How Big a Drawdown Can You Endure?
Emlyn Flint, Florence Chikurunhe and Luke van Schaik have another excellent paper out, Understanding Drawdowns: Evidence, Models and Decision Rules.
Their question is: Given a drawdown of X%, what is the probability of recovering to the previous high-water mark (HWM) within a specified period?
My earlier KnowRisk calculator asked a related but different question. Given a drawdown at some point over the next ten years, what subsequent return is required to still achieve the original ten-year target CAGR, set by default at 7.5%?
Flint et al. introduce a probabilistic recovery frontier using a regime-switching model. I've adapted that idea to address another important question:
How likely is it that the required recovery return will actually be achieved?
The updated widget compares three approaches:
1. Geometric Brownian Motion (GBM), with a constant assumed return and volatility.
2. Markov-Switching AR(1) (MSAR1), allowing returns and volatility to differ between calm and volatile regimes, adapted from Flint et al.
3. Historical Block Bootstrap (BLOCK), resampling actual returns in blocks to preserve some of their dependence.
The default example of four years of 14% annual growth, followed by a 35% drawdown over the fifth year, leaving five years to achieve the original 7.5% ten-year CAGR.
The required recovery CAGR is 13.43% annually. But how likely is that? (The simulation results below use 4,000 paths with a fixed seed.)
GBM: 35.2% chance of meeting the target; 69.8% of regaining the HWM.
MSAR1: 25.6% and 56.9% respectively.
BLOCK: 29.6% and 60.2% respectively.
For context, GBM assumes 10% forward geometric CAGR and 18% volatility. MSAR1 uses illustrative regime assumptions; BLOCK uses MSCI ACWI monthly returns from December 2006 to October 2026.
These are scenario probabilities, not calibrated forecasts. Importantly, the models have different underlying expected returns, so the gap between them should not be taken as evidence that regime switching alone improves the result.
The widget also calculates the maximum tolerable drawdown for a specified 50%, 75% or 90% chance of achieving the original return objective, allowing users to vary both the timing and magnitude of the loss.
The central insight is that recovering your previous high-water mark and recovering your original investment objective are two very different things.
Try the calculator tool here.
Appendix
User-adjustable regime assumptions
A note on the regime assumptions. Flint et al. calibrate their MS-AR(1) model to South African equity and bond markets. Our widget uses two illustrative regimes: calm, with 14% annual CAGR and 12% volatility; and volatile, with −12% CAGR and 32% volatility. The assumed average durations are six and two months, with a 25% starting probability of the volatile state. Users can change these, including the return-persistence settings, by selecting Regime-switching AR(1) in the Method menu and expanding Regime model assumptions.
The volatile-state return and volatility are close to Flint et al.'s fitted South African equity estimates (−11.8% and 30.5%). Our 25% starting volatile-state probability also happens to equal the long-run share implied by our default durations, close to Flint et al.'s fitted long-run volatile share of 25.6%. These are not the same type of probability in general, and the remaining settings are NOT their calibration. In particular, Flint et al. estimate average durations of 68 and 23 trading days (about three and one months), and their AR(1) coefficient of around 0.04 is daily, whereas ours operates monthly. The widget does not automatically infer or fit market regimes from uploaded returns.
Statistical qualification
A further modelling qualification. The three scenarios have different implied average returns: GBM assumes 10% geometric CAGR, the ACWI sample has delivered approximately 8.4%, and the default MSAR1 regime mix implies approximately 6.9% before simulation effects. At five years, most of the difference in P(target) is attributable to those differing assumptions rather than the switching mechanism itself.
The revised widget therefore adds a GBM comparison matched to each simulation's terminal mean and variance. For the default five-year MSAR1 example, that gives 25.7% versus 25.6% under MSAR1. For BLOCK it is about 27.8% versus 29.6%. These compare terminal probabilities, not matched drawdown paths or first passage, and are not proof of forecasting superiority.
With 4,000 simulated paths, differences smaller than about one percentage point are easily explained by simulation noise. GBM also monitors recovery to the old peak continuously, while MSAR1 and BLOCK check month ends. For more stable results, choose 10,000 paths, recognising that the estimates will change from the published 4,000-path figures. Probabilities are displayed to one decimal place.
Reference regime calibrations
Comparison for context only. Flint et al. use South African DAILY MS-AR(1); the other cited studies use US MONTHLY regime models. Return definitions and sampling periods differ.
Sources: Flint, Chikurunhe & van Schaik (2026), Understanding Drawdowns, Table 7 (uploaded research PDF); American Academy of Actuaries (December 2002), Report on C-3 Phase 2, Table 1; Feldman & Yang (August 2025), The Anatomy of Lost Stock Market Decades, Table 15 (return-based model).
Implementation notes
Editorial and implementation notes (not part of the LinkedIn post)
• Add the live KnowRisk widget URL before publishing. The draft deliberately contains no dummy link.
• The main post ends at the Paper URL. The regime and methodological notes are separate LinkedIn comments, not part of its character count.
• Published MSAR1 and BLOCK percentages are from 4,000 simulations at fixed seed 20261114; raising paths to 10,000 will change the displayed results. The GBM probabilities are analytic.
• The 2002 American Academy of Actuaries and 2025 Feldman/Yang research parameters have now been checked against their source documents. Figures previously attributed to Berberovic & Eriksson have been omitted pending independent verification.
• Workbook external links: Probability Engine B4/B5/B6/B8/B10 were incorrectly described in an earlier audit. In the checked FINAL(2).xlsm workbook their correct internal row-7 source references are K7/C7/E7/I7/G7 respectively. No VBA changes are required for this HTML/post update.
Comments