Quant GANs: Financial Time Series Generation

• Quant GANs are deep generative models that combine WaveNet-style TCNs with risk-neutral constraints to simulate arbitrage-free financial time series.
• They employ dilated-causal convolutions to capture long-range dependencies and heavy-tailed behaviors, effectively mimicking volatility clustering and leverage effects.
• They outperform traditional models like GARCH by accurately reproducing both marginal distributions and dependency structures in asset returns.

Quant GANs are a class of generative adversarial networks specifically tailored for the deep generation of financial time series, integrating temporal convolutional networks (TCNs) into the adversarial learning paradigm. This approach directly addresses the limitations of classical stochastic-process models in capturing the empirical complexities of financial data—such as volatility clustering, leverage effects, serial autocorrelations, and non-Gaussian heavy tails—by employing a data-driven neural generative mechanism that nonetheless embeds risk-neutral martingale constraints required for quantitative finance applications (Wiese et al., 2019).

1. Core Architecture and Model Structure

Quant GANs depart from standard feed-forward or recurrent architectures by leveraging WaveNet-style dilated-causal convolutional neural networks, both in the generator and discriminator. The generator is a functional mapping

$$g_\theta : \mathbb{R}^{N_Z \times rfs^{(g)}} \to \mathbb{R}^{N_X \times 1}$$

where $N_Z$ is the latent noise dimension (typically $3$), $rfs^{(g)}$ is the generator's receptive field (standardly $127$ time steps), and $N_X$ is the dimensionality of the output (often $1$, e.g., for univariate asset returns). The generator produces synthetic log-return sequences via

$$\tilde R_{t,\theta} = g_\theta(Z_{t-(rfs^{(g)}-1):t}),$$

where $\{Z_t\}$ are i.i.d. standard normal noise variables.

Quant GANs admit a constrained form called the stochastic-volatility neural network (SVNN), where the generator outputs at each step both a volatility $\sigma_t$ and drift $\mu_t$; synthetic return samples are then given by

$R_{t,\theta} = \sigma_t \odot \epsilon_t + \mu_t,$

where $\epsilon_t = g^{\epsilon}_\beta(Z_t)$ is an independent noise process.

The discriminator is likewise a TCN, mapping real or fake log-return blocks

$$d_\eta : \mathbb{R}^{N_X \times rfs^{(d)}} \to [0,1],$$

with $rfs^{(d)}$ typically set equal to the generator's receptive field. Both networks are implemented as sequences of temporal blocks with dilated-causal convolutions, PReLU activations, and skip connections, following the architectures given in Def. 2.9 and Table A.1 of the source (Wiese et al., 2019).

2. Risk-Neutral Generative Modeling

A critical innovation of Quant GANs is the explicit risk-neutral calibration of the generated stochastic process. For financial applications, synthetic asset-price paths must evolve as martingales under the risk-neutral measure (to avoid arbitrage in option pricing and hedging). The model imposes this constraint using an analytic correction to each generated return: $$S_t = S_{t-1} \cdot \exp(R_{t,\theta}), \quad \tilde S_t = e^{-rt} S_t$$ requiring

$$\mathbb{E}[\exp(R_{t,\theta} - r) \mid \mathcal{F}_{t-1}] = 1$$

or, for SVNNs,

$$R^M_{t,\theta} = R_{t,\theta} - \log h(\sigma_{t,\theta}, \mu_{t,\theta}) + r$$

where $h(\sigma,\mu) = \mathbb{E}_x[\exp(\sigma x + \mu)]$ and $x$ is distributed as the generator's internal noise. For Gaussian $\epsilon_t$, the correction $h(\sigma,\mu) = \exp(\mu + \sigma^2/2)$ gives

$$R^M_{t,\theta} = \sigma_t \epsilon_t - \frac{1}{2} \sigma_t^2 + r,$$

which enforces the risk-neutral martingale property in each simulated path.

3. Adversarial Objective and Training Regime

Quant GANs adopt the classical Goodfellow-style minimax adversarial loss: $$\mathcal{L}(\theta, \eta) = \mathbb{E}_X[\log d_\eta(X_{1:rfs^{(d)}})] + \mathbb{E}_Z[\log(1-d_\eta(\tilde X_{1:rfs^{(d)},\theta}))]$$ where $X_{1:rfs^{(d)}}$ are real blocks and $\tilde X_{1:rfs^{(d)},\theta}$ are fake blocks generated from noise. The generator is trained to minimize, and the discriminator to maximize, this objective. In practice, Quant GANs apply adversarial regularization heuristics for training stability (e.g., modifications from Mescheder et al. 2018), though no explicit Wasserstein critic or gradient penalty is imposed in the baseline setup.

4. Data Pipeline, Parameterization, and Implementation

The time-series data undergoes stringent preprocessing: calculation of log returns, standardization, application of the inverse Lambert W transform to induce marginal Gaussianity (for tractable learning of heavy tails), re-standardization, and windowing into overlapping rolling blocks. Training employs the Adam optimizer with typical learning rates $\alpha_g=10^{-4}$ for the generator and $\alpha_d=5\times 10^{-4}$ for the discriminator, batch size $M=64$, and a generator-to-discriminator update ratio $k=1:5$.

The convolutional structures use 7 dilated blocks with kernel sizes beginning at 1 (for block 1) and 2 for subsequent blocks; dilation doubles in each successive block (i.e., $1,2,4,8,16,32$). Hidden layers are of width $N_H=80$ (Wiese et al., 2019).

5. Empirical Performance: Statistical and Dependency Metrics

Quant GANs are benchmarked on S&P 500 log-returns (2,413 daily samples) and are evaluated using:

• Distributional metrics:
  • Earth-Mover Distance (EMD/Wasserstein-1) for lags $t = 1,5,20,100$.
  • DY-distance: $$DY(t)=\sum_{x} |\log P^h_t(A_{t,x})-\log P^g_t(A_{t,x})|$$ (between empirical and generated histograms).
• Dependence scores:
  • L2-norm between real and mean-generated autocorrelation functions for raw, absolute, and squared returns up to lag $S=250$.
  • Leverage score: L2-norm between real and generated lagged cross-correlations $Corr(r_t, r_{t+\tau}^2)$, for $\tau=1$ to $250$.

Quant GANs yield EMD and DY distances an order of magnitude lower than GARCH(1,1) baselines and nearly perfectly reproduce empirical autocorrelation and leverage structures: | Model | EMD(1) | EMD(100) | DY(1) | ACF²-score | Leverage | |-----------|--------|----------|--------|------------|----------| | TCN | 0.0039 | 0.0154 | 19.12 | 0.0214 | 0.3291 | | C-SVNN | 0.0040 | 0.0464 | 19.85 | 0.0245 | 0.3351 | | GARCH(1,1)| 0.0199 | 0.0935 | 32.71 | 0.0253 | 0.4636 |

These results indicate accurate reproduction of heavy tails (via preprocessing, not learned end-to-end), volatility clusters, and leverage effects (Wiese et al., 2019).

6. Significance, Limitations, and Outlook

Quant GANs' principal strengths are the capacity to model high-lag dependencies, articulate explicit risk-neutral corrections for arbitrage-free simulation, and empirically outperform GARCH in both marginal and dependence statistics. The convolutional design avoids both the memory limitations and the instability of RNN-based models for long memory.

Limitations include:

• Stability: GAN training remains sensitive to hyperparameters with no convergence guarantee.
• Tail behavior: The heavy-tailed character is "reintroduced" by preprocessing rather than being learned organically; end-to-end tail learning is unresolved.
• Receptive field: Maximum horizon limited to the TCN's receptive field (e.g., 127 days); longer-term dependencies would require architectural modifications (e.g., attention, residual blocks).
• Calibration: Current training is on historical data; direct calibration to option-implied (risk-neutral) distributions or other observables is not implemented.

Future work includes extending the generation horizon, improving tail learning, and integrating risk-neutral path constraints into end-to-end adversarial training (Wiese et al., 2019).

7. Critical Context and Relation to the Literature

Quant GANs are notable among neural stochastic-process models for their explicit integration of financial no-arbitrage requirements and their use of deep generative adversarial machinery to achieve realism in simulated paths. They highlight the competitive advantage of TCNs for long-memory time-series modeling over recurrent architectures and contribute to the growing trend of data-driven models in quantitative finance.

A plausible implication is that, by satisfying the martingale property through analytic constraints, Quant GANs facilitate deployment in option pricing and scenario simulation, types of applications where purely generative neural models without such economic guarantees have heretofore been regarded as unreliable. The empirical and architectural details established by Wiese et al. provide a clear blueprint for further developments in data-driven financial simulation (Wiese et al., 2019).