Neural Score-Driven Volatility and Risk Forecasting

summary

The gist

Neural score-driven models are developed to represent innovation density using a neural network within a Generalized Autoregressive Score (GAS) framework…

In short

The study developed GAS-NN, a neural score-driven volatility model using a compact neural network to represent innovation density scores within a Generalized Autoregressive Score framework. This method allows for flexible modeling of financial returns and volatility dynamics, successfully outperforming traditional parametric benchmarks during the COVID-19 period.

Key concepts

Generalized Autoregressive Score (GAS) Framework
This is a statistical model used to describe how the log variance of financial returns evolves over time. It dictates that the change in log variance depends on a constant term, a score contribution from past observations, and the current log variance itself. This framework provides a structured way to model volatility dynamics.
Neural Score-Driven Model (GAS-NN)
This model replaces the traditional score component in the GAS framework with a neural network. The network learns to represent the negative density score of innovation density. By using this neural representation, the model gains flexibility to capture complex features like asymmetry in financial returns, which standard models might miss.
Innovation Density Score ($\phi(z)$)
This score is defined as the negative derivative of the log probability density function with respect to its input. It explicitly shows how a new observation affects the volatility response. The model learns this score using a neural network, allowing it to adapt the volatility prediction based on how 'innovative' or extreme a new return observation is.
Score Matching
This estimation procedure is used to train the neural network by comparing its output scores against target scores derived from standardized returns. It helps learn the shape of the marginal standardized-return distribution, which is crucial for accurately modeling how volatility responds to new observations.

Terminology used across episodes

This episode discusses

The paper

Neural Score-Driven Volatility and Risk Forecasting · Read on arXiv

Qianli Zhao, Chao Wang, Minh-Ngoc Tran, Richard Gerlach

Transcript

Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.

Ruth: We're looking at the paper: "Neural Score-Driven Volatility and Risk Forecasting".

Warren: Here's the gist of the paper. Neural score-driven models are developed to represent innovation density using a neural network within a Generalized Autoregressive Score (GAS) framework, allowing for flexible modeling of financial return distributions and volatility dynamics This paper develops GAS-NN, a neural score-driven volatility model in which a compact feedforward network represents a raw negative density score The proposed method relies on a lightweight neural network to represent the negative score of the innovation density By integrating and normalizing the learned score and imposing location-scale identification, we recover a flexible innovation density that can accommodate asymmetry, including negative skewness commonly observed in financial returns Two estimation procedures are considered.

Ruth: A two-step procedure estimates the neural score by score matching and subsequently fits the volatility recursion with the identified density held fixed A one-step procedure jointly refines the neural and GAS parameters under the conditional likelihood The results show that the neural score-driven models are competitive with Gaussian and Student-t GAS benchmarks that require a prespecified parametric innovation distribution, with particularly strong performance during the COVID-19 period Neural Score-Driven Volatility and Risk Forecasting> Model Framework The underlying observation model is defined as rt = exp(ft/2)zt, where ft = log σ2t and zt i.i.d.

Warren: ∼ D(0, 1).

Ruth: The generalized autoregressive score (GAS) framework dictates that the log variance evolves according to ft+1 = ω + αst + βft, where st is the score contribution to ft+1 The resulting models retain a tractable observation-driven recursion and can be estimated by maximum likelihood The negative innovation-density score is defined as ϕ(z) = − ∂ log p(z) / ∂z This representation makes explicit how the innovation-density score determines the volatility response to a new observation The GAS-NN model replaces ϕ in Equation (8) with its neural representation NN and using unit scaling, St = 1, gives the proposed GAS-NN model The neural network selected for the empirical analysis has two hidden layers with eight neurons in each layer, one scalar input, and one scalar output, giving a 1–8–8–1 architecture The identified neural score NN(z) is obtained by relocating and rescaling the raw score g, given by NN(z) = a g(m + az), (16)> Estimation Procedures Two estimation procedures are considered for fitting the model parameters First, the two-step GAS-NN procedure estimates the raw neural score by regularized score matching on unconditionally standardized returns After density reconstruction and location–scale identification, the neural parameters are held fixed while the GAS parameters are estimated by conditional maximum likelihood Second, the one-step GAS-NN procedure jointly estimates the neural and GAS parameters using a regularized conditional likelihood, initialized from the corresponding two-step solution Joint refinement allows the innovation density and the standardized residuals implied by the volatility filter to adjust together The objective combines the mean negative conditional log likelihood in Equation (28) with the boundary and neural parameter penalties described in Section 4.4 Optimization is initialized at the two-step estimates from Section 4.5 The GAS parameters obey the restrictions in Section 2.1 Retained neural estimates must satisfy the strict endpoint sign conditions before they define a valid reconstructed density for likelihood evaluation Density Reconstruction and Identification The innovation density must satisfy Z R p(z) dz = 1, E(zt) = 0, Var(zt) = 1 The raw neural score g approximates this target The raw neural score g approximates this target The reconstructed raw density is given by H(u) = Z u ∫ g(v) dv, q(u) = exp(-H(u)) C, C = Z R exp(-H(v)) dv The mean m and variance a2 of the reconstructed raw density q are given by m = Z R u q(u) du, a2 = Z R (u − m)2 q(u) du Using the standardized coordinate z = (u − m)/a, we relocate and rescale the raw density q and its negative density score g to obtain the identified innovation density p with mean zero and unit variance and its corresponding negative density score NN: p(z) = aq(m + az), NN(z) = ϕ(z) = ag(m + az), (20)> Training Signal Construction The first-stage training inputs are constructed as ut = rt / (√T in T in PT in j=1 r2 j, (24)>The former use a single sample-wide scale, while the latter use the time-varying conditional standard deviation Score matching on ut learns the shape of the marginal standardized-return distribution, which generally differs from the conditional innovation distribution The population score-matching criterion is Jpop(θNN) = 1/2 Z R (g(u) − g∗(u))2 pU (u) du, (26)> Performance Evaluation We evaluate forecasting performance using predictive negative log-likelihood (PNLL), the quantile score (QS), and the asymmetric-Laplace joint loss (AL) PNLL:= 1/T out T in X + T out t=T in+1 bft2 − log p(zt) /, (33)>The AL joint loss is defined in equation (35) All three criteria are averaged over the corresponding evaluation period, and lower values indicate better forecast performance During the COVID period, GAS-NN-1 (30-30) attains the lowest predictive negative log-likelihood in all eight markets, ranging from 1.9381 for HSI to 2.2183 for DJI The consistent numerical ranking of GAS-NN-1 (30-30) therefore indicates strong crisis-period density performance, while the MCS results leave several neural alternatives statistically competitive The findings indicate that the amount of first-stage training and the subsequent joint refinement have different implications for density and tail-risk forecasts Likelihood refinement adjusts the complete predictive distribution, but it does not necessarily minimize realized losses focused on the lower tail The two-step specification remains useful when its fixed working density produces favourable risk forecasts, even when another model provides a better average density fit Several extensions follow from the present design.

Paper summary: Ruth: So we're talking about this paper, "Neural Score-Driven Volatility and Risk Forecasting," which basically develops a way to model volatility without needing to pick a specific distribution for the innovation density.

Warren: Right, the core idea is using a neural network to represent the negative score of that innovation density, and then fitting it into a Generalized Autoregressive Score framework.

Iris: It claims this lets you recover a flexible innovation density that can handle asymmetry, like negative skewness which we see in financial returns.

Ruth: That's the main appeal, right? They replace the traditional score function with this lightweight neural network to capture how volatility responds to new observations.

Warren: And they test it using eight international equity indices and also look at a high-volatility period during COVID-nineteen.

Iris: They compare their neural score-driven models against benchmarks like Gaussian and Student t Generalized Autoregressive Score models that require you to prespecify the distribution, showing that the neural score driven models are competitive with those.

Ruth: That's a big deal because it means we don't have to assume a certain shape for the innovation density when modeling financial returns.

Warren: The paper describes two estimation procedures: a two-step procedure where you estimate the neural score first and then fit the volatility recursion with that fixed density, and a one-step procedure that jointly refines both parameters under a conditional likelihood.

Iris: That joint refinement procedure seems interesting because it allows the innovation density and the standardized residuals implied by the volatility filter to adjust together.

Ruth: Exactly, so instead of picking one path or the other, you let them adapt simultaneously using that regularized conditional likelihood approach.

Warren: The paper also shows how they reconstruct a raw density from that neural score by integrating it and then normalizing it to get a valid probability distribution.

Iris: And they use this reconstructed density to define the mean and variance of the raw density, which is then used to define the identified innovation density with mean zero and unit variance.

Ruth: They construct two different training signals for their neural network: one based on a single sample-wide scale, and another that uses time-varying conditional standard deviation.

Warren: And they use score matching on that second signal to learn the shape of the marginal standardized return distribution, which is often different from what's driving the innovation distribution.

Iris: The population score matching criterion they use looks at how close the neural network's raw score is to a target raw score based on a reconstructed density.

Ruth: It seems like this whole setup is designed to capture the dynamics of volatility while letting the data determine the underlying structure of what drives that volatility.

Warren: The performance evaluation uses predictive negative log likelihood, quantile score, and an asymmetric Laplace joint loss to see how well they forecast risk.

Iris: And during that COVID period, the GAS-NN-one model achieved the lowest predictive negative log likelihood across all eight markets, ranging from one point nine three eight one for HSI to two point two one eight three for DJI.

Ruth: That consistent ranking suggests strong performance in crisis periods, which is exactly what we want to see when modeling extreme events.

Warren: But the authors also point out some limitations, specifically that the amount of first-stage training and the subsequent joint refinement have different implications for density and tail-risk forecasts.

Iris: They mention that likelihood refinement adjusts the complete predictive distribution but it doesn't necessarily minimize realized losses focused on the lower tail.

Ruth: It seems like they suggest that even if one procedure gives a better average density fit, you might still want to stick with the two-step specification if it produces favorable risk forecasts.

Warren: They also suggest extensions, like using preliminary filtered residuals for score matching inputs or exploring manifold learning for multivariate innovations.

Iris: And they note that further theoretical work is needed to establish stability and estimation properties for this neural representation of the innovation density within the GAS framework.

Ruth: So, looking at the whole thing, it’s about using a neural network to learn an innovation score within a structured volatility recursion, and whether that learning process helps us forecast risk better than fixed parametric models.

Warren: The paper provides empirical support for learning these innovation-density scores in this structured way, with gains depending on the specific market, evaluation period, and the target forecast.

Conclusion: Ruth: So, we're wrapping up this look at "Neural Score-Driven Volatility and Risk Forecasting." The main thing is they built a model that uses a neural network to figure out the shape of innovation density without having to guess it beforehand.

Warren: Right, so instead of forcing the data into a fixed distribution like Gaussian or Student t, they let the network learn how much the volatility spikes based on new information.

Iris: What this means is we get a more flexible way to model things like negative skewness in financial returns because the density shape can change with the data.

Ruth: Exactly, it moves away from those older models where you had to pick a distribution and then see if it fit. It lets the data dictate that underlying structure.

Warren: But they did this by fitting a compact feedforward network to represent the negative score of that innovation density, which is pretty neat for keeping things computationally light.

Iris: The paper shows two ways to estimate this—a two-step process where you find the neural score first and then fit the volatility recursion with it fixed, and a one-step process that lets both parameters adjust together.

Ruth: That joint refinement procedure is interesting because it suggests that the density shape and how volatility reacts can be learned at the same time under a specific likelihood function.

Warren: And when they tested this across eight different markets, especially during the COVID period, it actually showed a very strong performance in predicting risk compared to those traditional benchmarks.

Iris: That COVID result is notable because it shows that this neural approach can handle periods of extreme market stress quite well when compared to models that assume a fixed distribution.

Ruth: It really suggests that learning the innovation density score within the volatility recursion is a solid way to get better forecasts without relying on those older, more rigid assumptions.

Warren: So what does this mean for us out there? It means we might be able to build risk forecasts that adapt better to real-world market surprises instead of being stuck with a single mathematical assumption.

Iris: It opens up the door for modeling more complex dependencies in financial time series where the underlying distribution isn't static.

Ruth: And while it's competitive, they still point out that the way you train the network and whether you do that joint refinement matters a lot for what kind of forecast you actually get.

Warren: Exactly, so it’s not just one magic setup; there are trade-offs in how much training data you use and how tightly coupled those two estimation procedures are.

Iris: We still need more theoretical work to fully nail down the stability and estimation properties of this neural representation itself, though the empirical results are encouraging.

More episodes

← Home