Stock Correlation Analysis Skill

If all dependencies are already installed, skip the install step and proceed directly.

Step 2: Route to the Correct Sub-Skill

Classify the user's request and jump to the matching sub-skill section below.

If ambiguous, default to Sub-Skill A (Co-movement Discovery) for single tickers, or Sub-Skill B (Return Correlation) for two tickers.

Defaults for all sub-skills

Sub-Skill A: Co-movement Discovery

Goal: Given a single ticker, find stocks that move with it.

A1: Build the peer universe

You need 15-30 candidates. Do not use hardcoded ticker lists — build the universe dynamically at runtime. See references/sector_universes.md for the full implementation. The approach:

1. Screen same-industry stocks using yf.screen() + yf.EquityQuery to find stocks in the same industry as the target
2. Broaden to sector if the industry screen returns fewer than 10 peers
3. Add thematic/adjacent industries — read the target's longBusinessSummary and screen 1-2 related industries (e.g., a semiconductor company → also screen semiconductor equipment)
4. Combine, deduplicate, remove target ticker

A2: Compute correlations

import yfinance as yf
import pandas as pd
import numpy as np

def discover_comovement(target_ticker, peer_tickers, period="1y"):
    all_tickers = [target_ticker] + [t for t in peer_tickers if t != target_ticker]
    data = yf.download(all_tickers, period=period, auto_adjust=True, progress=False)

    # Extract close prices — yf.download returns MultiIndex (Price, Ticker) columns
    closes = data["Close"].dropna(axis=1, thresh=max(60, len(data) // 2))

    # Log returns
    returns = np.log(closes / closes.shift(1)).dropna()
    corr_series = returns.corr()[target_ticker].drop(target_ticker, errors="ignore")

    # Rank by absolute correlation
    ranked = corr_series.abs().sort_values(ascending=False)

    result = pd.DataFrame({
        "Ticker": ranked.index,
        "Correlation": [round(corr_series[t], 4) for t in ranked.index],
    })
    return result, returns

A3: Present results

Show a ranked table with company names and sectors (fetch via yf.Ticker(t).info.get("shortName")):

Include:

• Top 10 positively correlated stocks
• Any notable negatively correlated stocks (potential hedges)
• Brief explanation of why each might be linked (sector, supply chain, customer overlap)

Sub-Skill B: Return Correlation

Goal: Deep-dive into the relationship between two (or a few) specific tickers.

B1: Download and compute

import yfinance as yf
import pandas as pd
import numpy as np

def return_correlation(ticker_a, ticker_b, period="1y"):
    data = yf.download([ticker_a, ticker_b], period=period, auto_adjust=True, progress=False)
    closes = data["Close"][[ticker_a, ticker_b]].dropna()

    returns = np.log(closes / closes.shift(1)).dropna()
    corr = returns[ticker_a].corr(returns[ticker_b])

    # Beta: how much does B move per unit move of A
    cov_matrix = returns.cov()
    beta = cov_matrix.loc[ticker_b, ticker_a] / cov_matrix.loc[ticker_a, ticker_a]

    # R-squared
    r_squared = corr ** 2

    # Rolling 60-day correlation for stability
    rolling_corr = returns[ticker_a].rolling(60).corr(returns[ticker_b])

    # Spread (log price ratio) for mean-reversion
    spread = np.log(closes[ticker_a] / closes[ticker_b])
    spread_z = (spread - spread.mean()) / spread.std()

    return {
        "correlation": round(corr, 4),
        "beta": round(beta, 4),
        "r_squared": round(r_squared, 4),
        "rolling_corr_mean": round(rolling_corr.mean(), 4),
        "rolling_corr_std": round(rolling_corr.std(), 4),
        "rolling_corr_min": round(rolling_corr.min(), 4),
        "rolling_corr_max": round(rolling_corr.max(), 4),
        "spread_z_current": round(spread_z.iloc[-1], 4),
        "observations": len(returns),
    }

B2: Present results

Show a summary card:

Interpretation guide:

• Correlation > 0.80: Strong co-movement — these stocks are tightly linked
• Correlation 0.50–0.80: Moderate — shared sector drivers but independent factors too
• Correlation < 0.50: Weak — limited co-movement despite possible sector overlap
• High rolling std: Unstable relationship — correlation varies significantly over time
• Spread Z > |2|: Unusual divergence from historical relationship

Sub-Skill C: Sector Clustering

Goal: Given a group of tickers, show the full correlation structure and identify clusters.

C1: Build the correlation matrix

import yfinance as yf
import pandas as pd
import numpy as np

def sector_clustering(tickers, period="1y"):
    data = yf.download(tickers, period=period, auto_adjust=True, progress=False)

    # yf.download returns MultiIndex (Price, Ticker) columns
    closes = data["Close"].dropna(axis=1, thresh=max(60, len(data) // 2))
    returns = np.log(closes / closes.shift(1)).dropna()
    corr_matrix = returns.corr()

    # Hierarchical clustering order
    from scipy.cluster.hierarchy import linkage, leaves_list
    from scipy.spatial.distance import squareform

    dist_matrix = 1 - corr_matrix.abs()
    np.fill_diagonal(dist_matrix.values, 0)
    condensed = squareform(dist_matrix)
    linkage_matrix = linkage(condensed, method="ward")
    order = leaves_list(linkage_matrix)
    ordered_tickers = [corr_matrix.columns[i] for i in order]

    # Reorder matrix
    clustered = corr_matrix.loc[ordered_tickers, ordered_tickers]

    return clustered, returns

Note: if scipy is not available, fall back to sorting by average correlation instead of hierarchical clustering.

C2: Present results

1. Full correlation matrix — formatted as a table. For more than 8 tickers, show as a heatmap description or highlight only the strongest/weakest pairs.

2. Identified clusters — group tickers that have high intra-group correlation:

   • Cluster 1: [NVDA, AMD, AVGO] — avg intra-correlation 0.82
   • Cluster 2: [AAPL, MSFT] — avg intra-correlation 0.75

3. Outliers — tickers with low average correlation to the group (potential diversifiers).

4. Strongest pairs — top 5 highest-correlation pairs in the matrix.

5. Weakest pairs — top 5 lowest/negative-correlation pairs (hedging candidates).

Sub-Skill D: Realized Correlation

Goal: Show how correlation changes over time and under different market conditions.

D1: Rolling correlation

import yfinance as yf
import pandas as pd
import numpy as np

def realized_correlation(ticker_a, ticker_b, period="2y", windows=[20, 60, 120]):
    data = yf.download([ticker_a, ticker_b], period=period, auto_adjust=True, progress=False)
    closes = data["Close"][[ticker_a, ticker_b]].dropna()

    returns = np.log(closes / closes.shift(1)).dropna()

    rolling = {}
    for w in windows:
        rolling[f"{w}d"] = returns[ticker_a].rolling(w).corr(returns[ticker_b])

    return rolling, returns

D2: Regime-conditional correlation

def regime_correlation(returns, ticker_a, ticker_b, condition_ticker=None):
    """Compare correlation across up/down/volatile regimes."""
    if condition_ticker is None:
        condition_ticker = ticker_a

    ret = returns[condition_ticker]

    regimes = {
        "All Days": pd.Series(True, index=returns.index),
        "Up Days (target > 0)": ret > 0,
        "Down Days (target < 0)": ret < 0,
        "High Vol (top 25%)": ret.abs() > ret.abs().quantile(0.75),
        "Low Vol (bottom 25%)": ret.abs() < ret.abs().quantile(0.25),
        "Large Drawdown (< -2%)": ret < -0.02,
    }

    results = {}
    for name, mask in regimes.items():
        subset = returns[mask]
        if len(subset) >= 20:
            results[name] = {
                "correlation": round(subset[ticker_a].corr(subset[ticker_b]), 4),
                "days": int(mask.sum()),
            }

    return results

D3: Present results

1. Rolling correlation summary table:

2. Regime correlation table:

3. Key insight: Highlight whether correlation increases during sell-offs (very common — "correlations go to 1 in a crisis"). This is critical for risk management.

4. Trend: Is correlation trending higher or lower recently vs. its historical average?

Step 3: Respond to the User

After running the appropriate sub-skill, present results clearly:

Always include

• The lookback period and data interval used
• The number of observations (trading days)
• Any tickers dropped due to insufficient data

Always caveat

• Correlation is not causation — co-movement does not imply a causal link
• Past correlation does not guarantee future correlation — regimes shift
• Short lookback windows produce noisy estimates; longer windows smooth but may miss regime changes

Practical applications (mention when relevant)

• Sympathy plays: Stocks likely to follow a peer's earnings/news move
• Pair trading: High-correlation pairs where the spread has diverged from its mean
• Portfolio diversification: Finding low-correlation assets to reduce risk
• Hedging: Identifying inversely correlated instruments
• Sector rotation: Understanding which sectors move together
• Risk management: Correlation spikes during stress — diversification may fail when needed most

Important: Never recommend specific trades. Present data and let the user draw conclusions.

Reference Files

• references/sector_universes.md — Dynamic peer universe construction using yfinance Screener API

Read the reference file when you need to build a peer universe for a given ticker.