Paper: arXiv 2410.14587
Authors: Namid R. Stillman, Rory Baggott
Abstract
Deep generative models are becoming increasingly used as tools for financial analysis. However, it is unclear how these models will influence financial markets, especially when they infer financial value in a semi-autonomous way. In this work, we explore the interplay between deep generative models and market dynamics. We develop a form of virtual traders that use deep generative models to make buy/sell decisions, which we term neuro-symbolic traders, and expose them to a virtual market. Under our framework, neuro-symbolic traders are agents that use vision-language models to discover a model of the fundamental value of an asset. Agents develop this model as a stochastic differential equation, calibrated to market data using gradient descent. We test our neuro-symbolic traders on both synthetic data and real financial time series, including an equity stock, commodity, and a foreign exchange pair. We then expose several groups of neuro-symbolic traders to a virtual market environment. This market environment allows for feedback between the traders belief of the underlying value to the observed price dynamics. We find that this leads to price suppression compared to the historical data, highlighting a future risk to market stability. Our work is a first step towards quantifying the effect of deep generative agents on markets dynamics and sets out some of the potential risks and benefits of this approach in the future.
Complexity vs Empirical Score
- Math Complexity: 6.5/10
- Empirical Rigor: 7.0/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper employs advanced mathematical modeling including stochastic differential equations (SDEs) and gradient-based calibration, placing it in the high math complexity range, while also conducting experiments on real financial time series with a virtual market environment, indicating substantial empirical rigor.
Research Flowchart
flowchart TD
A["Research Goal:<br>Quantify Impact of Deep Generative<br>Agents on Market Dynamics"] --> B{"Methodology"}
B --> C["Data Inputs<br>(Synthetic & Real: Equity, Commodity, Forex)"]
B --> D["Create Neuro-Symbolic Traders:<br>Vision-Language Models calibrating<br>Stochastic Differential Equations via Gradient Descent"]
C --> E["Virtual Market Environment<br>with Price Feedback Loop"]
D --> E
E --> F["Computational Simulation:<br>Agents make buy/sell decisions<br>based on inferred fundamental value"]
F --> G["Key Findings/Outcomes:<br>1. Price suppression observed<br>2. Potential risk to market stability<br>3. Framework for future analysis"]
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes neuro‑symbolic traders that use vision‑language models to discover stochastic differential equation models of fundamental value and trade in a virtual market.
Source sentence
We develop a form of virtual traders that use deep generative models to make buy/sell decisions, which we term neuro-symbolic traders
Assumes 5
- The fundamental value Vt can be approximated by a stochastic differential equation (SDE). model
Source sentence
We assume that the fundamental value, 𝑉𝑡, can be approximated by a stochastic differential equation (SDE)
- Price impact is linear (Kyle’s lambda). market
Source sentence
We assume a linear price impact model, also known as Kyle’s lambda 𝜆𝐾
- Noise‑trader demand is drawn from a normal distribution with scale 0 and standard deviation 10⁻¹. market
Source sentence
For all results, we assume the demand from noise traders is sampled from a normal distribution with scale 0 and standard deviation 10−1
- Kyle’s lambda is fixed at 10⁻¹. market
Source sentence
We also fix Kyle’s lambda as 10−1
Show 1 more
- All agents use the Claude‑3.5‑Sonnet model with a fixed temperature. execution
Source sentence
All results in this work are with the temperature of…
Data
- Synthetic data synthetic unknown
Source sentence
We test our neuro-symbolic traders on both synthetic data and real financial time series
- Real financial time series (equity stock, commodity, FX pair) market-prices unknown
Source sentence
We test our neuro-symbolic traders on both synthetic data and real financial time series, including an equity stock, commodity, and a foreign exchange pair
Costs
Unclear
Finds 3
- Exposing neuro‑symbolic traders to the virtual market leads to price suppression compared to historical data, indicating a risk to market stability. negative
Source sentence
We find that this leads to price suppression compared to the historical data, highlighting a future risk to market stability
- Model calibration reduces moment error from 7.7×10⁻² to 7.0×10⁻² with gradient‑based calibration (or from 8.4×10⁻² to 8.2×10⁻² using the VLM directly). statistical
Source sentence
the error in the calculated moments decreases from 7.7 × 10−2 to 7.0 × 10−2 when using gradient-based calibration or from 8.4 × 10−2 to 8.2 × 10−2 when using the VLM directly
- Neuro‑symbolic agents consistently increase model complexity, adding terms such as square‑root drift, mean‑reversion, sinusoidal and time‑dependent volatility. other
Source sentence
the neuro-symbolic agent consistently increases the complexity of the model. In total, there are ten broad groups of different model extensions. The most commonly introduced are a square root drift term (mentioned 11 times), mean-reversion term (mentioned 9 times), sinusoidal term (5 times) and time-dependent volatilit
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 12 related papers). Coverage is partial, so this is neither replication nor refutation.
1 paper in the archive
Code
No repository by the authors found in the paper.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (50)
- Isaac Newton, Daniel Defoe and the dynamics of financial bubbles (2018)
- Generating synthetic data in finance: opportunities, challenges and pitfalls (2020)
- Simulation techniques in financial risk management (2015)
- A data-driven market simulator for small data environments (2020) — link
- An agent-based financial market simulator for evaluation of algorithmic trading strategies (2014)
- Traders’ long-run wealth in an artificial financial market (2003)
- High frequency trading strategies, market fragility and price spikes: an agent based model perspective (2019)
- Agent-based simulation of wealth, capital and asset distribution on stock markets (2017)
and 42 more
- Market Ecology: Trading Strategies and Market Volatility (2024)
- Volatility clustering in financial markets: a microsimulation of interacting agents (2000)
- Ai in finance: challenges, techniques, and opportunities (2022)
- Generative models: An interdisciplinary perspective (2023)
- Build, compute, critique, repeat: Data analysis with latent variable models (2014)
- On the measure of intelligence (2019) — link
- A useful method for model-building (1962)
- Black-box model risk in finance (2023)
- Deep Generative Modeling (2024)
- Neural stochastic agent-based limit order book simulation: A hybrid methodology — → 2303.00080
- Knowledge graph-based event embedding framework for financial quantitative investments (2020)
- Neurosymbolic artificial intelligence (why, what, and how) (2023)
- Improving coherence and consistency in neural sequence models with dual-system, neuro-symbolic reasoning (2021)
- What can we learn from collective human opinions on natural language inference data? (2020) — link
- The Wisdom of Partisan Crowds: Comparing Collective Intelligence in Humans and LLM-based Agents (2024)
- Complex economics: individual and collective rationality (2010)
- Large language model based multi-agents: A survey of progress and challenges (2024) — link
- Automated finance: The assumptions and behavioral aspects of algorithmic trading (2012)
- Herd behavior and contagion in financial markets (2008)
- Peripheral vision: Economic markets as calculative collective devices (2005)
- Claude 3.5 Sonnet Model Card Addendum (2024)
- Handbook of collective intelligence (2022)
- Information aggregation and collective intelligence beyond the wisdom of crowds (2022)
- Automated statistical model discovery with language models (2024) — link
- Agent-based modeling in economics and finance: Past, present, and future (2022)
- Agent-based simulation of a financial market (2001)
- Financial market simulation (2004)
- Stochastic processes from 1950 to the present (2009)
- Differences of opinion in financial markets (1989)
- Chain-of-thought prompting elicits reasoning in large language models (2022)
- Stochastic differential equations and applications (2007)
- Continuous martingales and Brownian motion (2013)
- Understanding uncertainty shocks and the role of black swans (2014)
- The COVID-19 black swan crisis: Reaction and recovery of various financial markets (2022)
- The frontier of simulation-based inference (2020)
- Black-box Bayesian inference for economic agent-based models (2022) — link
- Robust pricing and hedging via neural sdes (2020) — link
- O n N eural D ifferential E quations (2021)
- Co-existence of trend and value in financial markets: Estimating an extended Chiarella model (2020)
- Augmenting physical models with deep networks for complex dynamics forecasting (2021)
- Neural sdes as infinite-dimensional gans (2021)
- JAX : composable transformations of P ython+ N um P y programs (2018) — link
1 of 50 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.