Risk report

risk-report

You hand over a book — either inline weights or a positions JSON — and the skill returns the empirical risk picture: how volatile this book has been, how it co-moves with the benchmark, how bad the tail gets (VaR + Expected Shortfall), how deep the recent drawdown was, which historical days hurt most and which names did the damage, and which positions are doing the heavy lifting in the variance budget.

This is descriptive risk math on a current book. The script does NOT predict future returns. It tells you what the last N days of history say about how a book like the one you have just handed over has behaved.

When to invoke

• PM needs the daily risk snapshot on a current book
• Risk officer running a tail-risk review
• Researcher comparing book risk vs a benchmark
• Pre-trade check: "what does this proposed book look like?"
• Post-portfolio-mark follow-up: "we know what the book is worth; what could it lose?"

What you need

• A book: either inline --positions T=w,T=w,... or a --book book.json (see examples/sample-book.json)
• MASSIVE_API_KEY exported

Optional:

• Benchmark ticker (default SPY) for beta + tracking error + R²
• Lookback window in trading days (default 252)
• VaR confidence levels (default 0.95,0.99)
• Number of worst historical days to surface (default 5)
• Vol estimator (default realized; ewma for RiskMetrics EWMA with configurable λ, default 0.94, that responds faster to recent regime shifts)

What you get back

Two output layers from one analysis.

Layer 1: canonical JSON matching output-schema.json. Per-book stats (vol, return, Sharpe, beta, alpha, tracking error, correlation, R²). A full VaR block keyed by each requested confidence level with historical VaR, parametric VaR, historical ES, parametric ES. Max drawdown with peak/trough/duration/recovery. Worst-N stress days with per-name loss attribution. Per-position variance contribution and per-position beta to the benchmark. Concentration: top-1/3/5 weights, Herfindahl, effective N. tier_caveats for excluded names and the always-on methodology warnings.

Layer 2: rendered PM report. Header line with the book and lookback, a Portfolio statistics block, a VaR table (rows: historical, parametric, ES — columns: each confidence), one-liner max drawdown, the worst-N stress days with per-name attribution and the benchmark's return on the same date, a Position contribution table sorted by variance share, a Concentration line, then the adaptive Take. The Take reads what's actually true about the book (high beta, concentrated, big drawdown, low tracking error) and surfaces 2-3 of the most striking facts in plain English. See references/ for the full methodology.

How it works

1. Parse the book. Inline --positions (weights sum to ≤ 1.0; the residual is implicit cash) or --book JSON. The JSON format supports either weight per position or shares + price, in which case weights are computed from the value share.
2. Pull daily aggs per position and per benchmark over --lookback-days (default 252). Massive's /v2/aggs/ticker/{T}/range/1/day/{from}/{to}?adjusted=true so dividends and splits don't contaminate the vol estimate.
3. Compute log returns close-to-close. Align all series to the intersection of date indices so every metric reads from the same panel.
4. Drop short series. A position with fewer than 60 aligned trading days is excluded (surfaced in tier_caveats and positions_excluded); the equivalent weight folds into the cash bucket so the math stays consistent.
5. Per-name annualized vol via np.std(daily_returns, ddof=1) * sqrt(252).
6. Pairwise Pearson correlation matrix, 5%-shrunk toward identity for numerical safety (same pattern as position-sizer). Covariance from per-name vols × the shrunk correlation.
7. Portfolio daily returns as the weighted sum across names per day. This is the single time series every metric reads from.
8. Stats: annualized vol, annualized mean return, naïve Sharpe; beta + alpha + tracking error + correlation + R² vs the benchmark.
9. VaR + ES at each confidence:
   • Historical VaR = -percentile(returns, 100 × (1 - confidence))
   • Parametric VaR = z × σ - μ, where z = Φ⁻¹(confidence)
   • Historical ES = -mean(returns ≤ VaR threshold)
   • Parametric ES = -(μ - σ × φ(z) / (1 - confidence)) (Gaussian) See references/var-and-es.md.
10. Max drawdown on the cumulative NAV (reconstructed from log returns). Returns peak, trough, duration, and whether the series recovered to the peak. See references/max-drawdown.md.
11. Worst-N stress picks the N most-negative days in the portfolio return series and attributes each day's loss to individual names via w_i × r_i_on_that_day. See references/stress-scenarios.md.
12. Position variance contributions via the MRC_i = w_i × (Σw)_i decomposition normalized to sum to 1.
13. Concentration stats: top-1/3/5, Herfindahl (Σw²), effective N (1 / HHI). See references/concentration.md.
14. Adaptive Take. Reads beta, top variance contributor share, Herfindahl, drawdown, and tracking error. Surfaces the 2-3 that actually matter for this book, in plain English.

Foundations used

• massive-api-patterns for REST auth and rate limiting on the daily aggs pull.

Output mode: report

The render is a top-to-bottom report — header, stats block, VaR table, drawdown one-liner, stress days, position table, concentration line, take, caveats. Closer to a daily risk note than a pure data dump, because the PM/risk audience needs the take alongside the numbers. See references/ for the per-section methodology.

MC mode

N/A. risk-report uses empirical historical distributions; it does not run Monte Carlo. For distribution-of-outcomes sweeps see valuation-sanity-check --mc.

Endpoints used

• GET /v2/aggs/ticker/{ticker}/range/1/day/{from}/{to}?adjusted=true Daily aggregates per name + per benchmark. One call per ticker per run; results cached per ticker.

Verify endpoint paths against current docs at massive.com/docs before shipping; field names and versions shift.

Doesn't handle (yet)

• Single-regime lookback. All math reads from one window. A multi-window comparison (60-day vs 252-day VaR) would show regime sensitivity; queued.
• No factor decomposition. Variance contribution is at the position level, not the factor level. A factor-attributed decomposition (size, value, momentum, quality) needs the factor-research machinery wired in; obvious follow-up.
• Cornish-Fisher / Student-t parametric VaR. Parametric VaR assumes normality; surfaces in caveats. A higher-moment variant is a clean PR.
• No conditional / stressed-period VaR. Worst-N is the closest thing; a regime-conditional VaR (e.g., compute VaR only on days when SPY was down) is a follow-up.
• No correlation-shock stress. "What if every correlation goes to 0.9" is queued; the worst-N empirical stress carries the day for v1.
• Single benchmark. Beta is vs one ticker. Multi-benchmark (SPY + IWM + EFA + AGG) needs a multi-regression; queued.

These are clean PR extensions. The output schema reserves space for each so adding them later doesn't break consumers.