Source-linked AI summary
Reverse Engineering Financial Markets with Majority and Minority Games using Genetic Algorithms
J. Wiesinger, D. Sornette, J. Satinover
TL;DR
The paper addresses whether real financial time series can reveal the organization of interacting boundedly rational market agents and support prediction. It fits agent-based virtual markets with genetic algorithms to Nasdaq data, selecting reconstructed ensembles that match in-sample series and testing one-step directional predictions. The reconstructed models outperform random strategies significantly across all five ABM types, while the study remains limited by the computational cost of its large parameter search.
Problem
The paper asks whether aggregate financial time-series behavior can reveal the internal organization of interacting market agents and be used for prediction.
Method
A genetic algorithm selects agent-based virtual stock markets whose generated series best match Nasdaq data, using the resulting ensembles for out-of-sample directional prediction.
Results
The reconstructed 3PGs outperform random strategies significantly for all five ABM types, with p-values of 0.001, 0.01, or 0.02.
Takeaways & Limitations
Simple agent interactions selected by genetic algorithms can reproduce useful aspects of a real financial time series and provide novel insight into its properties.
Takeaways & Limitations
The number of predictions remains limited by the high computational processing costs of searching the large genetic-algorithm parameter space.
Abstract
from arXiv · showhide
Using virtual stock markets with artificial interacting software investors, aka agent-based models (ABMs), we present a method to reverse engineer real-world financial time series. We model financial markets as made of a large number of interacting boundedly rational agents. By optimizing the similarity between the actual data and that generated by the reconstructed virtual stock market, we obtain parameters and strategies, which reveal some of the inner workings of the target stock market. We validate our approach by out-of-sample predictions of directional moves of the Nasdaq Composite Index.
1 Introduction
The paper asks whether observing financial time series can reveal the organization of interacting market agents and support prediction. It reverse engineers markets with agent-based models and genetic algorithms, extending prior Minority Game approaches from synthetic to real financial data.
- 1 Introduction: The study frames market efficiency as an emergent property of interacting boundedly rational agents, motivating reconstruction from aggregate price behavior.ABMs aggregate simple micro-level interactions into nonlinear macro-level patterns and financial stylized facts.
- 1 Introduction: Reverse engineering seeks to infer internal market properties and organization from macroscopic financial time series, then use the reconstruction for prediction.The method assumes price changes can be modeled as aggregated outputs of many interacting agents with limited knowledge, strategies, actions, and adaptation.
- 1 Introduction: The paper extends financial ABM research from explaining stylized facts toward reconstructing specific markets and validating the reconstruction through simple predictability.Validation tests prediction rather than only how well the synthetic market reproduces stylized facts.
- 1 Introduction: Prior reverse engineering used Grand Canonical Minority Games to evolve artificial ensembles matching a known black-box ensemble, but did not report results on real financial time series.Its matching criterion was cross-correlation between generated and black-box time series.
- 1 Introduction: A genetic algorithm is applied to optimize an ensemble of agents and strategies reproducing an external time series, rather than to model individual agents’ learning.The large, rapidly growing parameter space and rugged solution landscape motivate the search procedure.
2 Model / Methodology
The method fits several agent-based market models to Nasdaq return windows with a genetic algorithm, then uses the best reconstructed ensembles for one-step out-of-sample prediction. It compares model types across market regimes and evaluates directional success against standard strategies.
- The Nasdaq Composite index as the input: The input is daily adjusted Nasdaq Composite closing-price data, with performance compared across upward, downward, and no-trend regimes against buy-and-hold, sell-and-hold, and random strategies.The regime analysis is used to infer which investor populations may be dominant under different market conditions.
- The Nasdaq Composite index as the input: Computational cost limits the sample despite its 606 predictions, because the genetic algorithm searches a large parameter space.The sample is substantially larger than prior work’s few tens of predictions but remains constrained by processing requirements.
- Description of the different types of ABMs: The five ABMs differ in incentives: minority payoff, majority payoff, delayed payoff, or a 50–50 mixture of Grand Canonical Minority and Majority Game agents.The delayed games compare the next-step return with the previous decision, while the mixed game combines equal shares of GCMG and GCMjG agents.
- Description of the genetic algorithm: The pipeline trains each selected 3PG on an in-sample Nasdaq window and uses its aggregate decisions to predict the following out-of-sample return.Figure 2 illustrates a 25-day in-sample period followed by a one-step prediction, repeated across 606 days and five ABM types.
3 Validation by the statistical significance of the success rate of next-day prediction
The reverse-engineered ABMs predict next-day return signs significantly better than random strategies overall, with performance varying across market regimes and game incentives.
- The reverse-engineered 3PGs outperform random strategies overall, with p < 0.001 for GCMjG, delGCMjG, and MixG, and p = 0.01 and 0.02 for GCMG and delGCMG.Table 1 aggregates success rates and p-values across all 606 test days.
- Success rates are highly significant during trending periods but less significant during non-trending periods, making the procedure a good trend-detection method.The test set contains 202 trending days and 404 non-trending days.
- GCMjG, delGCMjG, and MixG are expected to perform in trends through their majority incentive, whereas GCMG and delGCMG also detect cycles or alternating corrections within trends.The authors interpret minority-game success during trends as evidence that those periods contain more than simple monotonic movement.
- All ABMs perform substantially worse in non-trending periods, with GCMG performing best, consistent with oscillations associated with its minority payoff.The minority incentive is linked in the text to frustration and oscillatory prices.
4 Conclusion
The paper concludes that genetic-algorithm-selected simple ABMs can reverse engineer real financial time series and provide insight into their properties. It also argues that the approach may extend to complex systems with underlying multi-agent structures.
- Genetic-algorithm-selected simple ABMs can reverse engineer real financial time series and provide novel insight into their properties.The authors attribute the validation performance to sophisticated macro-level structures emerging from aggregated micro-level interactions.
- The method is presented as more generally applicable to predicting complex systems with an underlying multi-agent structure.