There are basically two ways to answer the question "how bad can this get," and I have watched people pick one, defend it like a religion, and get burned by the half they ignored. You can measure risk from the actual sample you have, which is the historical or empirical route. Or you can assume the returns follow some distribution, fit a couple of parameters, and read the risk off the formula, which is the parametric or probabilistic route. Most of the time these two agree closely enough that nobody notices the difference. The trouble is that they disagree exactly in the region you care about, which is the tail.
So it is worth being precise about what each one is actually doing, and where each one quietly lies to you.
What the two methods really measure
The historical approach is almost embarrassingly simple. You take your realized returns, sort them, and read the number off. Historical value at risk at the 95 percent level is just the 5th-percentile loss in your sample. Maximum drawdown is the deepest peak-to-trough you actually lived through. There is no model, no assumption about shape. You are reporting what happened.
The parametric approach fits a distribution instead. The usual default is the normal, where you estimate the mean and the standard deviation and then risk statistics fall out of the shape. Tail loss at a given confidence level becomes the standard deviation times a fixed multiplier. Drawdown gets estimated from volatility and horizon rather than read off a chart. It is clean, it is smooth, and it gives you an answer even for events your sample never contained.
That last part is the whole tension. The historical estimate can only tell you about losses at least as large as the worst thing in your window. The parametric estimate will happily hand you a number for a 1-in-500 day even if your data is 200 days long. One of them is silent where you need it, and the other is confident where it has no right to be.
Where each one lies
The historical method lies by omission. If your sample did not contain a crash, your empirical tail loss reflects a world without crashes. A strategy that ran through a calm couple of years will show a tame historical drawdown right up until the first real shock, at which point the "worst case" you were quoting gets doubled in an afternoon. The empirical record is only as honest as the window is representative, and short windows are almost never representative of the tail. This is also why maximum drawdown is such a treacherous single number. It is one draw from a distribution, the deepest hole you happened to fall into, and it has enormous sampling error. Two identical strategies can show very different max drawdowns purely by luck of timing.
The parametric method lies by assumption. Financial returns are not normal. They have fat tails, meaning extreme moves happen far more often than a bell curve predicts, and they cluster, meaning volatile days come in runs rather than sprinkled evenly. If you plug a normal distribution into a series that has fat tails, your tail-loss estimate will look reassuringly small and will be wrong in the dangerous direction. The formula does not know it is lying. It gives you a crisp number with a confidence level attached, which is exactly what makes it seductive and exactly what makes it dangerous.
A quick way to see which failure you are exposed to: compare the two estimates on the same data.
- If historical tail loss is much larger than the normal parametric estimate, your returns have fat tails and the clean formula is understating risk. Trust the empirical number more, and be suspicious of anything built on a normal assumption.
- If historical tail loss is much smaller than the parametric one, your sample probably just did not contain a bad enough event yet. The calm is doing the talking. Do not congratulate yourself.
- If they roughly agree, either your data is genuinely well behaved or your sample is too short to have revealed the disagreement. Check the length before you relax.
Why blending beats either alone
The practical fix is not to crown a winner. It is to use each one as a check on the other's blind spot. The historical estimate anchors you to reality, so you never talk yourself into a tail loss smaller than something that literally happened. The parametric estimate extrapolates past the edge of your sample, so you get a number for the event you have not seen yet instead of a shrug.
A few things I actually do rather than just admire the theory:
- Fatten the tails in the model. If you are going parametric, do not use the normal. A Student t distribution with low degrees of freedom, or a plain volatility multiplier bump, pushes the estimate toward the fat-tailed reality historical data keeps showing. This closes most of the gap in the honest direction.
- Take the worse of the two for sizing. When historical and parametric tail loss disagree, size the position off the larger number. You give up a little efficiency in calm periods and you stop getting cut in half in bad ones. That trade is almost always worth it.
- Bootstrap the drawdown instead of trusting the single realized max. Resample your returns many times and look at the distribution of drawdowns you could have had. Your realized max drawdown is one point in that cloud, and usually not the scary end of it. Plan around a high percentile of the resampled distribution, not the one path you happened to walk.
- Watch the volatility regime. Both methods assume the near future looks statistically like the window you fit. Volatility clusters, so if you are estimating in a quiet stretch, both estimates are borrowing calm they will have to pay back. Scale your risk budget down when realized volatility is unusually low, not up.
The failure mode I see most often is someone reporting a single risk number with a decimal point on it and treating the decimal as precision. It is not precision. It is one estimate from one method on one window, and there is a second method that would give a different answer for a reason worth understanding. When the two agree, fine, you can be a bit more relaxed. When they split, the size of the split is the actual information. It is telling you how much of your comfort is coming from the shape of the world versus the shortness of your memory.
Neither the empirical record nor the clean formula deserves your full trust. The record is honest about the past and silent about the future. The formula is fluent about the future and often wrong about it. Keeping both on the screen, and paying attention to when they part ways, is roughly as close to honest risk measurement as I have found.