Indicator Design Skill

Help the user turn a discretionary trading hypothesis into a vectorized, testable indicator.

Flow

1. Extract the hypothesis

Ask:

• What pattern are you trying to capture in one sentence?
• When is the pattern "on"? What market condition?
• What is the expected return when the pattern fires?
• Roughly how often does it fire historically?

2. Operationalize

Convert verbal description to measurable quantities:

• "High volume" → volume > rolling_quantile(volume, 90, window=20)
• "Breakout" → close > rolling_max(high, window=N).shift(1)
• "Exhaustion" → rsi(close, 14) > 80 and rsi(close, 14).shift(1) > 80

Push back on fuzzy language. Ask for concrete thresholds.

3. Implement — vectorized, no loops

import pandas as pd

def indicator(df: pd.DataFrame, window: int = 20, quantile: float = 0.9) -> pd.Series:
    """One-liner describing what this indicator outputs (signal series, bool series, score)."""
    # implementation — pure pandas, no .apply, no Python loops
    ...
    return signal

Rules:

• Pure pandas operations. No for loops over bars.
• Return pd.Series aligned with input index.
• Use .shift(1) for any rolling computation used at decision time — no look-ahead.
• Document whether the signal is {-1, 0, 1}, probability [0, 1], or raw score.

4. Validate

Before the user tests it, check:

• Does the signal fire at the expected frequency?
• Is it well-distributed across regimes?
• Does it correlate with known factors (trend, vol, momentum)?

Generate this sanity-check code:

print(f"fires: {signal.sum()} of {len(signal)} bars ({signal.mean():.2%})")
print(f"by year: {signal.groupby(df.index.year).mean()}")
print(f"corr with returns: {signal.corr(df['close'].pct_change().shift(-1))}")

5. Next step

Suggest the user feed the signal to backtest-review skill next.

Anti-patterns to reject

• "Use ML to detect the pattern" — no. Operationalize first, test edge, THEN consider ML.
• "Let the algorithm find the best threshold" — no. Threshold per hypothesis, test out-of-sample.
• Indicators with >3 parameters — push back, ask to simplify.
• Indicators combining >4 conditions with AND — almost guaranteed overfit; ask what each condition adds.