Разработка отчетов по бэктестам (P&L, Sharpe, drawdown, win rate)

Разработка отчётов по бэктестам (P&L, Sharpe, drawdown, win rate)

Бэктест без качественной отчётности — набор цифр без смысла. Хороший отчёт позволяет быстро понять сильные и слабые стороны стратегии, сравнивать несколько стратегий между собой и принимать обоснованные решения о дальнейшей разработке.

Структура отчёта

from dataclasses import dataclass
from typing import Optional
import pandas as pd
import numpy as np

@dataclass
class BacktestReport:
    # Сводные метрики
    initial_capital: float
    final_capital: float
    total_return_pct: float
    annual_return_pct: float
    
    # Risk-adjusted
    sharpe_ratio: float
    sortino_ratio: float
    calmar_ratio: float
    
    # Drawdown
    max_drawdown_pct: float
    avg_drawdown_pct: float
    max_drawdown_duration_days: int
    
    # Торговля
    total_trades: int
    win_rate: float
    profit_factor: float
    avg_win_pct: float
    avg_loss_pct: float
    best_trade_pct: float
    worst_trade_pct: float
    avg_trade_duration_hours: float
    
    # Комиссии
    total_commission: float
    commission_as_pct_of_pnl: float
    
    # Временные серии
    equity_curve: pd.Series
    monthly_returns: pd.DataFrame
    trade_list: pd.DataFrame

Расчёт метрик

def compute_all_metrics(equity_curve: pd.Series, trades: list[dict]) -> BacktestReport:
    returns = equity_curve.pct_change().dropna()
    annual_factor = 252

    # Базовые
    total_return = (equity_curve.iloc[-1] / equity_curve.iloc[0]) - 1
    days = (equity_curve.index[-1] - equity_curve.index[0]).days
    annual_return = (1 + total_return) ** (365 / max(days, 1)) - 1

    # Sharpe (risk-free rate = 0 для крипто)
    sharpe = (returns.mean() * annual_factor) / (returns.std() * np.sqrt(annual_factor)) if returns.std() > 0 else 0

    # Sortino (только downside volatility)
    downside = returns[returns < 0].std()
    sortino = (returns.mean() * annual_factor) / (downside * np.sqrt(annual_factor)) if downside > 0 else 0

    # Drawdown
    rolling_max = equity_curve.cummax()
    drawdown_series = (equity_curve - rolling_max) / rolling_max
    max_dd = drawdown_series.min()

    # Max drawdown duration
    in_drawdown = drawdown_series < 0
    dd_start = None
    max_duration = 0
    for date, is_dd in in_drawdown.items():
        if is_dd and dd_start is None:
            dd_start = date
        elif not is_dd and dd_start is not None:
            duration = (date - dd_start).days
            max_duration = max(max_duration, duration)
            dd_start = None

    # Calmar
    calmar = annual_return / abs(max_dd) if max_dd != 0 else 0

    # Trade-level
    trades_df = pd.DataFrame(trades)
    if not trades_df.empty:
        winning = trades_df[trades_df['pnl'] > 0]
        losing = trades_df[trades_df['pnl'] < 0]

        win_rate = len(winning) / len(trades_df)
        gross_profit = winning['pnl'].sum()
        gross_loss = abs(losing['pnl'].sum())
        profit_factor = gross_profit / gross_loss if gross_loss > 0 else float('inf')

        avg_win_pct = (winning['pnl'] / winning['entry_value'] * 100).mean() if not winning.empty else 0
        avg_loss_pct = (losing['pnl'] / losing['entry_value'] * 100).mean() if not losing.empty else 0
    else:
        win_rate = profit_factor = avg_win_pct = avg_loss_pct = 0

    return BacktestReport(
        initial_capital=equity_curve.iloc[0],
        final_capital=equity_curve.iloc[-1],
        total_return_pct=total_return * 100,
        annual_return_pct=annual_return * 100,
        sharpe_ratio=round(sharpe, 3),
        sortino_ratio=round(sortino, 3),
        calmar_ratio=round(calmar, 3),
        max_drawdown_pct=max_dd * 100,
        avg_drawdown_pct=drawdown_series[drawdown_series < 0].mean() * 100,
        max_drawdown_duration_days=max_duration,
        total_trades=len(trades),
        win_rate=win_rate,
        profit_factor=profit_factor,
        avg_win_pct=avg_win_pct,
        avg_loss_pct=avg_loss_pct,
        equity_curve=equity_curve,
        trade_list=trades_df,
    )

Monthly Returns Heatmap

def compute_monthly_returns(equity_curve: pd.Series) -> pd.DataFrame:
    """Создаём матрицу месячных доходностей для heatmap"""
    monthly = equity_curve.resample('ME').last()
    monthly_returns = monthly.pct_change().dropna()

    # Создаём матрицу год × месяц
    matrix = monthly_returns.groupby([
        monthly_returns.index.year,
        monthly_returns.index.month
    ]).first().unstack()

    matrix.columns = ['Jan', 'Feb', 'Mar', 'Apr', 'May', 'Jun',
                       'Jul', 'Aug', 'Sep', 'Oct', 'Nov', 'Dec']
    return matrix * 100  # в процентах

HTML отчёт

def generate_html_report(report: BacktestReport, strategy_name: str) -> str:
    import plotly.graph_objects as go
    from plotly.subplots import make_subplots

    fig = make_subplots(
        rows=3, cols=2,
        subplot_titles=['Equity Curve', 'Drawdown', 'Monthly Returns', 'Trade P&L Distribution', 'Win/Loss', 'Rolling Sharpe'],
    )

    # Equity curve
    fig.add_trace(go.Scatter(x=report.equity_curve.index, y=report.equity_curve.values,
                              name='Portfolio', line=dict(color='#00C853')), row=1, col=1)

    # Drawdown
    rolling_max = report.equity_curve.cummax()
    drawdown = (report.equity_curve - rolling_max) / rolling_max * 100
    fig.add_trace(go.Scatter(x=drawdown.index, y=drawdown.values,
                              fill='tozeroy', name='Drawdown', line=dict(color='#FF5252')), row=1, col=2)

    # P&L distribution
    if not report.trade_list.empty:
        pnl_pct = report.trade_list['pnl'] / report.trade_list['entry_value'] * 100
        fig.add_trace(go.Histogram(x=pnl_pct, name='Trade P&L %', nbinsx=30), row=2, col=1)

    fig.update_layout(
        title=f'Backtest Report: {strategy_name}',
        height=1000,
        showlegend=False,
        template='plotly_dark',
    )

    return fig.to_html(include_plotlyjs='cdn')

Интерпретация ключевых метрик

Важно: высокий win rate без хорошего risk/reward — бесполезен. Стратегия с 80% win rate и средним loss = 5× average win убыточна. Profit Factor и expectancy важнее голого win rate.