Correlation Aware Exposure Limits

When to Use

Invoke this whenever a portfolio trades multiple instruments within the same sector, asset class, or macro factor (e.g., holding multiple tech stocks NVDA, AMD, MSFT, or crypto assets BTC, ETH, SOL). Setting individual position limits per ticker alone creates hidden concentration risk: when correlated instruments fall in tandem during market sell-offs, total portfolio drawdown spikes unexpectedly. Estimating rolling return correlation matrices ($C_{i,j}$), grouping instruments into correlated clusters ($\rho \ge 0.70$), and capping total cluster exposure (e.g. $\le 30%$ NAV) before approving order executions is mandatory.

When NOT to Use

• As the only pre-trade control. Cluster caps bound concentration, not leverage, margin adequacy, drawdown, or per-symbol size. Compose with the risk skills under Related Skills; SEC Rule 15c3-5 expects a control suite, not a single check.
• When the correlation estimate cannot be trusted. Fewer than ~30 overlapping returns, a newly listed instrument, or a regime break makes the Pearson estimate too noisy to cluster on. The module fails closed on a missing matrix, but it cannot detect a statistically weak one — widen the lookback or fall back to sector-only clustering.
• For linear factor-risk budgeting. Connected-component clustering answers "which names move together enough to share a cap," not "how much of my variance is one factor." Use a factor/covariance model for the latter; a single 0.70 edge can chain a long clustering chain into one pocket.
• As a substitute for netting-aware margin math. Exposure here is deliberately GROSS, so it will not match broker margin, which does grant offsets. Do not drive collateral decisions from these numbers.
• For a single-instrument or deliberately paired mandate, unless caps are set to what the mandate actually authorises.

Prerequisites

• Historical return series for portfolio instruments over rolling lookback window (e.g. 60 days).
• Defined correlation threshold $\rho_{\text{threshold}}$ (default 0.70).
• Defined maximum allowed cluster exposure cap, expressed as an absolute notional (default 1,000,000; convert a NAV-percentage policy to notional as $\text{cap} = \text{pct} \times \text{NAV}$ at construction time).

Workflow

1. Estimate Rolling Correlation Matrix:

   • Compute pairwise Pearson correlation matrix $C$ over historical return vectors $R_1, R_2, \dots, R_K$: $$C_{i,j} = \frac{\text{Cov}(R_i, R_j)}{\sigma_i \cdot \sigma_j}$$
   • Price series must be chronological, positive, finite, and date-aligned at their most recent point; differing lengths are correlated over their most recent overlapping returns. Bad data (zero/negative/NaN prices) is rejected, never silently correlated.

2. Form Correlation Clusters:

   • Group assets into connected clusters $G_1, G_2, \dots, G_m$ where pairwise correlation $C_{i,j} \ge \rho_{\text{threshold}}$ (transitively: A–B and B–C at threshold join A, B, C even if A–C is below). Symbols sharing a sector_mapping label are forced into one cluster regardless of measured correlation — sector co-membership is treated as one risk pocket.

3. Compute Current Cluster Exposures:

   • Calculate current GROSS dollar exposure for cluster $k$ (sum of absolute notionals; netting longs against shorts inside a correlated cluster is deliberately not done, because correlations converge in stress and the hedge fails exactly when it matters): $$\text{Exposure}(G_k) = \sum_{i \in G_k} |w_i \cdot \text{Position}_i|$$ where $w_i$ is the underlying delta (options delta-adjusted to underlying-equivalent exposure). Divide by NAV for a percentage view.

4. Validate Proposed Order Execution:

   • Fail closed: if no correlation matrix has been built, evaluation raises — never approve orders against silently-empty clusters. A stale matrix (default > 7 days) either warns or blocks, per stale_matrix_policy.
   • Check the aggregate portfolio cap on post-trade GROSS notional, counting the proposed leg even when the symbol is not yet in the book — an opening order in a fresh symbol must consume portfolio headroom like any other. This cap is raw notional (not delta-adjusted); the cluster cap below is delta-adjusted.
   • For a proposed signed increment $V_{\text{proposed}}$ on symbol $i$, compute the exact post-trade cluster exposure: $\text{Exposure}(G_k) + |V_{\text{proposed}}|$ for a new position, or $|\text{Position}i + V{\text{proposed}}|$ netting against an existing one — risk-REDUCING orders are never vetoed, even when a cluster is already over cap (they are approved with a remediation flag).
   • If post-trade exposure exceeds the cap and the order does not reduce it, veto with RiskCheckResult(approved=False); allowed_notional carries the indicative downsized size.

Full step-by-step procedure with broker-specific detail: see references/workflows.md. Broker/framework coverage table for this skill: see references/standards.md. Printable pre-flight checklist: see assets/checklist.md.

Common Pitfalls

• Per-Ticker Limit Blind Spot: Assuming single-ticker limits (e.g. 5% per stock) prevent risk concentration across 8 tech stocks (40% total tech exposure).
• Static Correlation Assumptions: Using static historical correlations without updating rolling matrices, missing correlation breakdown during market crashes.
• Fail-Open Risk Gates: Approving orders when the correlation matrix is missing or stale turns every symbol into an uncorrelated singleton and silently disables cluster limits. Missing matrix must block; stale matrix should block in production (stale_matrix_policy="block").
• Vetoing De-Risking Orders: Adding $|V_{\text{proposed}}|$ on top of current exposure vetoes position REDUCTIONS when a cluster sits near its cap. Net the increment against the existing position and only block exposure-increasing orders.
• Asymmetric Delta Treatment: Delta-adjusting the proposed option order but counting existing options at full notional overstates exposure — apply underlying delta weights to both sides.
• Netting Longs Against Shorts in a Cluster: Netted cluster exposure assumes the correlation hedge holds through the crash; gross (sum of absolute) exposure is the conservative basis for concentration caps.
• Misaligned Return Windows: Correlating truncated prefixes of different-length histories compares returns from different dates. Align on the most recent overlapping returns, and reject (don't skip) bad price data.
• Aggregate Cap Blind to New Positions: Computing the post-trade portfolio total by iterating only over symbols already held silently exempts every opening order from the aggregate cap — the book grows past the limit one new symbol at a time while each individual cluster check still passes. Count the proposed leg explicitly when the symbol is absent from the current book.
• Check-Then-Trade Races: two concurrent orders can each pass against the same cap and jointly breach it. The manager serializes its own matrix/audit state, but the caller must serialize check-then-place sequences when orders can arrive from multiple threads.

Verification

• Submit returns for highly correlated assets (NVDA & AMD, $\rho = 0.85$) and verify CorrelationExposureManager groups them into the same cluster.
• Submit proposed order that breaches the cluster notional limit and verify it is vetoed: RiskCheckResult(approved=False) with the indicative allowed_notional for downsizing.
• Verify an opening order in a symbol not yet held is vetoed when it would push post-trade portfolio gross notional past max_portfolio_notional.
• Verify a risk-reducing order on an at-cap cluster is approved, and that evaluating any order before update_correlation_matrix() raises CorrelationMatrixUnavailableError.
• Run unit test suite python -m unittest discover -s skills/correlation-aware-exposure-limits/scripts and confirm 100% pass rate.

Related Skills

• broker-account-margin-call-handling
• survivorship-bias-free-universe-construction
• ensemble-signal-combination-without-overfitting