Limitations of Historical VaR for Trending and Tail-Risk Scenarios
Summary
The document raises concerns about using unfiltered historical Value at Risk as a desk-level measure for comparing strategies and allocating risk capital. It contrasts this with a preferred use of a common holding period and confidence level, robust backtesting, and a parametric form that is less sensitive to historical idiosyncrasies. The described firm instead uses configurable historical windows and confidence levels, including very high confidence thresholds and multi-day horizons.
The author argues that sparse tail observations, short samples, trends, and large shocks can make these estimates noisy or misleading. For linear positions, historical VaR can assign sharply different risk to long and short exposures, even where exchange margin requirements are symmetric; instrument-specific definitions may then weaken comparability. The document asks whether historical VaR can be adjusted to reflect martingale or risk-neutral forward dynamics, but provides no proposed correction or empirical evidence. It frames the issue rather than resolving it.
Key ideas
- Historical VaR estimates depend strongly on the chosen sample, confidence level, and holding period.
- Sparse tail data, trends, and shock events can make historical estimates noisy.
- Long and short linear positions can receive different historical VaR estimates even when exchange margin treatment is symmetric.
- Instrument-specific VaR definitions can reduce comparability across positions.
- The document asks about martingale-based corrections but does not identify or test a method.
Tags
Full text
# Martingale corrections to historical Value at Risk? # Martingale corrections to historical Value at Risk? I am looking for a bit of advice. I have recently used to a new firm, which uses Value at Risk in a manner that is unfamiliar from previous places I have worked that I find less than ideal. Previous, I have seen it used as a simple desk level tool as a basic measue of risk to translate across strategies/assets/instruments, as often as not used as the denominator in a reward/risk ratio for trade sizing or two allocate scarce risk capital. For this purpose it needs to (a) have a universal definition (holding period,confidence) across assets and (b) to have confidence limits high enough to be simply and robusts back tested (c) be of parametric form to cut out any historical idiosyncracies. Regulatory calcs were often done by a more refined method. At the new place, they use purely historcal calculations, exactly the same methodology as used for regulatory submissions andares completely unfiltered with a database that is configurable in terms of confidence limits, historical periods etc. There is a tendency to use really tailish confidence limits (e.g. 99.5%) and long holding periods (10 day etc.). As you can imagine this leads some noisy, less meaningful calculations when used to time series which are short, or particularly trending or have huge shock events. For instance, for linear positions, long and short positions can have drastically different VaR consumptions, at odds to how much margin the exchanges would demand for the two positions. To address some of the postions some risk managers use idiosyncratic defintions by asset/instrument etc. This to me undermines the usefulness of the numbers severely. So my question is, is there a frequently referenced method to convert historical methods to treat forward price dynamics as risk neutral/martingale (e.g. to moderate the differences between long and short linear positions). There is such a deep literature on VaR that I would have thought a quick google would find that quickly. I may be missing something obvious but couldn't see this.
Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.