Guide

Conditional Drawdown at Risk (CDaR) explained

Harbor Capital's balanced 60/40 sleeve finished 2022 with a maximum drawdown of −14.2% — inside its −15% policy limit. Limited partners still complained: the portfolio spent nine months underwater, with drawdown hovering between −8% and −14% for most of the bear market. MDD captured the single worst trough but ignored how deep and persistent intermediate pain was. When the risk team replayed monthly drawdown paths, the average drawdown on the worst 20% of months exceeded −11% — far worse than the median underwater episode. That statistic is close to what Conditional Drawdown at Risk (CDaR) formalizes: the expected drawdown conditional on being in the tail of the drawdown distribution at a chosen confidence level. Unlike Expected Shortfall (CVaR), which averages return losses beyond a VaR threshold, CDaR operates on the path-dependent equity curve investors actually experience. This guide defines CDaR and Drawdown at Risk (DaR), explains Chekhlov-Uryasev-style optimization as a tractable linear program, walks through Harbor Capital's defensive sleeve refactor, compares CDaR to vol targeting and regime overlays in a technique decision table, lists pitfalls allocators hit, and ends with a production checklist alongside our volatility targeting and regime switching guides.

Why return-based tail risk misses drawdown pain

Daily or monthly returns are the standard input for VaR and CVaR. A portfolio can post modest daily losses while its equity curve slowly grinds from peak to trough over many months — the path investors feel. Two sleeves with identical one-day 99% CVaR can have very different underwater curves: one snaps back in six weeks, the other bleeds for a year.

Maximum drawdown records only the deepest single hole. It is non-diversifiable across time in the sense that one catastrophic episode dominates the statistic. Allocators who cap MDD at −15% learn nothing about whether typical bad periods average −5% or −12% underwater.

Drawdown as a stochastic process

At each time t, track running peak P_t = max_{s ≤ t} V_s on portfolio value V. Instantaneous drawdown is D_t = (V_t − P_t) / P_t (negative while underwater). The drawdown process {D_t} is path-dependent: today's drawdown depends on the entire history of peaks, not just today's return. CDaR asks tail questions about {D_t}, not about {r_t}.

The Ulcer Index integrates squared drawdowns over time — another path-sensitive pain metric. CDaR instead follows the CVaR recipe: pick a confidence level α (often 0.95 or 0.99), identify the worst (1−α) fraction of drawdown observations, and average them.

CDaR and Drawdown at Risk defined

Chekhlov, Uryasev, and Zabarankin (2003) introduced a family of drawdown functionals for portfolio optimization. The building blocks:

Maximum Drawdown (MDD) — MDD = min_t D_t over the horizon (most negative drawdown).
Average Drawdown — mean of {D_t} while underwater or over all t; sensitive to sample frequency.
Drawdown at Risk (DaR_α) — the α-quantile of the drawdown distribution (analogous to VaR on drawdowns). Example: 95% DaR = −8% means drawdown exceeds −8% on 5% of observed days.
Conditional Drawdown at Risk (CDaR_α) — expected drawdown given drawdown is at or beyond DaR_α: CDaR_α = E[D | D ≤ DaR_α] (using the convention that drawdowns are negative). CDaR_α is always at least as deep as DaR_α and captures how bad tail underwater periods are on average.

At α → 1, CDaR approaches MDD. At moderate α (0.90–0.95), CDaR balances depth and frequency of pain — penalizing portfolios that linger moderately underwater, not only those that cliff-dive once.

Sample vs scenario CDaR

Historical CDaR sorts observed drawdowns from a backtest window and averages the worst tail. Scenario CDaR computes drawdown paths on Monte Carlo simulated return paths, then applies the same tail average. Scenario methods need realistic autocorrelation and fat tails; i.i.d. Gaussian daily returns understate CDaR for equity-heavy books.

CDaR optimization: why it is tractable

Minimizing CDaR subject to return or weight constraints looks non-linear because drawdown depends on running peaks. Chekhlov and Uryasev show the problem reformulates as a linear program (LP) or mixed-integer LP when cardinality constraints bind. Practitioners use auxiliary variables for peak value and drawdown at each date; modern solvers (CPLEX, Gurobi, open-source HiGHS) handle thousand-asset instances with monthly rebalancing grids.

Typical objective variants:

Minimize CDaR_α subject to minimum expected return or benchmark tracking error.
Maximize return subject to CDaR_α ≥ −L (tail drawdown budget).
Multi-objective — trade off expected return vs CDaR on an efficient frontier, analogous to mean-variance but with drawdown tails on the risk axis.

CDaR-optimal weights often differ from mean-variance or risk-parity solutions: the optimizer tilts toward assets that recover quickly from underwater episodes, not merely low-volatility names that can still drift down for quarters.

Worked example: Harbor Capital defensive sleeve refactor

Harbor Capital's policy portfolio held 60% global equities and 40% aggregate bonds via ETFs, rebalanced quarterly, with a hard MDD cap of −15% monitored monthly. After 2022, the team added a 95% CDaR budget of −10% on monthly drawdown observations estimated on a ten-year rolling window.

Before CDaR overlay

Static 60/40 weights; crisis response via discretionary cash raises when MDD breached −12%.
2022: MDD −14.2%, median underwater month −6.8%, 95% CDaR (ex post) −11.4%.
Nine consecutive months with drawdown worse than −5%.

CDaR optimization layer

Each quarter, solve: maximize expected return proxy (12-month momentum score on sleeves) subject to CDaR_0.95 ≥ −10% on simulated paths, weight bounds [0%, 70%] per sleeve, and sum-to-100%. Allowed sleeves: equities, bonds, gold, cash. When CDaR constraint binds, the solver cuts equity toward gold and cash even if bonds alone would satisfy vol targeting.

After (2023–2025 backtest, OOS from 2010)

95% CDaR improved from −11.4% to −9.6% on holdout windows; MDD worsened slightly (−14.2% → −15.1%) because CDaR at 95% does not minimize single worst trough.
Average months underwater per episode fell from 7.2 to 4.8.
CAGR sacrifice: ~0.4% vs static 60/40 over OOS period — disclosed as “drawdown insurance premium.”
Interaction rule: CDaR rebalance runs first; vol targeting scales gross exposure second (documented precedence avoids double de-risking).

Harbor paired CDaR with a Markov crisis overlay for tail months where filtered crisis probability exceeded 0.6 — CDaR handles slow grinds, regime filter handles volatility spikes.

Technique decision table

Approach	What it controls	Best when	Watch out for
CDaR optimization	Average tail drawdown depth	LPs care about sustained underwater pain, not single MDD spike	Estimation window; may not cap true worst-case MDD
Maximum drawdown cap	Single worst peak-to-trough	Hard mandate limits, fund terms with MDD triggers	Ignores duration and frequency of moderate drawdowns
CVaR / Expected Shortfall	Tail of return distribution	Trading book, daily risk, regulatory capital	Path-blind; misses slow equity grinds
Volatility targeting	Realized return volatility	CTA sleeves, scalable risk budgets	Vol ≠ drawdown; reactive to recent spikes
Ulcer Index minimization	Integrated squared drawdown	Ranking funds by investor-reported stress	Less standard for optimization; harder LP narrative
Regime switching overlay	Crisis-state allocation shift	Discrete bull/bear/crisis detection	Estimation risk; whipsaw without hysteresis
Calmar / Sterling ratios	Return per unit MDD (reporting)	Fund screening, manager comparison	Single-episode denominator; not a constraint tool

Common pitfalls

Confusing CDaR with CVaR — CVaR on daily returns does not equal CDaR on drawdowns; reporting both without labels confuses risk committees.
Wrong sampling frequency — Daily vs monthly drawdown series produce different DaR/CDaR; align with rebalance cadence and stick to one frequency.
Short estimation windows — Five-year samples miss 2008 or 2020 tail drawdowns; use decade-plus history or stress overlays.
Ignoring MDD entirely — 95% CDaR at −10% still allows a −25% single crash in theory; pair with a hard MDD stop for mandate compliance.
Look-ahead in peaks — Backtests must compute running peaks causally; using full-sample peaks inflates CDaR quality.
Survivorship in constituent returns — Index backtests without delisted names understate historical CDaR (see survivorship bias guide).
Double de-risking — CDaR cuts plus vol targeting plus regime overlay without precedence rules can over-allocate to cash and miss recoveries.
Gaussian scenarios — Monte Carlo CDaR from normal returns underestimates equity tail drawdown persistence; use fat tails or historical bootstrap.

Production checklist

Choose confidence α (0.95 common) and drawdown sampling frequency.
Define estimation window (10y+ rolling) and OOS validation protocol.
Compute historical DaR and CDaR on policy portfolio; document baseline tail pain.
Set CDaR budget (e.g. 95% CDaR ≥ −10%) alongside existing MDD mandate.
Implement LP optimizer with weight bounds, turnover, and sleeve constraints.
Document rule precedence vs vol targeting, TAA, and regime overlays.
Stress-test 2008, 2020, 2022 paths; report MDD, CDaR, and months underwater.
Disclose CAGR sacrifice vs static policy as insurance cost in LP reports.
Monitor live monthly drawdown series; alert when approaching DaR threshold.
Refit annually; log parameter drift and binding constraint frequency.

Key takeaways

CDaR averages drawdown in the tail of the underwater distribution — it captures sustained pain that MDD and CVaR can miss.
Drawdown is path-dependent; CDaR optimization uses LP reformulations to find weights that limit tail underwater episodes.
95% CDaR budgets trade a modest CAGR sacrifice for fewer and shallower prolonged drawdowns — complementary to hard MDD caps.
Harbor Capital's refactor layered CDaR on 60/40 policy, cutting average tail drawdown with documented rule precedence over vol targeting.
Pair CDaR with regime switching for crisis spikes and always validate on long OOS windows with realistic delisting data.