backtesting-py-oracle（backtesting PY Oracle 搜索）

backtesting.py Oracle Validation for Range Bar Patterns

Configuration and anti-patterns for using backtesting.py to validate ClickHouse SQL sweep results. Ensures bit-atomic replicability between SQL and Python trade evaluation.

Companion skills: clickhouse-antipatterns (SQL correctness, AP-16) | sweep-methodology (sweep design) | rangebar-eval-metrics (evaluation metrics)

Validated: Gen600 oracle verification (2026-02-12) — 3 assets, 5 gates, ALL PASS.

Self-Evolving Skill: This skill improves through use. If instructions are wrong, parameters drifted, or a workaround was needed — fix this file immediately, don't defer. Only update for real, reproducible issues.

Critical Configuration (NEVER omit)

from backtesting import Backtest

bt = Backtest(
    df,
    Strategy,
    cash=100_000,
    commission=0,
    hedging=True,           # REQUIRED: Multiple concurrent positions
    exclusive_orders=False,  # REQUIRED: Don't auto-close on new signal
)

Why: SQL evaluates each signal independently (overlapping trades allowed). Without hedging=True, backtesting.py skips signals while a position is open, producing fewer trades than SQL. This was discovered when SOLUSDT produced 105 Python trades vs 121 SQL trades — 16 signals were silently skipped.

Anti-Patterns (Ordered by Severity)

BP-01: Missing Multi-Position Mode (CRITICAL)

Symptom: Python produces fewer trades than SQL. Gate 1 (signal count) fails.

Root Cause: Default exclusive_orders=True prevents opening new positions while one is active.

Fix: Always use hedging=True, exclusive_orders=False.

BP-02: ExitTime Sort Order (CRITICAL)

Symptom: Entry prices appear mismatched (Gate 3 fails) even though both SQL and Python use the same price source.

Root Cause: stats._trades is sorted by ExitTime, not EntryTime. When overlapping trades exit in a different order than they entered, trade[i] no longer maps to signal[i].

Fix:

trades = stats._trades.sort_values("EntryTime").reset_index(drop=True)

BP-03: NaN Poisoning in Rolling Quantile (CRITICAL)

Symptom: Cross-asset tests fail with far fewer Python trades. Feature quantile becomes NaN and propagates forward indefinitely.

Root Cause: np.percentile with NaN inputs returns NaN. If even one NaN feature value enters the rolling window, all subsequent quantiles become NaN, making all subsequent filter comparisons fail.

Fix: Skip NaN values when building the signal window:

def _rolling_quantile_on_signals(feature_arr, is_signal_arr, quantile_pct, window=1000):
    result = np.full(len(feature_arr), np.nan)
    signal_values = []
    for i in range(len(feature_arr)):
        if is_signal_arr[i]:
            if len(signal_values) > 0:
                window_data = signal_values[-window:]
                result[i] = np.percentile(window_data, quantile_pct * 100)
            # Only append non-NaN values (matches SQL quantileExactExclusive NULL handling)
            if not np.isnan(feature_arr[i]):
                signal_values.append(feature_arr[i])
    return result

BP-04: Data Range Mismatch (MODERATE)

Symptom: Different signal counts between SQL and Python for assets with early data (BNB, XRP).

Root Cause: load_range_bars() defaults to start='2020-01-01' but SQL has no lower bound.

Fix: Always pass start='2017-01-01' to cover all available data.

BP-05: Margin Exhaustion with Overlapping Positions (MODERATE)

Symptom: Orders canceled with insufficient margin. Fewer trades than expected.

Root Cause: With hedging=True and default full-equity sizing, overlapping positions exhaust available margin.

Fix: Use fixed fractional sizing:

self.buy(size=0.01)  # 1% equity per trade

BP-06: Signal Timestamp vs Entry Timestamp (LOW)

Symptom: Gate 2 (timestamp match) fails because SQL uses signal bar timestamps while Python uses entry bar timestamps.

Root Cause: SQL outputs the signal detection bar's timestamp_ms. Python's EntryTime is the fill bar (next bar after signal). These differ by 1 bar.

Fix: Record signal bar timestamps in the strategy's next() method:

# Before calling self.buy()
self._signal_timestamps.append(int(self.data.index[-1].timestamp() * 1000))

5-Gate Oracle Validation Framework

Expected residual: 1-2 exit type mismatches per asset at TIME barrier boundary (bar 50). SQL uses fwd_closes[max_bars], backtesting.py closes at current bar price. Impact on Kelly < 0.006.

Strategy Architecture: Single vs Multi-Position

Multi-Position Strategy Template

class Gen600Strategy(Strategy):
    def next(self):
        current_bar = len(self.data) - 1

        # 1. Register newly filled trades and set barriers
        for trade in self.trades:
            tid = id(trade)
            if tid not in self._known_trades:
                self._known_trades.add(tid)
                self._trade_entry_bar[tid] = current_bar
                actual_entry = trade.entry_price
                if self.tp_mult > 0:
                    trade.tp = actual_entry * (1.0 + self.tp_mult * self.threshold_pct)
                if self.sl_mult > 0:
                    trade.sl = actual_entry * (1.0 - self.sl_mult * self.threshold_pct)

        # 2. Check time barrier for each open trade
        for trade in list(self.trades):
            tid = id(trade)
            entry_bar = self._trade_entry_bar.get(tid, current_bar)
            if self.max_bars > 0 and (current_bar - entry_bar) >= self.max_bars:
                trade.close()
                self._trade_entry_bar.pop(tid, None)

        # 3. Check for new signal (no position guard — overlapping allowed)
        if self._is_signal[current_bar]:
            self.buy(size=0.01)

Data Loading

from data_loader import load_range_bars

df = load_range_bars(
    symbol="SOLUSDT",
    threshold=1000,
    start="2017-01-01",      # Cover all available data
    end="2025-02-05",        # Match SQL cutoff
    extra_columns=["volume_per_trade", "lookback_price_range"],  # Gen600 features
)

HTML Equity Plot: Y-Axis Auto-Fit on Zoom (BP-07 to BP-10)

backtesting.py generates Bokeh HTML plots via bt.plot(). By default, the y-axis is fixed — zooming on the x-axis does NOT rescale the y-axis to fit visible data. This makes it impossible to inspect zoomed-in regions of equity curves that span 4+ orders of magnitude.

Reference implementation: scripts/gen800/plotting.py in opendeviationbar-patterns.

BP-07: NEVER Use LogScale for Equity (CRITICAL)

Symptom: Equity panel y-axis doesn't auto-fit on zoom. JS callbacks fail silently.

Root Cause: Bokeh's LogScale breaks CustomJS y-range callbacks in multi-panel linked x_range layouts. The js_on_change callback fires but y_range.start/y_range.end assignments are ignored by the LogScale renderer.

Proven via POC: LogScale + JS callback works in single-panel mode but fails when 5 panels share a linked x_range.

Fix: Transform equity data to log10() in Python, display on linear scale with a CustomJSTickFormatter that shows 10^tick as readable values:

import numpy as np
from bokeh.models import CustomJSTickFormatter, Range1d

# Transform data
raw = np.asarray(src.data["equity"], dtype=float)
raw = np.where(raw > 0, raw, 1e-10)
src.data["equity"] = np.log10(raw).tolist()

# Custom tick formatter (shows 1%, 100%, 10K%, 1M%)
child.yaxis[0].formatter = CustomJSTickFormatter(code="""
    const v = Math.pow(10, tick);
    if (v >= 1e6) return (v/1e6).toFixed(1) + 'M%';
    if (v >= 1e3) return (v/1e3).toFixed(0) + 'K%';
    if (v >= 1) return v.toFixed(0) + '%';
    return v.toFixed(2) + '%';
""")

# MUST set initial y_range explicitly (backtesting.py leaves it as NaN)
valid = eq[np.isfinite(eq)]
pad = (valid.max() - valid.min()) * 0.05
child.y_range = Range1d(start=valid.min() - pad, end=valid.max() + pad)

BP-08: Panel-Aware Column Matching (CRITICAL)

Symptom: Drawdown/P&L panels go blank when zooming. Shows millions of percent.

Root Cause: Multiple panels share the same ColumnDataSource (backtesting.py optimization). The Equity panel's Line renderer and the Drawdown panel's Line renderer both reference a source containing equity, drawdown, High, Low, etc. A naive "find first y column" approach picks equity for the Drawdown panel.

Fix: Match by panel_label FIRST, then by column name:

panel_label = child.yaxis[0].axis_label or ""

# Label-first matching (panels share data sources!)
if "Drawdown" in panel_label and "drawdown" in d:
    hi_col = lo_col = "drawdown"
elif "Equity" in panel_label and "equity" in d:
    hi_col = lo_col = "equity"
elif "High" in d and "Low" in d:  # OHLC
    hi_col, lo_col = "High", "Low"

BP-09: JS Callback Pattern for Y-Axis Auto-Fit (PROVEN)

Attach a CustomJS callback to x_range.js_on_change for each panel. The callback scans visible data and sets y_range directly:

from bokeh.models import CustomJS

cb = CustomJS(
    args={"source": src, "yr": child.y_range,
          "x_col": "index", "y_col": "equity"},
    code="""
    const xs = source.data[x_col];
    const ys = source.data[y_col];
    const x0 = cb_obj.start, x1 = cb_obj.end;
    let lo = Infinity, hi = -Infinity;
    for (let i = 0; i < xs.length; i++) {
        if (xs[i] >= x0 && xs[i] <= x1 && isFinite(ys[i])) {
            if (ys[i] < lo) lo = ys[i];
            if (ys[i] > hi) hi = ys[i];
        }
    }
    if (isFinite(lo) && isFinite(hi) && lo < hi) {
        const pad = (hi - lo) * 0.05 || 0.001;
        yr.start = lo - pad;
        yr.end = hi + pad;
    }
    """,
)
child.x_range.js_on_change("start", cb)
child.x_range.js_on_change("end", cb)

BP-10: DataRange1d(only_visible=True) Is Unreliable

Symptom: DataRange1d(only_visible=True) works for simple cases but fails with:

• LogScale panels (y-range computed in wrong space)
• VBar renderers (candlesticks — bounds not tracked correctly)
• Shared data sources across panels

Fix: Always use the JS callback pattern (BP-09) instead of DataRange1d(only_visible=True).

Panel Data Source Reference (backtesting.py internals)

Project Artifacts (rangebar-patterns repo)

Post-Execution Reflection

After this skill completes, check before closing:

1. Did the command succeed? — If not, fix the instruction or error table that caused the failure.
2. Did parameters or output change? — If the underlying tool's interface drifted, update Usage examples and Parameters table to match.
3. Was a workaround needed? — If you had to improvise (different flags, extra steps), update this SKILL.md so the next invocation doesn't need the same workaround.

Only update if the issue is real and reproducible — not speculative.