# Filtered Historical Simulation (FHS) GARCH VaR & ES Guide

> Build a Filtered Historical Simulation (FHS) VaR and Expected Shortfall engine by bootstrapping GARCH-standardized residuals across horizons.

Canonical: https://quantcalctools.com/guides/filtered-historical-simulation-garch-standardized-residual-var
Last updated: 2026-11-14T19:29:39.969Z

**Publication:** QuantCalc Tools
**Category:** Parametric Var Expected Shortfall
**Published:** 2026-11-14T13:20:25.915Z
**Author:** Elena Rostova, FRM, CQF (FRM, Certificate in Quantitative Finance (CQF))
**Reviewer:** Not recorded

## Answer-first summary

Barone-Adesi Filtered Historical Simulation (FHS) marries GARCH conditional volatility responsiveness with non-parametric empirical fat tails and cross-asset rank dependence.

## Key takeaways

- Filtered Historical Simulation divides historical returns by prevailing historical GARCH volatility (z_t = r_t / sigma_t) and multiplies those i.i.d. innovations by tomorrow’s forecast volatility.
- De-volatilizing uncovers severe structural shocks hidden inside low-volatility eras (such as a -2.10% move on a 0.55% vol day = -3.82 sigma) that unfiltered HS underweights.
- Drawing synchronous cross-sectional date rows Z_{d*, 1..N} during multi-day path bootstrapping preserves empirical crash co-dependence without estimating huge correlation matrices.
- Always re-center standardized residuals to zero mean and normalize variance to exactly 1.00 before bootstrapping so finite-sample GARCH estimation drift does not bias 10-day paths.

## Related financial entities

- **Basel III FRTB 97.5% Expected Shortfall (ES) Tail Metric** (FRTB-ES-975) — license: Basel III FRTB 97.5% Expected Shortfall (ES) Tail Metric (FRTB-ES-975) refers to an institutional financial benchmark and regulatory standard monitored by BIS Basel Committee MAR33 (INTL), currently calibrated at 97.5% (Confidence Interval) as of October 2026.
- **Cholesky Multi-Asset Correlated Monte Carlo Path Simulator — 2026 Standard** (MONTE-CARLO-CHOLESKY-STD) — product: MONTE-CARLO-CHOLESKY 2026 Institutional Benchmark Framework is defined as the standardized quantitative and compliance methodology published by OCC / CME Risk Architecture governing Cholesky Multi-Asset Correlated Monte Carlo Path Simulator across US capital markets (10,000 Simulated Paths, October 2026).
- **CME SPAN 2 Historical & Stress Value-at-Risk (VaR) Model** (CME-SPAN2-VAR) — product: CME SPAN 2 Historical & Stress Value-at-Risk (VaR) Model (CME-SPAN2-VAR) refers to an institutional financial benchmark and regulatory standard monitored by CME Clearing (US), currently calibrated at 6.26% (Initial Margin / Notional) as of October 2026.

## Analysis

Unfiltered Historical Simulation (HS) preserves empirical fat tails and non-linear asset co-movements without imposing Gaussian formulas, yet fails [Christoffersen Markov independence backtests](/guides/christoffersen-markov-independence-conditional-coverage-test) whenever market volatility shifts regimes: feeding a calm 2025 return window into an autumn volatility spike understates 1-day VaR by half, whereas feeding crisis returns into a quiet market overcharges capital for twelve months. Pioneered by Barone-Adesi, Giannopoulos, and Vosper (1999), **Filtered Historical Simulation (FHS)** eliminates that trade-off—and today powers clearinghouse margin engines including LCH ForexClear and components of [CME SPAN 2 Historical & Stress VaR](/entities/cme-span2-var-cme-span-2-historical-stress-value-at-risk-var-model). Benchmark parametric normal outputs in our [Value at Risk (VaR) calculator](/calculators/value-at-risk-calculator) and review our [Parametric VaR & Expected Shortfall calibration guide](/guides/parametric-var-expected-shortfall-fat-tail-calibration-guide).

## De-Volatilization and Re-Volatilization: Exact FHS Equations

Suppose we observe $T = 500$ historical daily returns $r_1, \dots, r_T$ and fit an asymmetric GJR-GARCH(1,1) conditional mean and variance path $(\hat{\mu}_t, \hat{\sigma}_t)$ for each day $t = 1, \dots, T$, plus tomorrow's conditional forecast $(\hat{\mu}_{T+1}, \hat{\sigma}_{T+1})$:

1. **Extract Standardized Empirical Innovations ($\hat{z}_t$)**:
 $$\hat{z}_t = \frac{r_t - \hat{\mu}_t}{\hat{\sigma}_t}, \qquad t = 1, \dots, T$$
 Dividing each historical return by the conditional volatility prevailing *on that historical date* strips out volatility clustering, transforming raw returns into approximately i.i.d. draws that retain true empirical skewness, kurtosis, and cross-asset crash co-dependence!
2. **Re-Scale to Tomorrow's Volatility State ($\tilde{r}_{T+1}^{(t)}$)**:
 $$\tilde{r}_{T+1}^{(t)} = \hat{\mu}_{T+1} + \hat{\sigma}_{T+1} \cdot \hat{z}_t, \qquad t = 1, \dots, T$$

## Worked Barone-Adesi GARCH De-Volatilization & Re-Scaling Audit

```formula
Worked 1-Day Filtered vs Unfiltered Historical Simulation ($40,000,000 GBP/USD & Gilts Book)
Historical Window: T = 500 Trading Days; Target Metric: 97.5% Expected Shortfall (Worst 2.5% = Top 12.5 Losses)
Current Market State: Post-BoE Surprise Volatility Spike -> Tomorrow's GARCH Forecast sigma_{T+1} = 1.65%
Historical Worst Tail Days (Raw Loss vs Prevailing Historical GARCH Volatility sigma_t):
 Obs #114 (Quiet Summer Flash Dip): r_114 = -2.10%, Prevailing sigma_114 = 0.55% => z_114 = -3.8182!
 Obs #302 (Prior Crisis Selloff): r_302 = -3.60%, Prevailing sigma_302 = 1.80% => z_302 = -2.0000!

1) Unfiltered Historical Simulation Distortion:
 Ranks Obs #302 (-3.60% = -$1,440,000) as far worse than Obs #114 (-2.10% = -$840,000),
 ignoring that tomorrow's vol (1.65%) is 3x higher than Day 114's vol!
 Unfiltered 97.5% ES across raw 500-day average vol (0.88%) = $896,000.

2) Filtered Historical Simulation Re-Scaling at sigma_{T+1} = 1.65%:
 Re-scaled Obs #114 = 1.65% * (-3.8182) = -6.3000% (-$2,520,000!)
 Re-scaled Obs #302 = 1.65% * (-2.0000) = -3.3000% (-$1,320,000 — drops out of worst 2.5% tail!)
 Average of Worst 12.5 Standardized Residuals E[-z | z <= q_0.025] = 2.7420
 FHS 1-Day 97.5% ES = $40,000,000 * 0.0165 * 2.7420 = $1,809,720 (+101.98% timely risk capture!)
```

Examine that reversal between Observation `#114` and Observation `#302`! A raw $-2.10\%$ drop occurring during a sleepy $0.55\%$-volatility summer was actually a **$-3.82\sigma$ structural shock**; scaling that innovation to today's $1.65\%$ conditional volatility reveals a latent $-6.30\%$ crash scenario that raw Historical Simulation completely buries!

![Multi-day Filtered Historical Simulation path bootstrap with synchronous cross-asset residual vectors](/api/figure?type=flow&title=Multi-Horizon%20Synchronous%20Vector%20Bootstrap%20FHS%20Architecture&subtitle=Preserving%20empirical%20cross-asset%20tail%20copulas%20while%20recursively%20updating%20GJR-GARCH%20variance%20paths.&steps=1.%20Univariate%20GJR%20Fit%3A%3ADe-volatilize%20each%20asset%20i%20into%20residual%20matrix%20Z_%7Bt%2Ci%7D%7C2.%20Draw%20Random%20Day%20d*%3A%3ASample%20full%20row%20Z_%7Bd*%2C%201..N%7D%20to%20keep%20rank%20correlation%7C3.%20Update%20GARCH%20Path%3A%3AStep%20sigma_%7BT%2Bk%2C%20i%7D%5E2%20forward%20using%20simulated%20shock%7C4.%20Revalue%2010d%20Horizon%3A%3ACompound%2010%20steps%20across%2010%2C000%20paths%20for%2097.5%25%20ES&metrics=RESIDUAL%20MEAN%20%2F%20VAR%3A%3A0.0%20%2F%201.00%3A%3ARe-center%20%26%20variance-normalize%20Z_t%7CCOPULA%20PRESERVATION%3A%3AExact%20Row%3A%3ASimultaneous%20date%20draw%20keeps%20tail%20jumps%7C10-DAY%20PATHS%3A%3A10%2C000%20Runs%3A%3AReplaces%20invalid%20sqrt%2810%29%20scaling%7CSAMPLE%20WINDOW%3A%3A500-1000d%3A%3ABalances%20tail%20richness%20%26%20stationarity)

## Multi-Day Recursive Path Bootstrapping & Synchronous Cross-Asset Vectors

How do you extend FHS from a 1-day risk horizon to a 10-day [Basel III FRTB 97.5% Expected Shortfall](/entities/frtb-es-975-basel-iii-frtb-97-5-expected-shortfall-es-tail-metric) horizon without invoking $\sqrt{10}$? You **cannot** just sort 1-day residuals! Instead, run $B = 10,000$ recursive path simulations of length $H = 10$ days:
- On step $k = 1, \dots, 10$ of path $b$, draw a single random historical date index $d^* \in \{1, \dots, T\}$ uniformly with replacement.
- Apply **the entire cross-sectional row vector** $(\hat{z}_{d^*, 1}, \hat{z}_{d^*, 2}, \dots, \hat{z}_{d^*, N})$ simultaneously across all $N$ assets! Drawing the same historical date $d^*$ across every asset automatically preserves non-Gaussian empirical tail dependence (copula crash synchronization) without inverting a $N \times N$ correlation matrix!
- Compute each asset's Day $T+k$ simulated return $r_{T+k, i}^{(b)} = \hat{\mu}_{T+k, i} + \hat{\sigma}_{T+k, i}^{(b)}\hat{z}_{d^*, i}$, feed $(r_{T+k, i}^{(b)})^2$ directly into asset $i$'s GJR-GARCH equation to obtain $\hat{\sigma}_{T+k+1, i}^{(b)}$ for the next step, and compound the 10-day path!

## Unfiltered Historical Simulation vs Parametric GARCH vs FHS Comparison

| Simulation / VaR Paradigm | Volatility Responsiveness | Tail Distribution Shape | Multi-Asset Crash Copula | Procyclicality Floor Needed? |
| :--- | :--- | :--- | :--- | :--- |
| Unfiltered Historical Simulation | Very Slow (Ghost effects) | Empirical (Capped at max loss) | Exact historical dates | Low (Sticky crisis memory) |
| Gaussian Parametric / EWMA | Fast ([EWMA vs GARCH](/guides/ewma-lambda-094-vs-garch-11-conditional-volatility-var-decay)) | Thin mesokurtic ($K_e=0$) | Linear Pearson $\Sigma$ only | High (Needs vol floor) |
| Filtered Historical Simulation (FHS) | **Fast (GJR-GARCH scaled)** | **Empirical standardized $z_t$** | **Exact synchronous row copula** | Moderate (EMIR 25% floor) |
| FHS + EVT GPD Hybrid (McNeil-Frey) | **Fast (GJR-GARCH scaled)** | **Smooth Pareto extrapolation** | Empirical + GPD marginals | Best-in-class extreme tail |
| Cholesky Normal Monte Carlo | Model-dependent | Parametric only | Elliptical [MONTE-CARLO-CHOLESKY](/entities/monte-carlo-cholesky-2026-benchmark-standard) | Model-dependent |

![250-day 99% VaR backtest exception comparison across Unfiltered HS, Gaussian EWMA, Standard FHS, and McNeil-Frey FHS-EVT](/api/figure?type=bars&title=Empirical%20250-Day%2099%25%20VaR%20Backtest%20Performance%20Across%20Regimes&subtitle=Comparing%20breach%20counts%20and%20Christoffersen%20LR_cc%20conditional%20coverage%20p-values%20during%20volatile%20macro%20years.&steps=Unfiltered%20250d%20HS%3A%3A7%20clustered%20breaches%20fail%20Christoffersen%20LR_cc%20%3D%2014.8%7CGaussian%20EWMA%200.94%3A%3A6%20fat-tail%20breaches%20despite%20fast%20vol%20decay%20%28LR_cc%20%3D%204.9%29%7CStandard%20GJR-FHS%3A%3A3%20isolated%20breaches%20pass%20Kupiec%20%26%20Markov%20%28LR_cc%20%3D%200.42%29%7CMcNeil-Frey%20FHS-EVT%3A%3A2%20isolated%20breaches%3B%20smooths%20sparse%2099.9%25%20quantiles&metrics=UNFILTERED%20HS%20HITS%3A%3A7%20%28Clustered%29%3A%3AFails%20independence%20p%20%3D%200.0006%7CGAUSSIAN%20EWMA%20HITS%3A%3A6%20%28Amber%29%3A%3AMisses%20non-normal%20jump%20kurtosis%7CSTANDARD%20FHS%20HITS%3A%3A3%20%28Green%29%3A%3AClean%20LR_cc%20%3D%200.42%20pass%7CFHS%20%2B%20EVT%2099.9%25%20TAIL%3A%3AExtrapolated%3A%3ASolves%20max%28z_t%29%20sample%20ceiling)

## Eliminating Sample Ceiling Bias via McNeil-Frey FHS-EVT

Standard FHS has one remaining limitation at ultra-high confidence levels (such as 99.9% CCP default-fund stress): a 500-day residual pool contains **zero** observations beyond $\min_t(\hat{z}_t)$. McNeil & Frey (2000) solved this permanently by fitting a Generalized Pareto Distribution (GPD) to standardized GARCH residuals exceeding the 90th or 95th loss percentile—see our companion derivation in [Extreme Value Theory Peaks-Over-Threshold GPD tail ES](/guides/extreme-value-theory-peaks-over-threshold-gpd-tail-es-model). Audit broker margin methodologies across venues like [BP Prime](/broker-lookup/bp-prime) and explore all 20 guides on the [Parametric VaR & Expected Shortfall Topic Hub](/topics/parametric-var-expected-shortfall).

## Frequently asked questions

### Why must standardized GARCH residuals be re-centered and variance-normalized before running FHS?

In finite samples of 500 days, quasi-maximum-likelihood GARCH residuals often exhibit a sample mean slightly off zero and sample standard deviation around 0.98 or 1.02; enforcing exact zero mean and unit variance prevents compounding bias across 10-day bootstrap horizons.

### How does Filtered Historical Simulation handle correlations across 200 portfolio assets?

Instead of estimating 19,900 pairwise correlations, FHS fits 200 univariate GJR-GARCH filters and bootstraps historical calendar dates synchronously across all 200 residual columns, preserving exact historical rank correlation and joint tail jumps.

### What happens if a newly listed stock or bond has fewer than 500 historical days in an FHS engine?

Missing pre-IPO history breaks synchronous date-row sampling unless backfilled via factor-model regression against sector/curve benchmarks plus idiosyncratic residual draws, as permitted under Basel MAR33 proxy-data rules.

### Why do central counterparties add a 25% anti-procyclicality floor to FHS initial margin?

During prolonged tranquil markets, GARCH volatility drops toward historic lows, shrinking FHS margin and encouraging excess leverage. EMIR RTS rules require CCPs to floor FHS volatility or blend a 25% buffer calibrated to stressed periods.


## Source URLs

- https://www.jstor.org/stable/2676226
- https://www.sciencedirect.com/science/article/abs/pii/S0927539800000128
- https://www.esma.europa.eu/sites/default/files/library/esma70-151-1154_final_report_emir_rts_apc_margin_measures_for_ccps.pdf

## Citation

- APA: Elena Rostova, FRM, CQF. (2026). Filtered Historical Simulation (FHS): GARCH Bootstrap VaR & ES Math Guide. QuantCalc Tools. https://quantcalctools.com/guides/filtered-historical-simulation-garch-standardized-residual-var
- Canonical article: [Filtered Historical Simulation (FHS): GARCH Bootstrap VaR & ES Math Guide](https://quantcalctools.com/guides/filtered-historical-simulation-garch-standardized-residual-var)
- Site research index: [QuantCalc Tools](https://quantcalctools.com/llms-full.txt)
