Explain why trading research is path-limited and how adaptive search and multiple testing can inflate apparent backtest performance.
Use classical simulation baselines, including bootstrap and stochastic volatility models, as interpretable benchmarks for synthetic data generation.
Select a synthetic-data approach that matches the data structure and downstream objective, including learned generators for time series and tabular financial data.
Diagnose generated data using stylized-fact, dependence, and task-based evaluation methods, including Train-Synthetic-Test-Real comparisons.
Assess privacy and generator-specific risks, including leakage, bias amplification, overfitting to the generator, and limited scenario novelty.
5.1
The quant’s dilemma
Quantitative strategy development must make inferences from a single realized history with limited crises, regime shifts, and correlation breakdowns. After testing 100 configurations with zero true Sharpe, the expected maximum in-sample Sharpe exceeds 2.5 (Bailey et al. 2015), making backtest overfitting near-certain. Synthetic data is positioned as simulation infrastructure that turns one realized history into a distribution of plausible histories — but the generator must reproduce tail outcomes and dependence structure, not just bulk distributional similarity.
5.2
Evaluating synthetic financial data
Bootstrap methods and parametric stochastic models (GBM, jump-diffusion, Heston, GARCH) remain strong reference points for evaluating learned generators. Bootstrap variants range from IID (preserves marginals, destroys temporal dependence) through stationary bootstrap (preserves short-range dependence). Any deep generative model should at least outperform these baselines on the diagnostics that matter for the downstream task.
5.3
Classical simulation baselines
Four families of learned generators are introduced: variational autoencoders (stable but may oversmooth), GANs (sharp but unstable), diffusion models (stable with iterative denoising), and LLM-based tabular generators (serialize rows as text). The key distinction is that learned generators can represent complex dependence structures that are difficult to specify parametrically, but can also fail silently by smoothing tails or collapsing modes.
5.4
Generative model taxonomy
Four GAN variants address specific limitations: TimeGAN adds supervised temporal objectives (TSTR ratio 1.70), Tail-GAN augments with VaR/ES penalties (reducing VaR error from 102% to 11.5%), Sig-CWGAN uses path-signature kernels for temporal fidelity (TSTR ratio 0.97), and GT-GAN uses neural ODEs for irregularly sampled data. Honest results and shared challenges (mode collapse, training instability) are documented.
5.5
GANs for financial time series
Diffusion-TS uses a Transformer encoder-decoder with trend-plus-seasonal decomposition, achieving KS statistic 0.06 and TSTR ratio 1.00 on 20 ETFs. Classifier-guided conditional generation enables regime-specific scenario production with a 2.6x volatility ratio between generated low and high-volatility samples. The key tradeoff is computational cost versus training stability.
5.6
Diffusion models for financial time series
Autoregressive language models generate synthetic tabular data by serializing rows as text. Using the GReaT framework with distilgpt2, fine-tuned in ~10 minutes on GPU, the approach achieves TSTR AUC-ROC of 0.84 and KS statistics below 0.035. Specific failure modes include invalid records from autoregressive generation and numerical fidelity limitations from token-level optimization.
5.7
LLMs for structured financial data
A three-dimensional evaluation framework: fidelity (marginal + temporal structure), utility (TSTR benchmarks), and privacy (empirical leakage + differential privacy). Applied to DP-GAN, strong privacy (epsilon=1) degrades fidelity by 6x while epsilon in [5,10] offers a practical sweet spot. The minimum validation checklist requires one distributional metric, one task benchmark, and one leakage test.
5.8
Applying the Fidelity–Utility–Privacy framework
5.9
Summary
00 Classical Simulation
01 Timegan
02 Tailgan Tail Risk
03 Sigcwgan Signatures
04 Gtgan Irregular
05 Diffusion Ts
06 Llm Tabular Great
07 Dp Gan
6 primer topics providing foundational concepts for this chapter.
Tim Bollerslev (1986) — Journal of Econometrics · 23212 citations
This paper introduces the Generalized Autoregressive Conditional Heteroskedasticity (GARCH) model, a significant extension of the ARCH model that allows for more flexible and parsimonious modeling of time-varying volatility by incorporating past conditional variances.
R. Cont (2001) — Quantitative Finance · 3429 citations
This paper reviews the 'stylized facts' of asset returns across various markets and instruments, highlighting statistical properties like heavy tails, volatility clustering, and the absence of linear autocorrelations, while also discussing the statistical issues that arise when analyzing financial time series.
This paper reviews the empirical stylized facts of financial time series and proposes a parsimonious agent-based model where volatility clustering emerges endogenously from investors updating their reaction thresholds asynchronously.
David H. Bailey and Marcos Lopez de Prado (2014) · 110 citations
The paper introduces the Deflated Sharpe Ratio (DSR), a statistic that adjusts performance metrics for the probability of backtest overfitting caused by multiple testing and non-normal returns.
Ian J. Goodfellow et al. (2014) — arXiv:1406.2661 [cs, stat]
Introduces GANs: a generative model trained via a two-player game where a generator tries to fool a discriminator, enabling sample generation using only backpropagation and no MCMC/inference machinery.
This paper proposes a practical, model-free way to estimate how likely your “best” backtest is a false positive caused by trying many strategy variations on the same data.
Jinsung Yoon et al. (2019) — Curran Associates, Inc. · 1342 citations
TimeGAN is a GAN for multivariate time series that adds a step-by-step supervised loss in a learned latent embedding space so generated sequences match both the marginal distributions and the temporal transition dynamics of real data.
This paper introduces CTGAN, a conditional generative adversarial network designed to address the challenges of modeling mixed discrete and continuous tabular data, particularly non-Gaussian, multimodal distributions and imbalanced categorical columns, outperforming existing GANs and Bayesian network baselines.
CorGAN generates privacy-preserving synthetic EHR-like records by using 1D CNN-based GANs plus a convolutional autoencoder decoder to better capture feature correlations than MLP-based baselines like medGAN.
This survey organizes and evaluates time-series-specific data augmentation methods (time, frequency, time-frequency, decomposition, generative, and automated policies) to help deep models generalize better when labeled time-series data is scarce or imbalanced.
Hao Ni et al. (2020) — arXiv:2006.05421 [cs, stat] · 38 citations
The paper proposes a conditional time-series GAN whose discriminator is an explicit, signature-based Wasserstein-1 proxy (Sig-W1), enabling stable training and better matching of temporal dependence than prior time-series GAN baselines.
Jonathan Ho et al. (2020) — arXiv:2006.11239 [cs, stat] · 28738 citations
This paper shows diffusion probabilistic models can generate high-quality images by learning to reverse a fixed Gaussian noising process, with a particularly effective training parameterization that predicts the added noise (ε) and connects diffusion training to denoising score matching and sampling to annealed Langevin dynamics.
Florian Eckerli and Joerg Osterrieder (2021) · 61 citations
This overview explains how GANs can generate statistically realistic synthetic financial data (especially time series) to mitigate data scarcity, and it includes a proof-of-concept test showing that off-the-shelf GANs can reproduce several financial “stylized facts” on S&P 500 log-returns.
Marcos López de Prado (2022) — The Journal of Portfolio Management
The paper derives closed-form (analytic) formulas for Sharpe-ratio Type I and Type II error rates under multiple testing (familywise), showing how selection among many backtests can dramatically inflate false discoveries and how power depends on sample size and non-normality.
Jinsung Jeon et al. (2022) — Advances in Neural Information Processing Systems · 64 citations
This paper introduces GT-GAN, a novel generative adversarial network that can synthesize both regular and irregular time series data without model changes, outperforming existing specialized methods across various datasets and levels of data irregularity.
This paper introduces GReaT, a novel method that leverages large language models (LLMs) to generate realistic synthetic tabular data by converting tabular data into a textual representation, fine-tuning an LLM, and sampling from it, achieving state-of-the-art performance.
Tomonori Takahashi and Takayuki Mizuno (2024) · 10 citations
This paper proposes using denoising diffusion probabilistic models (DDPMs) combined with wavelet transformation to generate synthetic financial time series that accurately replicate all known stylized facts, including fat tails, volatility clustering, intraday seasonality, and cross-correlations, outperforming existing GAN and VAE-based methods.
The paper tests whether vanilla GANs can reproduce key “stylized facts” of financial time series (random walk, mean reversion, jumps, stochastic volatility) and finds they often match marginal return distributions but struggle with detailed temporal structure and multivariate dependence—highly sensitive to generator architecture.
TSGM is an open-source framework that standardizes how to generate and evaluate synthetic time series using a wide range of generative, simulation-based, and augmentation methods, with a broad metric suite covering utility, similarity, privacy, fairness, and diversity.
Adil Rengim Cetingoz and Charles-Albert Lehalle (2025)
The paper argues that synthetic financial return data from generic generative models can mislead portfolio/risk conclusions—especially for long-short portfolios—because (i) you cannot “create information” beyond the original sample size and (ii) standard generative losses learn the wrong directions (high-variance PCs) for portfolio optimization.
Tail-GAN is a GAN-based scenario generator trained to match portfolio tail risk (VaR/ES) for user-chosen static and dynamic trading strategies, producing more realistic tail-loss scenarios than standard GAN losses (e.g., Wasserstein) especially out-of-sample.
The paper proposes a factor-conditioned diffusion model that generates next-day cross-sectional return distributions for many stocks and uses those samples to drive daily mean–variance portfolio optimization, improving performance on China A-shares versus standard moment estimators.
CORRGAN: Sampling Realistic Financial Correlation Matrices Using Generative Adversarial Networks
G. Marti
The paper introduces CorrGAN, a GAN-based method to generate synthetic but realistic financial correlation matrices that reproduce key empirical “stylized facts,” enabling better stress testing and more falsifiable empirical comparisons.
The paper proposes using Wasserstein GANs to generate realistic, data-anchored Monte Carlo simulations—reducing researcher discretion in simulation design and enabling finite-sample evaluation of causal estimators on distributions that mimic a target dataset.
The paper argues for replacing standard Walk-Forward backtests with Monte Carlo simulations on synthetic data to create 'tactical' algorithms optimized for specific market regimes rather than 'all-weather' strategies.
The paper proposes a “market generator” that trains a conditional VAE on path signatures (rather than raw returns) and evaluates realism with a signature-kernel MMD test, aiming to generate realistic financial paths even with small, irregular, or oscillatory datasets.
ABIDES is an open-source agent-based market simulation environment designed to facilitate research in market applications by simulating the interactions of thousands of trading agents with an exchange, incorporating realistic network latency and message protocols.
Petter N. Kolm and Nicholas Westray (2022) — The Journal of Portfolio Management
The paper proposes a fast, practitioner-friendly simulator that generates realistic meta-order flow (sizes and execution speeds/PoV) using mean-variance portfolio optimization plus a calibrated PoV model, and shows it can be used to estimate how long it takes to statistically detect whether a client’s order flow has alpha.
Koray D. Simsek (2023) — The Journal of Portfolio Management
The paper is a practitioner-oriented overview of Monte Carlo simulation—how to generate scenarios, model uncertainty and correlation, and summarize outputs—with concrete finance examples in option pricing, portfolio insurance, and VaR/ES risk measurement.
VolGAN is a conditional GAN trained on SPX options data to simulate realistic one-day-ahead joint scenarios of the SPX return and the full implied volatility surface while controlling static arbitrage via scenario reweighting, and it can be used to build data-driven option hedges that outperform delta and delta–vega hedging in their case study.
This paper introduces novel GAN architectures incorporating stochastic masking, multi-frequency data, Hidden Markov Models, and vector quantization to generate diverse and realistic financial time series data, specifically for simulating benchmarks and futures data.
TransFusion introduces a novel diffusion model combined with transformers to generate high-quality, long-sequence synthetic time series data, outperforming existing GAN-based and RNN-based methods and addressing their limitations like mode collapse and inability to capture long-term dependencies.
The paper replaces model-based “Greek” hedging with a scenario-based hedge computed from a conditional generative model trained on market data, producing sparse, transaction-cost-aware hedge ratios that adapt to market conditions.
Ben R. Marshall et al. — Journal of Banking & Finance
A rigorous empirical test of 28 candlestick patterns on DJIA stocks using a bootstrap methodology finds no evidence of profitability, supporting market efficiency in large-cap US equities.
A roundtable of leading academics and Man Group practitioners discussing the specific challenges of applying Generative AI to quantitative finance, identifying look-ahead bias in backtesting as the primary obstacle.
Conditional Generative Adversarial Nets
Mehdi Mirza and Simon Osindero (2014) — ArXiv
The paper introduces conditional GANs (cGANs), a simple extension of GANs that generates samples controlled by side information (e.g., class labels or image features), enabling targeted generation and one-to-many predictions like image tagging.
David Buckle and Marielle de Jong (2024) — The Journal of Portfolio Management
The paper shows how Monte Carlo simulation can be used (even in Excel) to estimate portfolio risk measures and optimize portfolios under non-normal return distributions using copulas and expected-utility optimization.
