage buy, sell, stop enter/exit
a) Find first buy index
b) Discover first stop and sell hit after buy index
c) Chose first instance as trade exit
Phase 2
2) Manage dynamic Stop and ROI Exit
a) Create trade slice from 1
b) search within trade slice for dynamice stop hit
c) search within trade slice for ROI hit
'''
if debug_timing:
st = s()
'''
0 - Find next buy entry
Finds index for first (buy = 1) flag
Requires: np_buy_arr - a 1D array of the 'buy' column. To find next "1"
Required: t_exit_ind - Either 0, first loop. Or The index we last exited on
Requires: np_buy_arr_len - length of pair array.
Requires: stops_stops - number of stops allowed before stop trading a pair
Requires: stop_stop_counts - count of stops hit in the pair
Provides: The next "buy" index after t_exit_ind
If -1 is returned no buy has been found in remainder of array, skip to exit loop
'''
t_open_ind = self.np_get_t_open_ind(np_buy_arr, t_exit_ind, np_buy_arr_len, stop_stops, stop_stops_count)
if debug:
print("\n(0) numpy debug \nnp_get_t_open, has returned the next valid buy index as", t_open_ind)
print("If -1 there are no valid buys in the remainder of ticker data. Skipping to end of loop")
if debug_timing:
t_t = f(st)
print("0-numpy", str.format('{0:.17f}', t_t))
st = s()
if t_open_ind != -1:
"""
1 - Create views to search within for our open trade
The views are our search space for the next Stop or Sell
Numpy view is employed as:
1,000 faster than pandas searches
Pandas cannot assure it will always return a view, it may make a slow copy.
The view contains columns:
buy 0 - open 1 - close 2 - sell 3 - high 4 - low 5
Requires: np_bslap is our numpy array of the ticker DataFrame
Requires: t_open_ind is the index row with the buy.
Provides: np_t_open_v View of array after buy.
Provides: np_t_open_v_stop View of array after buy +1
(Stop will search in here to prevent stopping in the past)
"""
np_t_open_v = np_bslap[t_open_ind:]
np_t_open_v_stop = np_bslap[t_open_ind + 1:]
if debug:
print("\n(1) numpy debug \nNumpy view row 0 is now Ticker_Data Index", t_open_ind)
print("Numpy View: Buy - Open - Close - Sell - High - Low")
print("Row 0", np_t_open_v[0])
print("Row 1", np_t_open_v[1], )
if debug_timing:
t_t = f(st)
print("2-numpy", str.format('{0:.17f}', t_t))
st = s()
'''
2 - Calculate our stop-loss price
As stop is based on buy price of our trade
- (BTO)Buys are Triggered On np_bto, typically the CLOSE of candle
- (BCO)Buys are Calculated On np_bco, default is OPEN of the next candle.
This is as we only see the CLOSE after it has happened.
The back test assumption is we have bought at first available price, the OPEN
Requires: np_bslap - is our numpy array of the ticker DataFrame
Requires: t_open_ind - is the index row with the first buy.
Requires: p_stop - is the stop rate, ie. 0.99 is -1%
Provides: np_t_stop_pri - The value stop-loss will be triggered on
'''
np_t_stop_pri = (np_bslap[t_open_ind + 1, np_bco] * p_stop)
if debug:
print("\n(2) numpy debug\nStop-Loss has been calculated at:", np_t_stop_pri)
if debug_timing:
t_t = f(st)
print("2-numpy", str.format('{0:.17f}', t_t))
st = s()
'''
3 - Find candle STO is under Stop-Loss After Trade opened.
where [np_sto] (stop tiggered on variable: "close", "low" etc) < np_t_stop_pri
Requires: np_t_open_v_stop Numpy view of ticker_data after buy row +1 (when trade was opened)
Requires: np_sto User Var(STO)StopTriggeredOn. Typically set to "low" or "close"
Requires: np_t_stop_pri The stop-loss price STO must fall under to trigger stop
Provides: np_t_stop_ind The first candle after trade open where STO is under stop-loss
'''
np_t_stop_ind = utf1st.find_1st(np_t_open_v_stop[:, np_sto],
np_t_stop_pri,
utf1st.cmp_smaller)
# plus 1 as np_t_open_v_stop is 1 ahead of view np_t_open_v, used from here on out.
np_t_stop_ind = np_t_stop_ind + 1
if debug:
print("\n(3) numpy debug\nNext view index with STO (stop trigger on) under Stop-Loss is",
np_t_stop_ind - 1,
". STO is using field", np_sto,
"\nFrom key: buy 0 - open 1 - close 2 - sell 3 - high 4 - low 5\n")
print(
"If -1 or 0 returned there is no stop found to end of view, then next two array lines are garbage")
print("Row", np_t_stop_ind - 1, np_t_open_v[np_t_stop_ind])
print("Row", np_t_stop_ind, np_t_open_v[np_t_stop_ind + 1])
if debug_timing:
t_t = f(st)
print("3-numpy", str.format('{0:.17f}', t_t))
st = s()
'''
4 - Find first sell index after trade open
First index in the view np_t_open_v where ['sell'] = 1
Requires: np_t_open_v - view of ticker_data from buy onwards
Requires: no_sell - integer '3', the buy column in the array
Provides: np_t_sell_ind index of view where first sell=1 after buy
'''
# Use numpy array for faster search for sell
# Sell uses column 3.
# buy 0 - open 1 - close 2 - sell 3 - high 4 - low 5
# Numpy searches 25-35x quicker than pandas on this data
np_t_sell_ind = utf1st.find_1st(np_t_open_v[:, np_sell],
1, utf1st.cmp_equal)
if debug:
print("\n(4) numpy debug\nNext view index with sell = 1 is ", np_t_sell_ind)
print("If 0 or less is returned there is no sell found to end of view, then next lines garbage")
print("Row", np_t_sell_ind, np_t_open_v[np_t_sell_ind])
print("Row", np_t_sell_ind + 1, np_t_open_v[np_t_sell_ind + 1])
if debug_timing:
t_t = f(st)
print("4-numpy", str.format('{0:.17f}', t_t))
st = s()
'''
5 - Determine which was hit first a stop or sell
To then use as exit index price-field (sell on buy, stop on stop)
STOP takes priority over SELL as would be 'in candle' from tick data
Sell would use Open from Next candle.
So in a draw Stop would be hit first on ticker data in live
Validity of when types of trades may be executed can be summarised as:
Tick View
index index Buy Sell open low close high Stop price
open 2am 94 -1 0 0 ----- ------ ------ ----- -----
open 3am 95 0 1 0 ----- ------ trg buy ----- -----
open 4am 96 1 0 1 Enter trgstop trg sel ROI out Stop out
open 5am 97 2 0 0 Exit ------ ------- ----- -----
open 6am 98 3 0 0 ----- ------ ------- ----- -----
-1 means not found till end of view i.e no valid Stop found. Exclude from match.
Stop tiggering and closing in 96-1, the candle we bought at OPEN in, is valid.
Buys and sells are triggered at candle close
Both will open their postions at the open of the next candle. i/e + 1 index
Stop and buy Indexes are on the view. To map to the ticker dataframe
the t_open_ind index should be summed.
np_t_stop_ind: Stop Found index in view
t_exit_ind : Sell found in view
t_open_ind : Where view was started on ticker_data
TODO: fix this frig for logic test,, case/switch/dictionary would be better...
more so when later testing many options, dynamic stop / roi etc
cludge - Setting np_t_sell_ind as 9999999999 when -1 (not found)
cludge - Setting np_t_stop_ind as 9999999999 when -1 (not found)
'''
if debug:
print("\n(5) numpy debug\nStop or Sell Logic Processing")
# cludge for logic test (-1) means it was not found, set crazy high to lose < test
np_t_sell_ind = 99999999 if np_t_sell_ind <= 0 else np_t_sell_ind
np_t_stop_ind = 99999999 if np_t_stop_ind <= 0 else np_t_stop_ind
# Stoploss trigger found before a sell =1
if np_t_stop_ind < 99999999 and np_t_stop_ind <= np_t_sell_ind:
t_exit_ind = t_open_ind + np_t_stop_ind # Set Exit row index
t_exit_type = SellType.STOP_LOSS # Set Exit type (stop)
np_t_exit_pri = np_sco # The price field our STOP exit will use
if debug:
print("Type STOP is first exit condition. "
"At view index:", np_t_stop_ind, ". Ticker data exit index is", t_exit_ind)
# Buy = 1 found before a stoploss triggered
elif np_t_sell_ind < 99999999 and np_t_sell_ind < np_t_stop_ind:
# move sell onto next candle, we only look back on sell
# will use the open price later.
t_exit_ind = t_open_ind + np_t_sell_ind + 1 # Set Exit row index
t_exit_type = SellType.SELL_SIGNAL # Set Exit type (sell)
np_t_exit_pri = np_open # The price field our SELL exit will use
if debug:
print("Type SELL is first exit condition. "
"At view index", np_t_sell_ind, ". Ticker data exit index is", t_exit_ind)
# No stop or buy left in view - set t_exit_last -1 to handle gracefully
else:
t_exit_last: int = -1 # Signal loop to exit, no buys or sells found.
t_exit_type = SellType.NONE
np_t_exit_pri = 999 # field price should be calculated on. 999 a non-existent column
if debug:
print("No valid STOP or SELL found. Signalling t_exit_last to gracefully exit")
# TODO: fix having to cludge/uncludge this ..
# Undo cludge
np_t_sell_ind = -1 if np_t_sell_ind == 99999999 else np_t_sell_ind
np_t_stop_ind = -1 if np_t_stop_ind == 99999999 else np_t_stop_ind
if debug_timing:
t_t = f(st)
print("5-logic", str.format('{0:.17f}', t_t))
st = s()
if debug:
'''
Print out the buys, stops, sells
Include Line before and after to for easy
Human verification
'''
# Combine the np_t_stop_pri value to bslap dataframe to make debug
# life easy. This is the current stop price based on buy price_
# This is slow but don't care about performance in debug
#
# When referencing equiv np_column, as example np_sto, its 5 in numpy and 6 in df, so +1
# as there is no data column in the numpy array.
bslap['np_stop_pri'] = np_t_stop_pri
# Buy
print("\n\nDATAFRAME DEBUG =================== BUY ", pair)
print("Numpy Array BUY Index is:", 0)
print("DataFrame BUY Index is:", t_open_ind, "displaying DF \n")
print("HINT, BUY trade should use OPEN price from next candle, i.e ", t_open_ind + 1)
op_is = t_open_ind - 1 # Print open index start, line before
op_if = t_open_ind + 3 # Print open index finish, line after
print(bslap.iloc[op_is:op_if], "\n")
# Stop - Stops trigger price np_sto (+1 for pandas column), and price received np_sco +1. (Stop Trigger|Calculated On)
if np_t_stop_ind < 0:
print("DATAFRAME DEBUG =================== STOP ", pair)
print("No STOPS were found until the end of ticker data file\n")
else:
print("DATAFRAME DEBUG =================== STOP ", pair)
print("Numpy Array STOP Index is:", np_t_stop_ind, "View starts at index", t_open_ind)
df_stop_index = (t_open_ind + np_t_stop_ind)
print("DataFrame STOP Index is:", df_stop_index, "displaying DF \n")
print("First Stoploss trigger after Trade entered at OPEN in candle", t_open_ind + 1, "is ",
df_stop_index, ": \n",
str.format('{0:.17f}', bslap.iloc[df_stop_index][np_sto + 1]),
"is less than", str.format('{0:.17f}', np_t_stop_pri))
print("A stoploss exit will be calculated at rate:",
str.format('{0:.17f}', bslap.iloc[df_stop_index][np_sco + 1]))
print("\nHINT, STOPs should exit in-candle, i.e", df_stop_index,
": As live STOPs are not linked to O-C times")
st_is = df_stop_index - 1 # Print stop index start, line before
st_if = df_stop_index + 2 # Print stop index finish, line after
print(bslap.iloc[st_is:st_if], "\n")
# Sell
if np_t_sell_ind < 0:
print("DATAFRAME DEBUG =================== SELL ", pair)
print("No SELLS were found till the end of ticker data file\n")
else:
print("DATAFRAME DEBUG =================== SELL ", pair)
print("Numpy View SELL Index is:", np_t_sell_ind, "View starts at index", t_open_ind)
df_sell_index = (t_open_ind + np_t_sell_ind)
print("DataFrame SELL Index is:", df_sell_index, "displaying DF \n")
print("First Sell Index after Trade open is in candle", df_sell_index)
print("HINT, if exit is SELL (not stop) trade should use OPEN price from next candle",
df_sell_index + 1)
sl_is = df_sell_index - 1 # Print sell index start, line before
sl_if = df_sell_index + 3 # Print sell index finish, line after
print(bslap.iloc[sl_is:sl_if], "\n")
# Chosen Exit (stop or sell)
print("DATAFRAME DEBUG =================== EXIT ", pair)
print("Exit type is :", t_exit_type)
print("trade exit price field is", np_t_exit_pri, "\n")
if debug_timing:
t_t = f(st)
print("6-depra", str.format('{0:.17f}', t_t))
st = s()
## use numpy view "np_t_open_v" for speed. Columns are
# buy 0 - open 1 - close 2 - sell 3 - high 4 - low 5
# exception is 6 which is use the stop value.
# TODO no! this is hard coded bleh fix this open
np_trade_enter_price = np_bslap[t_open_ind + 1, np_open]
if t_exit_type == SellType.STOP_LOSS:
if np_t_exit_pri == 6:
np_trade_exit_price = np_t_stop_pri
else:
np_trade_exit_price = np_bslap[t_exit_ind, np_t_exit_pri]
if t_exit_type == SellType.SELL_SIGNAL:
np_trade_exit_price = np_bslap[t_exit_ind, np_t_exit_pri]
# Catch no exit found
if t_exit_type == SellType.NONE:
np_trade_exit_price = 0
if debug_timing:
t_t = f(st)
print("7-numpy", str.format('{0:.17f}', t_t))
st = s()
if debug:
print("//////////////////////////////////////////////")
print("+++++++++++++++++++++++++++++++++ Trade Enter ")
print("np_trade Enter Price is ", str.format('{0:.17f}', np_trade_enter_price))
print("--------------------------------- Trade Exit ")
print("Trade Exit Type is ", t_exit_type)
print("np_trade Exit Price is", str.format('{0:.17f}', np_trade_exit_price))
print("//////////////////////////////////////////////")
else: # no buys were found, step 0 returned -1
# Gracefully exit the loop
t_exit_last == -1
if debug:
print("\n(E) No buys were found in remaining ticker file. Exiting", pair)
# Loop control - catch no closed trades.
if debug:
print("---------------------------------------- end of loop", loop,
" Dataframe Exit Index is: ", t_exit_ind)
print("Exit Index Last, Exit Index Now Are: ", t_exit_last, t_exit_ind)
if t_exit_last >= t_exit_ind or t_exit_last == -1:
"""
Break loop and go on to next pair.
When last trade exit equals index of last exit, there is no
opportunity to close any more trades.
"""
# TODO :add handing here to record none closed open trades
if debug:
print(bslap_pair_results)
break
else:
"""
Add trade to backtest looking results list of dicts
Loop back to look for more trades.
"""
# Build trade dictionary
## In general if a field can be calculated later from other fields leave blank here
## Its X(number of trades faster) to calc all in a single vector than 1 trade at a time
# create a new dict
close_index: int = t_exit_ind
bslap_result = {} # Must have at start or we end up with a list of multiple same last result
bslap_result["pair"] = pair
bslap_result["profit_percent"] = "" # To be 1 vector calc across trades when loop complete
bslap_result["profit_abs"] = "" # To be 1 vector calc across trades when loop complete
bslap_result["open_time"] = np_bslap_dates[t_open_ind + 1] # use numpy array, pandas 20x slower
bslap_result["close_time"] = np_bslap_dates[close_index] # use numpy array, pandas 20x slower
bslap_result["open_index"] = t_open_ind + 1 # +1 as we buy on next.
bslap_result["close_index"] = close_index
bslap_result["trade_duration"] = "" # To be 1 vector calc across trades when loop complete
bslap_result["open_at_end"] = False
bslap_result["open_rate"] = round(np_trade_enter_price, 15)
bslap_result["close_rate"] = round(np_trade_exit_price, 15)
bslap_result["exit_type"] = t_exit_type
bslap_result["sell_reason"] = t_exit_type #duplicated, but I don't care
# append the dict to the list and print list
bslap_pair_results.append(bslap_result)
if t_exit_type is SellType.STOP_LOSS:
stop_stops_count = stop_stops_count + 1
if debug:
print("The trade dict is: \n", bslap_result)
print("Trades dicts in list after append are: \n ", bslap_pair_results)
"""
Loop back to start. t_exit_last becomes where loop
will seek to open new trades from.
Push index on 1 to not open on close
"""
t_exit_last = t_exit_ind + 1
if debug_timing:
t_t = f(st)
print("8+trade", str.format('{0:.17f}', t_t))
# Send back List of trade dicts
return bslap_pair_results

230
freqtrade/optimize/backtesting.py
View File

@ -10,7 +10,7 @@ from datetime import datetime, timedelta
from typing import Any, Dict, List, NamedTuple, Optional, Tuple from typing import Any, Dict, List, NamedTuple, Optional, Tuple
import arrow import arrow
from pandas import DataFrame from pandas import DataFrame, to_datetime
from tabulate import tabulate from tabulate import tabulate
import freqtrade.optimize as optimize import freqtrade.optimize as optimize
@ -19,9 +19,14 @@ from freqtrade.arguments import Arguments
from freqtrade.configuration import Configuration from freqtrade.configuration import Configuration
from freqtrade.exchange import Exchange from freqtrade.exchange import Exchange
from freqtrade.misc import file_dump_json from freqtrade.misc import file_dump_json
from freqtrade.optimize.backslapping import Backslapping
from freqtrade.persistence import Trade from freqtrade.persistence import Trade
from freqtrade.strategy.interface import SellType from freqtrade.strategy.interface import SellType
from freqtrade.strategy.resolver import IStrategy, StrategyResolver from freqtrade.strategy.resolver import IStrategy, StrategyResolver
from profilehooks import profile
from collections import OrderedDict
import timeit
from time import sleep
logger = logging.getLogger(__name__) logger = logging.getLogger(__name__)
@ -52,6 +57,7 @@ class Backtesting(object):
backtesting = Backtesting(config) backtesting = Backtesting(config)
backtesting.start() backtesting.start()
""" """
def __init__(self, config: Dict[str, Any]) -> None: def __init__(self, config: Dict[str, Any]) -> None:
self.config = config self.config = config
self.strategy: IStrategy = StrategyResolver(self.config).strategy self.strategy: IStrategy = StrategyResolver(self.config).strategy
@ -69,6 +75,49 @@ class Backtesting(object):
self.exchange = Exchange(self.config) self.exchange = Exchange(self.config)
self.fee = self.exchange.get_fee() self.fee = self.exchange.get_fee()
self.stop_loss_value = self.strategy.stoploss
#### backslap config
'''
Numpy arrays are used for 100x speed up
We requires setting Int values for
buy stop triggers and stop calculated on
# buy 0 - open 1 - close 2 - sell 3 - high 4 - low 5 - stop 6
'''
self.np_buy: int = 0
self.np_open: int = 1
self.np_close: int = 2
self.np_sell: int = 3
self.np_high: int = 4
self.np_low: int = 5
self.np_stop: int = 6
self.np_bto: int = self.np_close # buys_triggered_on - should be close
self.np_bco: int = self.np_open # buys calculated on - open of the next candle.
self.np_sto: int = self.np_low # stops_triggered_on - Should be low, FT uses close
self.np_sco: int = self.np_stop # stops_calculated_on - Should be stop, FT uses close
# self.np_sto: int = self.np_close # stops_triggered_on - Should be low, FT uses close
# self.np_sco: int = self.np_close # stops_calculated_on - Should be stop, FT uses close
if 'backslap' in config:
self.use_backslap = config['backslap'] # Enable backslap - if false Orginal code is executed.
else:
self.use_backslap = False
logger.info("using backslap: {}".format(self.use_backslap))
self.debug = False # Main debug enable, very print heavy, enable 2 loops recommended
self.debug_timing = False # Stages within Backslap
self.debug_2loops = False # Limit each pair to two loops, useful when debugging
self.debug_vector = False # Debug vector calcs
self.debug_timing_main_loop = False # print overall timing per pair - works in Backtest and Backslap
self.backslap_show_trades = False # prints trades in addition to summary report
self.backslap_save_trades = True # saves trades as a pretty table to backslap.txt
self.stop_stops: int = 9999 # stop back testing any pair with this many stops, set to 999999 to not hit
self.backslap = Backslapping(config)
@staticmethod @staticmethod
def get_timeframe(data: Dict[str, DataFrame]) -> Tuple[arrow.Arrow, arrow.Arrow]: def get_timeframe(data: Dict[str, DataFrame]) -> Tuple[arrow.Arrow, arrow.Arrow]:
""" """
@ -81,7 +130,7 @@ class Backtesting(object):
for frame in data.values() for frame in data.values()
] ]
return min(timeframe, key=operator.itemgetter(0))[0], \ return min(timeframe, key=operator.itemgetter(0))[0], \
max(timeframe, key=operator.itemgetter(1))[1] max(timeframe, key=operator.itemgetter(1))[1]
def _generate_text_table(self, data: Dict[str, Dict], results: DataFrame) -> str: def _generate_text_table(self, data: Dict[str, Dict], results: DataFrame) -> str:
""" """
@ -126,10 +175,11 @@ class Backtesting(object):
""" """
Generate small table outlining Backtest results Generate small table outlining Backtest results
""" """
tabular_data = [] tabular_data = []
headers = ['Sell Reason', 'Count'] headers = ['Sell Reason', 'Count']
for reason, count in results['sell_reason'].value_counts().iteritems(): for reason, count in results['sell_reason'].value_counts().iteritems():
tabular_data.append([reason.value, count]) tabular_data.append([reason.value, count])
return tabulate(tabular_data, headers=headers, tablefmt="pipe") return tabulate(tabular_data, headers=headers, tablefmt="pipe")
def _store_backtest_result(self, recordfilename: Optional[str], results: DataFrame) -> None: def _store_backtest_result(self, recordfilename: Optional[str], results: DataFrame) -> None:
@ -168,14 +218,13 @@ class Backtesting(object):
sell = self.strategy.should_sell(trade, sell_row.open, sell_row.date, buy_signal, sell = self.strategy.should_sell(trade, sell_row.open, sell_row.date, buy_signal,
sell_row.sell) sell_row.sell)
if sell.sell_flag: if sell.sell_flag:
return BacktestResult(pair=pair, return BacktestResult(pair=pair,
profit_percent=trade.calc_profit_percent(rate=sell_row.open), profit_percent=trade.calc_profit_percent(rate=sell_row.open),
profit_abs=trade.calc_profit(rate=sell_row.open), profit_abs=trade.calc_profit(rate=sell_row.open),
open_time=buy_row.date, open_time=buy_row.date,
close_time=sell_row.date, close_time=sell_row.date,
trade_duration=int(( trade_duration=int((
sell_row.date - buy_row.date).total_seconds() // 60), sell_row.date - buy_row.date).total_seconds() // 60),
open_index=buy_row.Index, open_index=buy_row.Index,
close_index=sell_row.Index, close_index=sell_row.Index,
open_at_end=False, open_at_end=False,
@ -192,7 +241,7 @@ class Backtesting(object):
open_time=buy_row.date, open_time=buy_row.date,
close_time=sell_row.date, close_time=sell_row.date,
trade_duration=int(( trade_duration=int((
sell_row.date - buy_row.date).total_seconds() // 60), sell_row.date - buy_row.date).total_seconds() // 60),
open_index=buy_row.Index, open_index=buy_row.Index,
close_index=sell_row.Index, close_index=sell_row.Index,
open_at_end=True, open_at_end=True,
@ -205,6 +254,13 @@ class Backtesting(object):
return btr return btr
return None return None
def s(self):
st = timeit.default_timer()
return st
def f(self, st):
return (timeit.default_timer() - st)
def backtest(self, args: Dict) -> DataFrame: def backtest(self, args: Dict) -> DataFrame:
""" """
Implements backtesting functionality Implements backtesting functionality
@ -220,32 +276,50 @@ class Backtesting(object):
position_stacking: do we allow position stacking? (default: False) position_stacking: do we allow position stacking? (default: False)
:return: DataFrame :return: DataFrame
""" """
headers = ['date', 'buy', 'open', 'close', 'sell']
processed = args['processed']
max_open_trades = args.get('max_open_trades', 0)
position_stacking = args.get('position_stacking', False)
trades = []
trade_count_lock: Dict = {}
for pair, pair_data in processed.items():
pair_data['buy'], pair_data['sell'] = 0, 0 # cleanup from previous run
ticker_data = self.advise_sell( use_backslap = self.use_backslap
debug_timing = self.debug_timing_main_loop
if use_backslap: # Use Back Slap code
return self.backslap.run(args)
else: # use Original Back test code
########################## Original BT loop
headers = ['date', 'buy', 'open', 'close', 'sell']
processed = args['processed']
max_open_trades = args.get('max_open_trades', 0)
position_stacking = args.get('position_stacking', False)
trades = []
trade_count_lock: Dict = {}
for pair, pair_data in processed.items():
if debug_timing: # Start timer
fl = self.s()
pair_data['buy'], pair_data['sell'] = 0, 0 # cleanup from previous run
ticker_data = self.advise_sell(
self.advise_buy(pair_data, {'pair': pair}), {'pair': pair})[headers].copy() self.advise_buy(pair_data, {'pair': pair}), {'pair': pair})[headers].copy()
# to avoid using data from future, we buy/sell with signal from previous candle # to avoid using data from future, we buy/sell with signal from previous candle
ticker_data.loc[:, 'buy'] = ticker_data['buy'].shift(1) ticker_data.loc[:, 'buy'] = ticker_data['buy'].shift(1)
ticker_data.loc[:, 'sell'] = ticker_data['sell'].shift(1) ticker_data.loc[:, 'sell'] = ticker_data['sell'].shift(1)
ticker_data.drop(ticker_data.head(1).index, inplace=True) ticker_data.drop(ticker_data.head(1).index, inplace=True)
# Convert from Pandas to list for performance reasons if debug_timing: # print time taken
# (Looping Pandas is slow.) flt = self.f(fl)
ticker = [x for x in ticker_data.itertuples()] # print("populate_buy_trend:", pair, round(flt, 10))
st = self.s()
lock_pair_until = None # Convert from Pandas to list for performance reasons
for index, row in enumerate(ticker): # (Looping Pandas is slow.)
if row.buy == 0 or row.sell == 1: ticker = [x for x in ticker_data.itertuples()]
continue # skip rows where no buy signal or that would immediately sell off
lock_pair_until = None
for index, row in enumerate(ticker):
if row.buy == 0 or row.sell == 1:
continue # skip rows where no buy signal or that would immediately sell off
if not position_stacking: if not position_stacking:
if lock_pair_until is not None and row.date <= lock_pair_until: if lock_pair_until is not None and row.date <= lock_pair_until:
@ -255,20 +329,26 @@ class Backtesting(object):
if not trade_count_lock.get(row.date, 0) < max_open_trades: if not trade_count_lock.get(row.date, 0) < max_open_trades:
continue continue
trade_count_lock[row.date] = trade_count_lock.get(row.date, 0) + 1 trade_count_lock[row.date] = trade_count_lock.get(row.date, 0) + 1
trade_entry = self._get_sell_trade_entry(pair, row, ticker[index + 1:], trade_entry = self._get_sell_trade_entry(pair, row, ticker[index + 1:],
trade_count_lock, args) trade_count_lock, args)
if trade_entry: if trade_entry:
lock_pair_until = trade_entry.close_time lock_pair_until = trade_entry.close_time
trades.append(trade_entry) trades.append(trade_entry)
else: else:
# Set lock_pair_until to end of testing period if trade could not be closed # Set lock_pair_until to end of testing period if trade could not be closed
# This happens only if the buy-signal was with the last candle # This happens only if the buy-signal was with the last candle
lock_pair_until = ticker_data.iloc[-1].date lock_pair_until = ticker_data.iloc[-1].date
return DataFrame.from_records(trades, columns=BacktestResult._fields) if debug_timing: # print time taken
tt = self.f(st)
print("Time to BackTest :", pair, round(tt, 10))
print("-----------------------")
return DataFrame.from_records(trades, columns=BacktestResult._fields)
####################### Original BT loop end
def start(self) -> None: def start(self) -> None:
""" """
@ -289,6 +369,7 @@ class Backtesting(object):
timerange = Arguments.parse_timerange(None if self.config.get( timerange = Arguments.parse_timerange(None if self.config.get(
'timerange') is None else str(self.config.get('timerange'))) 'timerange') is None else str(self.config.get('timerange')))
data = optimize.load_data( data = optimize.load_data(
self.config['datadir'], self.config['datadir'],
pairs=pairs, pairs=pairs,
@ -298,6 +379,7 @@ class Backtesting(object):
timerange=timerange timerange=timerange
) )
ld_files = self.s()
if not data: if not data:
logger.critical("No data found. Terminating.") logger.critical("No data found. Terminating.")
return return
@ -309,6 +391,8 @@ class Backtesting(object):
max_open_trades = 0 max_open_trades = 0
preprocessed = self.tickerdata_to_dataframe(data) preprocessed = self.tickerdata_to_dataframe(data)
t_t = self.f(ld_files)
print("Load from json to file to df in mem took", t_t)
# Print timeframe # Print timeframe
min_date, max_date = self.get_timeframe(preprocessed) min_date, max_date = self.get_timeframe(preprocessed)
@ -332,27 +416,61 @@ class Backtesting(object):
if self.config.get('export', False): if self.config.get('export', False):
self._store_backtest_result(self.config.get('exportfilename'), results) self._store_backtest_result(self.config.get('exportfilename'), results)
logger.info( if self.use_backslap:
'\n' + '=' * 49 + logger.info(
' BACKTESTING REPORT ' + '\n====================================================== '
'=' * 50 + '\n' 'BackSLAP REPORT'
'%s', ' =======================================================\n'
self._generate_text_table( '%s',
data, self._generate_text_table(
results data,
results
)
) )
) # optional print trades
# logger.info( if self.backslap_show_trades:
# results[['sell_reason']].groupby('sell_reason').count() TradesFrame = results.filter(['open_time', 'pair', 'exit_type', 'profit_percent', 'profit_abs',
# ) 'buy_spend', 'sell_take', 'trade_duration', 'close_time'], axis=1)
logger.info( def to_fwf(df, fname):
'\n' + content = tabulate(df.values.tolist(), list(df.columns), floatfmt=".8f", tablefmt='psql')
' SELL READON STATS '.center(119, '=') + print(content)
'\n%s \n',
self._generate_text_table_sell_reason(data, results)
) DataFrame.to_fwf = to_fwf(TradesFrame, "backslap.txt")
# optional save trades
if self.backslap_save_trades:
TradesFrame = results.filter(['open_time', 'pair', 'exit_type', 'profit_percent', 'profit_abs',
'buy_spend', 'sell_take', 'trade_duration', 'close_time'], axis=1)
def to_fwf(df, fname):
content = tabulate(df.values.tolist(), list(df.columns), floatfmt=".8f", tablefmt='psql')
open(fname, "w").write(content)
DataFrame.to_fwf = to_fwf(TradesFrame, "backslap.txt")
else:
logger.info(
'\n================================================= '
'BACKTEST REPORT'
' ==================================================\n'
'%s',
self._generate_text_table(
data,
results
)
)
if 'sell_reason' in results.columns:
logger.info(
'\n' +
' SELL READON STATS '.center(119, '=') +
'\n%s \n',
self._generate_text_table_sell_reason(data, results)
)
else:
logger.info("no sell reasons available!")
logger.info( logger.info(
'\n' + '\n' +
@ -377,7 +495,7 @@ def setup_configuration(args: Namespace) -> Dict[str, Any]:
# Ensure we do not use Exchange credentials # Ensure we do not use Exchange credentials
config['exchange']['key'] = '' config['exchange']['key'] = ''
config['exchange']['secret'] = '' config['exchange']['secret'] = ''
config['backslap'] = args.backslap
if config['stake_amount'] == constants.UNLIMITED_STAKE_AMOUNT: if config['stake_amount'] == constants.UNLIMITED_STAKE_AMOUNT:
raise DependencyException('stake amount could not be "%s" for backtesting' % raise DependencyException('stake amount could not be "%s" for backtesting' %
constants.UNLIMITED_STAKE_AMOUNT) constants.UNLIMITED_STAKE_AMOUNT)

7
requirements.txt
View File

@ -23,3 +23,10 @@ scikit-optimize==0.5.2
# Required for plotting data # Required for plotting data
#plotly==3.0.0 #plotly==3.0.0
# find first, C search in arrays
py_find_1st==1.1.1
#Load ticker files 30% faster
ujson==1.35

2
setup.py
View File

@ -37,6 +37,8 @@ setup(name='freqtrade',
'cachetools', 'cachetools',
'coinmarketcap', 'coinmarketcap',
'scikit-optimize', 'scikit-optimize',
'ujson',
'py_find_1st'
], ],
include_package_data=True, include_package_data=True,
zip_safe=False, zip_safe=False,