📊 statistics

An outlier-robust dynamic semiparametric model for jointly forecasting value-at-risk and expected shortfall

This paper proposes an outlier-robust joint dynamic semiparametric model that integrates bounded innovation propagation into the CAViaR framework to improve the accuracy and reliability of Value-at-Risk and Expected Shortfall forecasts, particularly in the presence of extreme observations and heavy-tailed distributions.

Original authors: Jie Wang, Buyun Xu

Published 2026-08-14

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Jie Wang, Buyun Xu

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a weather forecaster trying to predict the worst possible storm of the year. You don't just want to know when the rain starts; you need to know how deep the floodwaters will get if the levees break. In the world of finance, this is exactly what risk managers do. They use two special tools to measure the danger of a financial market crashing. The first tool, called Value-at-Risk (VaR), is like a flood gauge that tells you the water level you are 99% sure won't be exceeded. It's a popular "line in the sand." The second tool, Expected Shortfall (ES), is even more critical: it asks, "Okay, we know the water might go over the line, but if it does, how deep will the water actually get?" This is the average depth of the worst-case scenarios.

For decades, mathematicians have built complex models to predict these levels. However, these models have a secret weakness: they panic when they see a "weird" data point. In finance, a "weird" point is an outlier—a sudden, massive price jump or crash that doesn't fit the usual pattern, like a meteor hitting a calm lake. Standard models treat these meteors as if they were just slightly bigger raindrops, letting them distort the entire forecast for the future. This paper, written by Jie Wang and Buyun Xu, introduces a new way to build these models that acts like a shock absorber. Instead of letting a single meteor crash the whole system, their new model learns to ignore the extreme spikes, ensuring the flood gauge stays accurate even when the weather gets chaotic.

The Problem: When One Bad Apple Spoils the Forecast

Financial markets are rarely calm. They are full of surprises. Sometimes, a stock price drops 10% in a day because of a sudden scandal, or jumps 15% because of a surprise earnings report. In the language of statistics, these are outliers.

The standard models used to predict VaR and ES are like a very sensitive, high-tech thermostat. If you throw a bucket of ice water on it, the thermostat doesn't just adjust the temperature; it freaks out, thinks the whole house is freezing, and turns the heat off completely. In the real world, this means a single, massive market shock can cause the model to overreact, predicting a disaster that might never happen, or worse, failing to see the next real danger because it's too busy reacting to the last one.

The authors of this paper noticed that while these models are great at handling normal market "weather," they break down when the "weather" gets extreme. They wanted to build a model that could say, "Okay, that was a huge spike, but let's not let it rewrite the entire history of how this market behaves."

The Solution: The "Bounded Innovation Propagation" Shock Absorber

To fix this, the researchers introduced a concept called Bounded Innovation Propagation (BIP). Think of this as a shock absorber for the model's brain.

Imagine you are driving a car down a bumpy road. The standard model is like a car with no suspension; every time you hit a pothole (an outlier), the whole car shakes violently, and the driver loses control. The new BIP model is like a car with heavy-duty suspension. When you hit a massive pothole, the suspension compresses to absorb the impact. The car still feels the bump, but it doesn't flip over.

In technical terms, the model uses a "bounded function." This is a mathematical rule that says, "No matter how big the shock is, its effect on our future predictions will never exceed a certain limit." If a stock price drops by 50% (a massive outlier), the model treats it as if it only dropped by, say, 10% for the purpose of updating its future predictions. It caps the damage.

This new model is called BIP-CAViaR-FZ. It combines three things:

CAViaR: A flexible way to track how risk changes over time.
FZ Loss: A special scoring system that checks if both the VaR (the line) and the ES (the depth) are correct at the same time.
BIP: The shock absorber that stops outliers from ruining the score.

What the Experiments Showed

The authors didn't just guess this would work; they put it through rigorous testing using two methods: computer simulations and real-world data.

1. The Simulation Lab
First, they created thousands of fake financial markets in a computer. They programmed these markets to have different types of "weather": some were calm (normal distribution), some had heavy tails (frequent big shocks), and some were skewed (more big drops than big jumps). They even injected specific "meteors" (outliers) into the data.

The results were clear:

In calm weather: The new model performed just as well as the old models. It didn't break anything.
In stormy weather: When the data was full of outliers, the old models got confused and made big mistakes. The new BIP model, however, kept its cool. It predicted the risk levels much more accurately.
The "Heavy Tail" Effect: The advantage was most obvious when the market had "heavy tails" (meaning extreme events happen more often than a normal bell curve would suggest) and when looking at very rare events (like a 1% chance of a crash). In these tough scenarios, the new model was significantly more accurate.

2. The Real-World Test
Next, they tested the model on real stock data from five major U.S. technology giants: Apple, Amazon, Alphabet (Google), Intel, and Microsoft. They looked at daily returns from 2006 to 2025.

They found that these stock prices are full of outliers. For example, Amazon's stock history shows several massive spikes and drops that look like meteors. When they compared the new BIP model against the standard models:

Smoother Forecasts: The standard models reacted violently to these spikes, causing their risk predictions to jump up and down wildly. The BIP model produced much smoother, more stable predictions.
Better Accuracy: When they checked how well the models predicted the future, the BIP models made fewer mistakes. They were better at catching the true risk levels without being thrown off by the noise.
The Quadratic Models: The improvement was especially huge for the "quadratic" versions of the models (which are more complex). These complex models were the most sensitive to outliers, and the BIP shock absorber saved them from falling apart.

The Verdict

The paper suggests that by adding a "shock absorber" to the way we calculate financial risk, we can make our predictions much more reliable during times of market chaos. The authors found that the new BIP-CAViaR-FZ model doesn't just survive outliers; it thrives in their presence, offering a more robust way to forecast both the "line in the sand" (VaR) and the "depth of the flood" (ES).

While the standard models are still useful for calm days, the authors argue that in the real world—where financial markets are messy and full of surprises—this new approach is a smarter, safer bet. It doesn't ignore the bad days; it just makes sure one bad day doesn't ruin the whole forecast.

Technical Summary: An Outlier-Robust Dynamic Semiparametric Model for Jointly Forecasting Value-at-Risk and Expected Shortfall

Problem Statement
Financial time series frequently contain outliers—large, isolated shocks that deviate significantly from prevailing patterns. While the joint forecasting of Value-at-Risk (VaR) and Expected Shortfall (ES) has advanced through the Conditional Autoregressive Value-at-Risk (CAViaR) framework combined with Fissler–Ziegel (FZ) loss minimization, existing models typically lack robustness against these extreme observations. Since VaR and ES forecasts are updated recursively, even a small number of extreme observations can propagate through the updating equations, distorting parameter estimation and adversely affecting subsequent risk forecasts. Outliers may inflate dispersion, exaggerate skewness or tail heaviness, and lead to larger in-sample estimation errors and poorer out-of-sample forecast performance. Despite the prevalence of outliers in financial data, the CAViaR-FZ literature has paid limited attention to this issue.

Methodology
To address the lack of robustness, the authors propose an outlier-robust extension of the CAViaR-FZ framework, termed the BIP-CAViaR-FZ model. The core innovation is the incorporation of Bounded Innovation Propagation (BIP) into the dynamic updating equations for VaR.

Framework: The model retains the semiparametric nature of the original CAViaR-FZ framework, where VaR dynamics are specified parametrically, the conditional distribution of returns is left unspecific, and ES is modeled as a scalar multiple of VaR ( $e_t = (1+\gamma)v_t$ ). Parameter estimation is conducted by minimizing the Fissler–Ziegel (FZ0) loss function, which is strictly consistent for the joint estimation of VaR and ES.
BIP Mechanism: The authors modify standard CAViaR specifications (Symmetric Absolute Value, Asymmetric Slope, Indirect GARCH, and Indirect GJR-GARCH) by introducing a bounded function, $\eta(x; c) = \min(x, c)$ , into the innovation term. For example, in the BIP-SAV specification, the updating equation becomes:
$|v_t| = \beta_0 + \beta_1|v_{t-1}| + \beta_2 \min(|y_{t-1}|, \beta_3|v_{t-1}|)$
Here, $\beta_3$ acts as a threshold parameter. When the absolute return $|y_{t-1}|$ exceeds the threshold $\beta_3|v_{t-1}|$ , its contribution to the updating equation is capped. This limits the influence of extreme innovations on future conditional risk measures.
Estimation: The model is estimated via a four-stage procedure involving global search and local optimization to minimize the average FZ0 loss. The gradient of the objective function is derived to facilitate optimization. The constraint $\beta_3 \geq 1$ is imposed to ensure that ordinary shocks retain their natural influence while unusually large shocks remain bounded.

Key Contributions

Robust Extension: The paper introduces the first explicit incorporation of BIP into joint dynamic semiparametric models for VaR and ES.
Nested Structure: The standard non-robust CAViaR-FZ model is nested as a special case (as $\beta_3 \to \infty$ ), facilitating direct comparison and implementation.
Semiparametric Flexibility: The proposed model maintains the flexibility of the original framework by avoiding direct estimation of the conditional density, thereby reducing the risk of distributional misspecification while adding robustness.

Results
The paper evaluates the model through extensive simulation studies and an empirical analysis of five major U.S. technology stocks (AAPL, AMZN, GOOGL, INTC, MSFT).

Simulation Findings:
- Outlier Environments: In the presence of outliers, particularly under skewed heavy-tailed distributions (e.g., Hansen's skew-t) and at lower probability levels (1%), the BIP-CAViaR-FZ model significantly outperforms the standard CAViaR-FZ model.
- Accuracy: The BIP specifications yield lower Root Mean Square Errors (RMSE) for both parameter estimation and risk forecasting when the data-generating process includes outliers. The advantage is most pronounced for quadratic specifications (IG and IGJR), which are more sensitive to extreme observations.
- Efficiency: When data are generated without outliers (e.g., from a normal distribution), the BIP specifications show only marginal efficiency losses compared to the standard models, indicating that the robustness mechanism does not severely penalize performance in clean data environments.
Empirical Findings:
- Data Characteristics: The selected stock return series exhibit significant excess kurtosis and skewness, with numerous potential outliers identified via boxplot rules.
- Performance: In full-sample risk estimation (measured by FZ0 loss) and out-of-sample forecasting (measured by traditional backtests like UC, CC, DQ, and comparative loss rankings), the BIP-CAViaR-FZ models consistently outperform or match their non-robust counterparts.
- Visual Evidence: The BIP models produce smoother VaR and ES paths that are less reactive to isolated spikes in returns compared to the sharp, abrupt movements observed in standard models.

Significance
The paper claims that incorporating bounded innovation propagation into joint VaR–ES modeling provides a "useful and practically relevant way to enhance robustness in the presence of outliers." The significance lies in the model's ability to mitigate the distortion of risk forecasts caused by extreme observations without sacrificing the flexibility of the semiparametric approach. The results suggest that the proposed specification is particularly effective in challenging environments characterized by heavy tails, asymmetry, and small sample sizes, offering a more reliable tool for financial risk management and regulatory compliance (e.g., Basel III requirements for ES). The authors note that future research could extend this framework to high-frequency and mixed-frequency settings.

The Problem: When One Bad Apple Spoils the Forecast

The Solution: The "Bounded Innovation Propagation" Shock Absorber

What the Experiments Showed

The Verdict

More like this