Methodology · Tool · Last updated 2026-05-08
How Drawdown-Recovery Markov Simulator works
Monte Carlo with Cornish-Fisher quantiles to estimate time-to-recover from a drawdown — the engine behind the Drawdown-Recovery Markov simulator.
Cornish-Fisher quantile expansion
For a target distribution with skew γ₃ and excess kurtosis (γ₄ − 3), the moment-matched quantile from a standard normal z is:
q ≈ z + (z² − 1)·γ₃/6 + (z³ − 3z)·(γ₄ − 3)/24 − (2z³ − 5z)·γ₃²/36 This is the standard fourth-order expansion (Cornish-Fisher 1937). It is exact at z = 0 and accurate for moderate skew (|γ₃| ≲ 1) and excess kurtosis (γ₄ − 3 ≲ 6). Beyond those bounds the approximation distorts.
Simulation procedure
μ_m = monthly_sharpe × 0.04 ← assume σ_m = 4%
For each path p of P:
value = 1 − drawdown_threshold
For t = 1..max_months:
z = standard_normal_sample()
q = cornish_fisher(z, γ₃, γ₄ − 3)
r = μ_m + 0.04 · q
value *= 1 + r
If value ≥ 1:
record recovery_time = t; break
If never recovers, record never += 1 Outputs
- 25th / 50th / 75th / 95th percentile of recovery times.
- Probability of recovering within 12 months.
- Probability of never recovering within the simulation cap (240 months).
Why σ_m is fixed
Time-to-recover for a given drawdown threshold depends on μ_m / σ_m (the monthly Sharpe), not σ_m alone — fixing σ_m = 4% is convenient and does not affect the result. Doubling both μ and σ yields the same recovery times.
References
- Cornish, E. A., Fisher, R. A. (1937). "Moments and cumulants in the specification of distributions." Revue de l'Institut International de Statistique 5(4): 307–320. DOI: 10.2307/1400905.
- Maillard, D. (2012). "A user's guide to the Cornish-Fisher expansion." SSRN 1997178.
- Magdon-Ismail, M., Atiya, A. F. (2004). "Maximum drawdown." Risk Magazine 17(10): 99–102.
- Burghardt, G., Duncan, R., Liu, L. (2003). "Deciphering Drawdown." Risk Magazine, September.
Limitations
- Returns are i.i.d. by assumption — no regime switching, no autocorrelation.
- Cornish-Fisher accuracy degrades for extreme skew/kurtosis. For very fat tails consider a t-distribution sampler.
- Simulation cap of 240 months means very-low-Sharpe inputs will report ∞-by-cap.