---
title: Quantile Estimation via Residual Simulation (QRS)
url: https://www.emergentmind.com/topics/quantile-estimation-through-residual-simulation-qrs
type: topic
---

# Quantile Estimation via Residual Simulation (QRS)

Quantile Estimation through Residual Simulation (QRS) is a probabilistic forecasting method for converting one-step-ahead point forecasts into conditional quantile forecasts without building a full parametric distribution of the target variable. In "Probabilistic Forecasting Cryptocurrencies Volatility: From Point to Quantile Forecasts" [2508.15922], QRS is formulated for a univariate volatility series, specifically daily realized variance \(y_t\), or its log-transformation \(z_t=\ln y_t\). The method approximates the predictive distribution at forecast origin \(\tau\) by combining the current point forecast with an empirical distribution of in-sample residuals, thereby producing a full set of quantile forecasts from a wide range of base models, including statistical and machine learning algorithms [2508.15922].

## 1. Definition and theoretical basis

QRS addresses the quantile estimation problem for a future time \(\tau\): estimate the conditional \(q\)-quantile
\[
Q_{q,\tau}=F^{-1}_{Y_\tau\mid\mathcal I_{\tau-1}}(q)
\]
without specifying a full parametric law for \(Y_\tau\) [2508.15922]. The method starts from a point-forecasting model that produces one-step-ahead forecasts \(\hat y_t\) or, on the log scale, \(\hat z_t\).

Its residual-simulation framework assumes
\[
Y_t = \hat Y_t + \varepsilon_t,
\]
with in-sample residuals
\[
\varepsilon_t = Y_t - \hat Y_t,\quad t=1,\dots,\tau-1.
\]
These residuals are treated as an empirical approximation to the future forecast-error distribution. Defining the empirical residual set
\[
\mathcal E = \{\varepsilon_1,\ldots,\varepsilon_{\tau-1}\},
\]
QRS approximates the predictive distribution at time \(\tau\) as
\[
Y_\tau \approx \hat Y_\tau + \varepsilon,\quad \varepsilon\sim\mathcal E.
\]

The method is therefore nonparametric at the error-distribution stage and model-agnostic at the point-forecast stage. A plausible implication is that its performance depends directly on the quality of the underlying point forecast and on how informative the historical residuals are about future forecast errors.

## 2. Algorithmic construction

The procedure described in [2508.15922] converts a point forecast into a set of simulated future values and then extracts empirical quantiles from those simulations.

First, an optional preprocessing step applies the log-transformation \(z_t=\ln y_t\). This is used to stabilize variance and enforce positivity after back-transformation. Next, a base model is fitted on historical data \(\{y_t\}_{t=1}^{\tau-1}\) or \(\{z_t\}\) to generate in-sample one-step-ahead forecasts \(\hat y_t\) for \(t=1,\dots,\tau-1\) and the next-day forecast \(\hat y_\tau\).

The in-sample residuals are then computed as
\[
e_t = y_t - \hat y_t,\quad t=1,\dots,\tau-1.
\]
From these residuals, the empirical residual set \(\mathcal E=\{e_1,\dots,e_{\tau-1}\}\) is formed. The paper describes three variants for generating forecast-error draws: directly reusing each \(e_t\) as a simulated error, sampling with replacement \(M\) times from \(\mathcal E\), or fitting a kernel-density estimator \(\hat f_e\) on \(\mathcal E\) using a Gaussian kernel with cross-validated bandwidth and drawing \(M\) samples from \(\hat f_e\) [2508.15922].

Each simulated residual \(e^{(i)}\) is added to the current point forecast:
\[
\tilde y_\tau^{(i)} = \hat y_\tau + e^{(i)},\quad i=1,\ldots,M.
\]
If the model is estimated on the log scale, the simulated values are generated as
\[
\tilde z_\tau^{(i)}=\hat z_\tau+e^{(i)},
\]
followed by back-transformation
\[
\tilde y_\tau^{(i)}=\exp(\tilde z_\tau^{(i)}).
\]

The recommended number of simulations depends on the residual-generation strategy. If in-sample residuals are simply recycled, then \(M\) equals the training-window length \((\tau-1)\). If residuals are sampled with replacement or drawn from a fitted kernel, the recommended range is \(M\ge1{,}000\)–\(5{,}000\) in order to ensure stable quantile estimates [2508.15922].

## 3. Quantile extraction and predictive distribution

Once the simulated ensemble \(\{\tilde y_\tau^{(i)}\}_{i=1}^M\) is available, QRS estimates the empirical \(q\)-quantile as
\[
\hat Q_{q,\tau}
=
\inf\Big\{\,y\;\big|\;\frac{1}{M}\sum_{i=1}^M\mathbf1\{\tilde y_\tau^{(i)}\le y\}\ge q\Big\}
\approx
\tilde y_\tau^{(\lceil Mq\rceil)}
\quad\text{(sorted).}
\]
Operationally, one sorts the simulated values in ascending order and selects the \(\lceil Mq\rceil\)-th order statistic. If a continuous CDF \(\hat F\) has been fitted, the quantile may also be obtained by numerical inversion:
\[
\hat Q_{q,\tau} = \hat F^{-1}(q).
\]

This extraction mechanism makes QRS a direct simulation-based route from a point forecast to a predictive quantile function. The method does not estimate quantiles through separate quantile-specific models at this stage; instead, all forecast quantiles are induced from a single predictive simulation ensemble. This suggests a structural reason why QRS avoids quantile-crossing when the simulated sample is sorted into a valid empirical distribution.

## 4. Base models, ensembles, and stacking

QRS is explicitly described as model-agnostic. In the reported application it can be applied to residuals from statistical models such as HAR, HAR-R, ARFIMA, and GARCH, and from machine learning models such as LASSO, RR, SVR-G, SVR-L, MLP, FNM, RF, and LSTM [2508.15922]. The same mechanism can be used on a single base model or on an ensemble point forecast, such as the cross-model mean
\[
\hat y_\tau^{\rm EnsMean}=\tfrac1n\sum_{i=1}^n\hat y_{i,\tau},
\]
or a cross-model median.

The paper contrasts QRS with a separate framework termed probabilistic stacking. In probabilistic stacking, the vector of base point forecasts
\[
\hat{\mathbf y}_\tau=[\hat y_{1,\tau},\dots,\hat y_{n,\tau}]
\]
is optionally augmented with additional predictors, including lagged daily, weekly, and monthly realized variance, and then passed to a quantile-regression meta-model \(f_q(\,\cdot\,;\theta_q)\) that directly outputs
\[
\hat Q_{q,\tau}=f_q(\hat{\mathbf y}_\tau,\mathbf x_\tau;\theta_q).
\]

Two specific meta-models are highlighted. Quantile Linear Regression (QLR) is defined as
\[
\hat Q_{q,\tau}=\sum_{i=1}^n a_{i,q}\,\hat y_{i,\tau}+a_{0,q},
\]
with estimation by minimizing the pinball loss \(\sum_jL_q(y_j,\hat Q_{q,j})\). Quantile Regression Forest (QRF) is described as an extension of Random Forest that retains the full empirical distribution of \(y\) in each leaf and estimates conditional quantiles as weighted sums of leaf-node responses [2508.15922].

Within this comparison, QRS functions as a residual-based post-processing method, whereas probabilistic stacking is a quantile-level meta-modeling approach. The empirical discussion in [2508.15922] reports that QRS on single strong base models, especially linear models on \(\ln\)-RV, often outperforms or matches these more sophisticated stacking solutions in overall CRPS, calibration, and interval-score metrics.

## 5. Empirical evaluation in cryptocurrency volatility forecasting

The reported application concerns Bitcoin realized-variance forecasting. The study states that it is, to the best of the authors’ knowledge, the first in the literature to propose and systematically evaluate probabilistic forecasts of variance in cryptocurrency markets based on predictions derived from multiple base models [2508.15922].

The evaluation uses several probabilistic and point-forecast proxy metrics. Continuous Ranked Probability Score (CRPS) is computed as a sum of pinball losses over a dense grid of quantiles:
\[
\mathrm{CRPS}(F,y)\approx\sum_{q\in\Pi}L_q\bigl(y,\hat Q_{q}\bigr).
\]
Calibration is assessed through relative frequency (coverage),
\[
\mathrm{ReFr}(q)=\tfrac1N\sum_{i=1}^N\mathbf1\{y_i\le\hat Q_{q,i}\},
\]
with nominal target \(\mathrm{ReFr}(q)=q\), and through Mean Absolute ReFr Error (MARFE),
\[
\frac1{|\Pi|}\sum_{q\in\Pi}|\mathrm{ReFr}(q)-q|.
\]
The evaluation also includes the Winkler Score (WS) for a \((1-\alpha)\times100\%\) interval \([\hat Q_{q_l},\hat Q_{q_u}]\), and the point-forecast proxies \(\mathrm{MAE\!-\!Q}\) and \(\mathrm{MSE\!-\!Q}\), both based on \(\hat Q_{0.5}\) [2508.15922].

The principal empirical findings reported for Bitcoin are summarized below.

| Finding | Reported result |
|---|---|
| QRS on \(\ln\)-RV linear models | Lowest mean CRPS \(\approx9.45\times10^{-4}\); MARFE \(0.026\text{–}0.049\) |
| Raw-RV QRS | Negative lower quantiles sometimes occurred; wider CRPS dispersion |
| Probabilistic stacking | QLR-l and QRF-l mean CRPS \(9.57\text{–}9.78\times10^{-4}\) |
| Quantile-crossing | QLR in \(\approx12\%\) of days; QRF and QRS never |
| 90% interval score | Best values around \(1.11\text{–}1.12\times10^{-2}\) |

More specifically, QRS applied to linear base models on \(\ln\)-RV, including HAR-l, HAR-R-l, ARFIMA-l, RR-l, and SVR-L-l, achieved the lowest mean CRPS and competitive calibration. Probabilistic stacking via QLR-l and QRF-l produced mean CRPS in the same general range and similar interval scores, but did not substantially outperform the simplest QRS-l specifications. The best 90% interval scores were around \(1.11\text{–}1.12\times10^{-2}\), reached by QRF-l, QLR-l, and some QRS-l instantiations [2508.15922].

These results support a narrow but important conclusion: in the reported Bitcoin setting, residual simulation around strong point forecasts—especially linear forecasts on logarithmized volatility—was sufficient to match or exceed more elaborate quantile-level meta-models.

## 6. Practical considerations, limitations, and interpretation

Several implementation considerations are explicit in the source description. Base-model choice matters: QRS is recommended for models with low point-forecast error and reasonably stable, near-IID residuals. In the reported study, the best QRS partners were simple linear autoregressions on \(\ln\)-RV [2508.15922]. Data transformation also matters. Applying \(\ln\) to realized variance avoids negative simulated draws and mitigates outliers, whereas on raw RV the method may generate inadmissible negative lower quantiles.

The method is computationally light. The reported runtime is approximately \(0.015\,\mathrm{s}\) per daily date for 99 quantiles, because QRS avoids retraining. By comparison, QLR takes approximately \(0.8\,\mathrm{s}\), described as one LP per quantile, and QRF takes approximately \(0.4\,\mathrm{s}\), described as a single forest [2508.15922]. This computational profile is consistent with the architecture of QRS: once a point forecast and residual history are available, quantiles are produced by simulation and sorting rather than repeated optimization.

The paper also identifies several pitfalls. QRS inherits any systematic bias in the base model and assumes stationarity of the residual distribution. Kernel-density fitting may fail to converge at extreme \(q\), in which case interpolation may be required. The method does not adapt to time-varying heteroscedasticity in forecast errors [2508.15922]. These caveats limit the circumstances under which historical residuals can reliably approximate future forecast uncertainty.

A common misconception would be to treat QRS as a full generative model for volatility. The formulation in [2508.15922] does not present it that way. It is a residual-based approximation scheme for conditional quantiles built on top of an existing point forecast. Another possible misconception is that more sophisticated quantile meta-models necessarily dominate residual simulation. The reported Bitcoin results do not support that generalization: QRS on well-specified linear models operating on \(\ln\)-transformed realized volatility often matched or exceeded QLR and QRF under CRPS, calibration, and interval-score criteria [2508.15922].

## 7. Position within probabilistic volatility forecasting

Within the framework of cryptocurrency volatility forecasting, QRS occupies the transition from deterministic to probabilistic prediction. The paper introduces probabilistic forecasting methods that leverage point forecasts from a wide range of base models to estimate conditional quantiles of cryptocurrency realized variance, and QRS is one of the central mechanisms for doing so [2508.15922]. In this setting, the method supplies a distributional layer over existing forecasters rather than replacing them.

The broader significance claimed in the source is domain-specific: traditional deterministic forecasting methods are described as inadequate for capturing the full spectrum of potential volatility outcomes in cryptocurrency markets, while probabilistic methods provide more comprehensive insight into uncertainty and risk. The empirical outcome that QRS performed particularly well when paired with linear base models on log-transformed realized volatility suggests that, in at least this application, the fidelity of residual behavior and the effect of scale transformation may be more consequential than increasing model complexity.

In that sense, QRS can be understood as a parsimonious probabilistic wrapper around point forecasts. In the reported evidence, its simplicity, model-agnosticity, and low computational cost coincide with strong empirical performance in Bitcoin realized-variance forecasting, especially relative to more elaborate stacking-based alternatives [2508.15922].

Source: https://www.emergentmind.com/topics/quantile-estimation-through-residual-simulation-qrs