What Regime-Aware Monte Carlo Does to a Retirement Plan: Sequence Risk, Quantified
In the companion post we fitted a two-state regime model to 155 years of monthly real S&P total returns: a calm state (9.8% annualized volatility, 83% of months) and a stressed state (25.0% volatility, negative drift, episodes that persist for months). This post answers the practical question: if you make a Monte Carlo retirement simulation draw returns from that structure — without changing the portfolio's expected return — what happens to the plan?
Short answer: for a mid-career saver, probability of success falls from 81% to 62%. For a near-retiree, the age by which the worst tenth of outcomes runs out of money moves six years earlier. Same expected return. Same volatility in four months out of five. The difference is entirely when the bad months arrive.
The setup: three return engines, one expected return
Run your own numbers — FREE
10,000 Monte Carlo simulations. Forward-looking forecasts from BlackRock, JPMorgan, Vanguard, GMO, Schwab, Invesco. No account needed.
Try QuantCalc Free →We ran the same plans through QuantCalc's simulation engine under three configurations, 10,000 paths each, identical random seed:
- Single-regime baseline. Standard lognormal monthly draws from forward-looking capital market expectations. No regime structure.
- Post-war stress overlay. The engine's long-standing hand-calibrated regime defaults: 1.8× volatility in stressed months, ~6% long-run stressed share, no drift shift.
- Fitted (1871–2026) preset. The parameters from the historical fit: 2.55× stressed-state volatility, monthly transition probabilities 0.031 (calm→stressed) and 0.151 (stressed→calm), implying a 17% long-run stressed share, plus the fitted bull–bear drift spread of −2.78%/month (log), re-centered under the chain's stationary distribution.
Two design choices keep the comparison clean — both documented in the methodology:
- The drift spread is re-centered, not bolted on. Forward-looking expected returns are unconditional — historical stressed episodes are already in the average. Applying the fitted stressed-state drag on top would double-count them and silently lower expected return by almost 6%/yr. Instead, every month receives a small offset so the calm state sits slightly above the unconditional mean, the stressed state carries the full spread, and the probability-weighted average matches your capital market inputs.
- The volatility multiplier is variance-compensated. Under lognormal growth, naively multiplying shocks by 2.55× in stressed months would mechanically raise arithmetic expected returns through convexity (by roughly 1.2%/yr at these parameters). The engine cancels that term exactly, so regime switching widens the distribution without quietly improving its mean. We verified both properties numerically: on a contribution-only test plan, mean final wealth under the fitted preset matches the single-regime baseline within 0.5%.
In other words: all three engines agree on the expected return. They disagree about clustering.
Plan A: mid-career accumulator
Age 40, retiring at 65, planning to 95. $250,000 saved, $2,000/month contributions, $5,000/month retirement spending, $2,200/month Social Security from 67, 2.5% inflation, 60/10/25/3/2 stock-heavy allocation.
| Return engine | Success rate | Median at 95 | 90th percentile |
|---|---|---|---|
| Single-regime baseline | 81.0% | $1.83M | $8.6M |
| Post-war stress overlay | 77.4% | $1.58M | $9.0M |
| Fitted (1871–2026) | 62.5% | $0.95M | $15.5M |
The fitted preset removes 18.5 percentage points of success probability relative to the baseline — and simultaneously raises the 90th percentile by 80%. That pattern is the signature of regime structure: both tails fatten. Paths that traverse retirement without a long stressed episode compound at the calm state's higher drift and finish far richer. Paths that hit a 2001-style or 2008-style episode early in the withdrawal phase — months of −20%-annualized drift at 25% volatility, exactly when withdrawals are forced — fail at rates the single-regime model cannot produce.
A single-regime model with the same mean and even the same unconditional variance spreads its bad months evenly across 55 years, where contributions and time diversify them away. The regime model concentrates them. Concentration is what sequence risk actually is.
Plan B: five years from retirement
Age 60, retiring at 65, planning to 95. $900,000 saved, $1,500/month contributions until retirement, $5,500/month spending, $2,400/month Social Security from 67, 55/10/30/3/2 allocation. This plan is deliberately tight — baseline success is roughly a coin flip, which is where modeling assumptions matter most.
| Return engine | Success rate | Age by which 10% of paths failed | Age by which 25% failed |
|---|---|---|---|
| Single-regime baseline | 50.6% | 82 | 87 |
| Post-war stress overlay | 49.1% | 81 | 85 |
| Fitted (1871–2026) | 46.6% | 76 | 81 |
The headline success rate barely moves — four points. The failure timing moves dramatically: under the fitted preset, the worst tenth of outcomes is broke by 76 instead of 82. For a 60-year-old, that is the difference between a problem that surfaces at an age where spending can still adjust and one that arrives after the adjustment window has closed. Plans that look equivalent on probability-of-success can carry very different early-failure profiles — one reason we think success probability alone undersells what Monte Carlo output contains.
Which engine should you believe?
None of these is the truth; they are three different priors about clustering, and the historical record sits closest to the third. What we would actually defend:
- The baseline is the right default for ranking decisions — allocation A vs allocation B, claim Social Security at 67 vs 70. Clustering affects both sides of those comparisons similarly.
- The fitted preset is the right stress view for withdrawal-phase questions — spending levels, cash buffers, when failure would surface. Those answers genuinely depend on drawdowns arriving in clusters, because that is how they arrived for 155 years.
- The gap between the two is information. A plan that holds 80%+ success under both engines is robust to the clustering assumption. A plan that drops 18 points is exposed to it, and the exposure is concentrated in the first decade of withdrawals.
The fitted preset ships in QuantCalc's PRO regime panel ("Fitted to 1871–2026 monthly data"); the underlying regime probabilities update monthly on the Market Regime Monitor. Run your own plan under both engines in the calculator — the comparison takes two clicks, and the delta between them is the most useful number this post can't compute for you.
All figures from QuantCalc engine runs: 10,000 Monte Carlo paths per configuration, identical random seed, forward-looking capital market expectations held identical across configurations. The reference-plan definitions and full percentile tables are published in the research repository alongside the regime fit.