Merge pull request #5391 from freqtrade/fix/get_analyzed_pair

Fix recently introduced lookahead bias in backtesting
This commit is contained in:
Matthias 2021-08-09 15:25:12 +02:00 committed by GitHub
commit 8bb42a07ce
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 17 additions and 5 deletions

View File

@ -466,6 +466,8 @@ class Backtesting:
for i, pair in enumerate(data):
row_index = indexes[pair]
try:
# Row is treated as "current incomplete candle".
# Buy / sell signals are shifted by 1 to compensate for this.
row = data[pair][row_index]
except IndexError:
# missing Data for one pair at the end.
@ -476,8 +478,8 @@ class Backtesting:
if row[DATE_IDX] > tmp:
continue
row_index += 1
self.dataprovider._set_dataframe_max_index(row_index)
row_index += 1
indexes[pair] = row_index
# without positionstacking, we can only have one open trade per pair.

View File

@ -1,6 +1,7 @@
# pragma pylint: disable=missing-docstring, W0212, line-too-long, C0103, unused-argument
import random
from datetime import timedelta
from pathlib import Path
from unittest.mock import MagicMock, PropertyMock
@ -741,8 +742,13 @@ def test_backtest_alternate_buy_sell(default_conf, fee, mocker, testdatadir):
# 100 buys signals
results = result['results']
assert len(results) == 100
# Cached data should be 200 (no change since required_startup is 0)
assert len(backtesting.dataprovider.get_analyzed_dataframe('UNITTEST/BTC', '1m')[0]) == 200
# Cached data should be 199 (missing 1 candle at the start)
analyzed_df = backtesting.dataprovider.get_analyzed_dataframe('UNITTEST/BTC', '1m')[0]
assert len(analyzed_df) == 199
# Expect last candle to be 1 below end date (as the last candle is assumed as "incomplete"
# during backtesting)
expected_last_candle_date = backtest_conf['end_date'] - timedelta(minutes=1)
assert analyzed_df.iloc[-1]['date'].to_pydatetime() == expected_last_candle_date
# One trade was force-closed at the end
assert len(results.loc[results['is_open']]) == 0
@ -774,6 +780,7 @@ def test_backtest_multi_pair(default_conf, fee, mocker, tres, pair, testdatadir)
data = trim_dictlist(data, -500)
# Remove data for one pair from the beginning of the data
if tres > 0:
data[pair] = data[pair][tres:].reset_index()
default_conf['timeframe'] = '5m'
@ -800,8 +807,11 @@ def test_backtest_multi_pair(default_conf, fee, mocker, tres, pair, testdatadir)
assert len(evaluate_result_multi(results['results'], '5m', 3)) == 0
# Cached data correctly removed amounts
removed_candles = len(data[pair]) - 1 - backtesting.strategy.startup_candle_count
offset = 2 if tres == 0 else 1
removed_candles = len(data[pair]) - offset - backtesting.strategy.startup_candle_count
assert len(backtesting.dataprovider.get_analyzed_dataframe(pair, '5m')[0]) == removed_candles
assert len(backtesting.dataprovider.get_analyzed_dataframe(
'NXT/BTC', '5m')[0]) == len(data['NXT/BTC']) - 2 - backtesting.strategy.startup_candle_count
backtest_conf = {
'processed': processed,