Controllable Financial Market Generation with Diffusion Guided Meta Agent
Source
- Raw Markdown: paper_diga-2026.md
- PDF: paper_diga-2026.pdf
- Preprint: arXiv 2408.12991
- Official code: microsoft/TimeCraft/DiGA
Status And Credibility
The paper was first posted to arXiv on 2024-08-23 and the inspected arXiv metadata reports version 3, updated 2026-01-15. The arXiv comment says it is to appear at AAAI 2026 as an oral paper. The official code is in the Microsoft TimeCraft repository.
Core Claim
DiGA, called DigMA in the paper text, formulates controllable financial market generation as conditional generation over market state and order flow. A diffusion meta-controller generates time-evolving distribution parameters, while a meta agent with financial priors turns those states into orders.
Key Contributions
- Defines control targets such as asset return, volatility, and rare market events.
- Uses a conditional diffusion model over market-state distribution parameters.
- Uses a meta agent with economic priors to generate order flow.
- Evaluates controllability, fidelity, high-frequency trading utility, and computational efficiency.
Evidence And Results
The paper reports low errors between target indicators and generated order-flow statistics, improved stylized-fact fidelity versus baselines, and downstream trading-agent experiments. The generated data are not just passive samples; they are used as a financial simulation environment.
Limitations
- The paper focuses on single-stock order flow and notes multi-asset correlation as future work.
- Market simulation is a specialized domain with strong microstructure assumptions.
- The generated environment supports trading experiments, but the causal validity of generated market impact depends on simulator assumptions.
- Financial actions and injected orders are closer to control inputs than ordinary exogenous variables, but they are not a general time-series control benchmark.
Foundation TSFM Relevance
| Agenda slot | Verdict | Evidence | Missing pieces |
|---|---|---|---|
| Control and counterfactuals | partially closes | Generates market trajectories under scenario controls and evaluates trading-agent utility. | Needs multi-asset, realistic agent-interaction, and causal market-impact validation. |
| Time-series generation | partially closes | Diffusion meta-controller plus meta agent generates order-flow sequences. | Domain-specific order-level interface, not generic numeric time-series generation. |
| Digital-world robots | adjacent | Generated markets can train or test trading agents in a risk-free environment. | Needs explicit action-observation logs and safety constraints for broader digital systems. |