Why backtest at all when you already have Monte Carlo
A Monte Carlo engine generates markets that never existed. That is its strength: it explores the space of plausible futures rather than the single path history happened to take. It is also its weakness, because the generated markets inherit whatever assumptions the model was given.
Historical backtesting has the opposite trade-off. There is exactly one sequence of real returns, so the sample is small and the future need not resemble it. But nothing is assumed: the correlations, the inflation spikes, the fifteen-year flat stretches and the crashes that arrived in the wrong order are all real.
Running both is how you find out whether a plan is robust or merely lucky in one model. If a plan survives 1966 and 1929 starts and also clears a high probability of success in simulation, the two methods are agreeing for different reasons.
How the backtest works
Killion takes your current allocation, your contribution or withdrawal schedule and your horizon, then replays them starting in each year for which a full horizon of data exists. A 30-year horizon starting in 1928 ends in 1958; starting in 1929 it ends in 1959, and so on to the most recent complete window.
Each of those runs is a "world". A world either survives the horizon or runs out of money, and the share that survive is the survival rate. Because the same start years and the same withdrawals are applied to every allocation being compared, differences in the result come from the allocation rather than from luck in the sample.
- Every start year, not a sample. The full set of complete windows is tested, so no year is quietly excluded.
- Identical treatment across strategies. Your mix, 100% stocks, 60/40 and 30/70 face the same years and the same spending.
- Open any world. Drill into a single start year to see the path, the worst drawdown and each annual return.
- Log scale by default. Decades of compounding read as a line rather than flattening into the axis.
Reading the result without fooling yourself
A survival rate is not a probability of your future. It is the share of historical windows in which a plan shaped like yours made it. Those windows overlap heavily (1929 and 1930 share 29 of 30 years), so they are far from independent samples, and the effective sample size is much smaller than the count suggests.
The number is still useful, especially at the extremes. A plan that fails in a third of historical windows is fragile in a way no amount of modelling assumption will fix. A plan that survives all of them may still fail in the future, but it has at least cleared everything that has actually happened.
The research behind it
- Is the 4% Rule Still Safe? We Ran 5,000 Simulated Retirements
- Sequence of Returns Risk: What If the Market Crashes the Year You Retire?
Related features
- Withdrawal strategies: Pick a spending rule (the 4% rule, guardrails, a fixed percentage) and watch the projection follow it.
- What-if scenarios: Retire two years earlier, switch the spending rule, add a market shock, then compare up to four plans side by side.