Plotted data and printed stats don't match

[Pilipets]

Max drawdown duration is 273 days from the plot, but 248 days from the stats.

[kernc]

The issue is not observed in the simple examples that accompany the distribution.

Are you, perchance, using more than:

backtesting.py/backtesting/_plotting.py

Line 92 in 0a76e96

_MAX_CANDLES = 10_000

bars of input data, with some downsampling (bt.plot(resample=)) taking place?

There are indeed two different code paths: stats:

backtesting.py/backtesting/backtesting.py

Line 1594 in 0a76e96

s.loc['Max. Drawdown Duration'] = _round_timedelta(dd_dur.max())

using:

backtesting.py/backtesting/backtesting.py

Lines 1493 to 1513 in 0a76e96

	@staticmethod
	def _compute_drawdown_duration_peaks(dd: pd.Series):
	iloc = np.unique(np.r_[(dd == 0).values.nonzero()[0], len(dd) - 1])
	iloc = pd.Series(iloc, index=dd.index[iloc])
	df = iloc.to_frame('iloc').assign(prev=iloc.shift())
	df = df[df['iloc'] > df['prev'] + 1].astype(int)
	# If no drawdown since no trade, avoid below for pandas sake and return nan series
	if not len(df):
	return (dd.replace(0, np.nan),) * 2
	df['duration'] = df['iloc'].map(dd.index.__getitem__) - df['prev'].map(dd.index.__getitem__)
	df['peak_dd'] = df.apply(lambda row: dd.iloc[row['prev']:row['iloc'] + 1].max(), axis=1)
	df = df.reindex(dd.index)
	return df['duration'], df['peak_dd']
	def _compute_stats(self, broker: _Broker, strategy: Strategy) -> pd.Series:
	data = self._data
	index = data.index
	equity = pd.Series(broker._equity).bfill().fillna(broker._cash).values
	dd = 1 - equity / np.maximum.accumulate(equity)
	dd_dur, dd_peaks = self._compute_drawdown_duration_peaks(pd.Series(dd, index=index))

And plot:

backtesting.py/backtesting/_plotting.py

Lines 285 to 294 in 0a76e96

	dd_end = equity_data['DrawdownDuration'].idxmax()
	if np.isnan(dd_end):
	dd_start = dd_end = equity.index[0]
	else:
	dd_start = equity[:dd_end].idxmax()
	# If DD not extending into the future, get exact point of intersection with equity
	if dd_end != equity.index[-1]:
	dd_end = np.interp(equity[dd_start],
	(equity[dd_end - 1], equity[dd_end]),
	(dd_end - 1, dd_end))

backtesting.py/backtesting/_plotting.py

Lines 362 to 367 in 0a76e96

	dd_timedelta_label = df['datetime'].iloc[int(round(dd_end))] - df['datetime'].iloc[dd_start]
	fig.line([dd_start, dd_end], equity.iloc[dd_start],
	line_color='red', line_width=2,
	legend_label=f'Max Dd Dur. ({dd_timedelta_label})'
	.replace(' 00:00:00', '')
	.replace('(0 days ', '('))

I'd consider stats value the more reliable one.

Appreciate any further investigation you can do.

What version is this anyway? Something similar had been fixed in #162.

[Pilipets]

Here is the version Backtesting==0.3.1
I was using a CSV file with 1 hour ohlc data - around 10k rows.

Thanks for your quick reply - I understand that the issue needs to be investigated on my side, but I created it in case anyone else will experience such.

[kernc]

It's most likely that the issue is a rounding error that stems from plot resampling (Backtest.plot(resample=)). Plotting resamples very early in its pipeline:

backtesting.py/backtesting/_plotting.py

Lines 186 to 189 in 0a76e96

	# Limit data to max_candles
	if is_datetime_index:
	df, indicators, equity_data, trades = _maybe_resample_data(
	resample, df, indicators, equity_data, trades)

[Pilipets]

It's most likely that the issue is a rounding error that stems from plot resampling (Backtest.plot(resample=)). Plotting resamples very early in its pipeline:

backtesting.py/backtesting/_plotting.py

Lines 186 to 189 in 0a76e96

	# Limit data to max_candles
	if is_datetime_index:
	df, indicators, equity_data, trades = _maybe_resample_data(
	resample, df, indicators, equity_data, trades)

backtest = Backtest(df, *, cash=*.cash, commission=.002, hedging=True)
print(backtest.run())
backtest.plot(resample=True, filename="temp2.html")

I'm using resampling as above, where len(df) == 7825.

Alright, as far as I understood from your response, this is expected behavior due to the resampling taking place. Let me know if I should close the issue.

[Pilipets]

But why trades amount is different as well - 68 vs 175?
This is a larger dataset, but is it because of the resampling as well?

[kernc]

I'm fairly confident it's due to brefore-plot resampling. See how trades are resampled and weighted aggregated on the bottom here:

backtesting.py/backtesting/_plotting.py

Lines 92 to 146 in 0a76e96

	_MAX_CANDLES = 10_000
	def _maybe_resample_data(resample_rule, df, indicators, equity_data, trades):
	if isinstance(resample_rule, str):
	freq = resample_rule
	else:
	if resample_rule is False or len(df) <= _MAX_CANDLES:
	return df, indicators, equity_data, trades
	from_index = dict(day=-2, hour=-6, minute=1, second=0, millisecond=0,
	microsecond=0, nanosecond=0)[df.index.resolution]
	FREQS = ('1T', '5T', '10T', '15T', '30T', '1H', '2H', '4H', '8H', '1D', '1W', '1M')
	freq = next((f for f in FREQS[from_index:]
	if len(df.resample(f)) <= _MAX_CANDLES), FREQS[-1])
	warnings.warn(f"Data contains too many candlesticks to plot; downsampling to {freq!r}. "
	"See `Backtest.plot(resample=...)`")
	from .lib import OHLCV_AGG, TRADES_AGG, _EQUITY_AGG
	df = df.resample(freq, label='right').agg(OHLCV_AGG).dropna()
	indicators = [_Indicator(i.df.resample(freq, label='right').mean()
	.dropna().reindex(df.index).values.T,
	**dict(i._opts, name=i.name,
	# Replace saved index with the resampled one
	index=df.index))
	for i in indicators]
	assert not indicators or indicators[0].df.index.equals(df.index)
	equity_data = equity_data.resample(freq, label='right').agg(_EQUITY_AGG).dropna(how='all')
	assert equity_data.index.equals(df.index)
	def _weighted_returns(s, trades=trades):
	df = trades.loc[s.index]
	return ((df['Size'].abs() * df['ReturnPct']) / df['Size'].abs().sum()).sum()
	def _group_trades(column):
	def f(s, new_index=df.index.astype(np.int64), bars=trades[column]):
	if s.size:
	# Via int64 because on pandas recently broken datetime
	mean_time = int(bars.loc[s.index].view('i8').mean())
	new_bar_idx = new_index.get_loc(mean_time, method='nearest')
	return new_bar_idx
	return f
	if len(trades): # Avoid pandas "resampling on Int64 index" error
	trades = trades.assign(count=1).resample(freq, on='ExitTime', label='right').agg(dict(
	TRADES_AGG,
	ReturnPct=_weighted_returns,
	count='sum',
	EntryBar=_group_trades('EntryTime'),
	ExitBar=_group_trades('ExitTime'),
	)).dropna()
	return df, indicators, equity_data, trades

Plot trade count is then inferred simply as length of the data frame:

backtesting.py/backtesting/_plotting.py

Lines 478 to 481 in 0a76e96

	fig_ohlc.multi_line(xs='position_lines_xs', ys='position_lines_ys',
	source=trade_source, line_color=trades_cmap,
	legend_label=f'Trades ({len(trades)})',
	line_width=8, line_alpha=1, line_dash='dotted')

Improvements are always welcome! 😁

If you avoid resampling by running:

backtest.plot(resample=False, filename="temp2.html")

or:

backtest.plot(resample=1_000_000, filename="temp2.html")  # Some large number

you should see values more aligned with those expected.