Research map / Time series / Causal and generative time series
Market-GAN: Adding Control to Financial Market Data Generation with Semantic Context
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Market-GAN, a GAN-based hybrid generator that produces context-controlled financial OHLC price sequences conditioned on market dynamics, stock ticker, and history, trained in two stages on a new Contextual Market Dataset built from DJI stocks.
Problem
Existing financial simulators and time-series GANs cannot generate high-fidelity market feature data controlled by semantic context (market dynamics, stock ticker, history), and financial datasets lack context labels.
Approach
Combines a TimeGAN-style autoencoder and GAN with context supervisors that transfer context classification knowledge to the generator, an OHLC-constrained data transformation layer, a C-TimesBlock (RNN plus TimesNet-style 2D blocks), and a two-stage pre-training then adversarial training scheme; dynamics labels come from a linear-regression plus clustering market dynamics modeling method.
Key results
- Market-GAN with C-TimesBlock achieves context alignment loss L_d of 0.023 versus 0.055 for real data and 2.4-3.3 for baselines.
- Market-GAN has 0% OHLC fact violations, while baselines without the data transformation layer violate 15-50% of samples.
- Market-GAN (RNN variant) gives lowest SMAPE across TimesNet, TCN, LSTM, and GRU downstream predictors (e.g., 1.11 TimesNet vs 1.76 real-only), though Market-GAN(CTB) has a higher L_D (8.05) than TimeGAN with transformation (5.67 to 6.52).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use C-TimesBlock instead of RNN (ablation) | Market-GAN · DJI stocks, context alignment | L_d: 0.626 -> 0.023 | Helped |
| Use C-TimesBlock instead of RNN (ablation) | Market-GAN · DJI stocks, context alignment | L_l: 0.501 -> 0.100 | Helped |
| Use C-TimesBlock instead of RNN (ablation) | Market-GAN · DJI stocks, fidelity | L_D: 11.85 \pm 9.89 -> 8.05 \pm 5.60 | Helped |
| Use C-TimesBlock instead of RNN (ablation) | Market-GAN · DJI stocks, TimesNet prediction | SMAPE: 1.11\pm0.66 -> 1.26\pm0.75 | Hurt |
| Use C-TimesBlock instead of RNN (ablation) | Market-GAN · DJI stocks, TCN prediction | SMAPE: 6.39 \pm 0.92 -> 6.48 \pm 0.73 | Hurt |
| Use C-TimesBlock instead of RNN (ablation) | Market-GAN · DJI stocks, LSTM prediction | SMAPE: 9.89 \pm 14.9 -> 12.9 \pm 16.7 | Hurt |
| Use C-TimesBlock instead of RNN (ablation) | Market-GAN · DJI stocks, GRU prediction | SMAPE: 8.41 \pm 12.4 -> 11.1 \pm 14.2 | Hurt |
| Add data transformation layer (ablation) | TimeGAN · DJI stocks, context alignment | L_d: 2.847 -> 1.839 | Helped |
| Add data transformation layer (ablation) | TimeGAN · DJI stocks, context alignment | L_l: 0.533 -> 0.490 | Helped |
| Add data transformation layer (ablation) | TimeGAN · DJI stocks, fidelity | L_D: 5.67\mathbf{\pm} 2.53 -> 6.52 \pm 5.25 | Hurt |
| Add data transformation layer (ablation) | TimeGAN · DJI stocks, TimesNet prediction | SMAPE: 1.73\pm0.72 -> 1.36\pm0.48 | Helped |
| Add data transformation layer (ablation) | TimeGAN · DJI stocks, TCN prediction | SMAPE: 6.59\pm 0.70 -> 6.52 \pm 0.72 | Helped |
Surprising findings
- Applying the data transformation layer reduced fidelity-discriminator separability for some baselines (TimeGAN, RCGAN) yet increased L_D for others, illustrating a trade-off between context alignment and fidelity.
- The RNN variant of Market-GAN outperformed the proposed C-TimesBlock variant on downstream SMAPE and fidelity, despite CTB winning on context alignment.
Limitations
- Context is limited to market dynamics and stock ticker plus history; fundamental factors are not included.
- Fidelity (L_D) of the best Market-GAN variant is not the lowest; some baselines replicate real data more closely.
- Evaluation uses only 29 DJI stocks and a fixed window length of 30.
Experimental setup
Datasets: Dow Jones Industrial Average stocks daily OHLC (29 tickers, 2000-2023)
Benchmarks: Contextual Market Dataset (DJI OHLC 2000-2023)
Metrics: cross-entropy context alignment loss discriminator accuracy deviation L_D (fidelity) OHLC violation rate SMAPE prediction loss for data usability
Baselines: TimeGAN RCGAN SigCWGAN CGMMN
Builds on: TimeGAN TimesNet RCGAN SigCWGAN CGMMN
Closest papers by meaning
- MarS: a Financial Market Simulation Engine Powered by Generative Foundation Model (ICLR 2025 · similarity 0.87)
- Towards Financial World Modeling (NeurIPS 2026 · similarity 0.87)
- Generating Financial Time Series by Matching Random Convolutional Features (NeurIPS 2026 · similarity 0.87)
- LOB-Bench: Benchmarking Generative AI for Finance - an Application to Limit Order Book Data (ICML 2025 · similarity 0.87)
- A Set-Sequence Model for Time Series (NeurIPS 2026 · similarity 0.85)
- FinBen: A Holistic Financial Benchmark for Large Language Models (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.