Alleviating Nonidentifiability: A High-Fidelity Calibration Objective for Financial Market Simulation With Multivariate Time Series Data

Ke Tang, Chenkai Wang, Junji Ren, Peng YangIEEE Trans. Comput. Soc. Syst 2025 · DOI

TL;DR

This paper tackles the "non-identifiability" problem in financial market simulations, where different parameter settings can produce indistinguishable results, undermining model reliability. They propose a novel calibration objective that leverages multiple time series features, theoretically and empirically demonstrating that this approach significantly reduces non-identifiability and improves simulation fidelity.

What Problem Does This Paper Address?

Agent-based models (ABMs) are powerful tools for understanding complex social systems like financial markets. However, their usefulness hinges on "model fidelity"—how accurately they represent reality. A key challenge in achieving this fidelity is "non-identifiability" during model calibration. This occurs when multiple distinct sets of model parameters yield simulated data that are statistically indistinguishable from real-world data, according to a given discrepancy metric. This ambiguity means that even if a model fits the data, we can't be sure which parameter set is the "true" one, leading to unreliable "what-if" scenario analyses. Existing calibration methods often rely on a single time series feature (like mid-price), which exacerbates this problem. This paper aims to address this by developing a method to improve parameter identifiability and thus simulation fidelity.

Methods

The core contribution is a new calibration objective function designed to alleviate non-identifiability. The authors first theoretically define non-identifiability as the probability of a randomly sampled parameter falling into a "non-identifiable set." They then prove that incorporating more distinct time series features into the calibration process exponentially reduces this probability.

To implement this, they propose a maximization-based aggregation function. This function takes the maximum of individual discrepancy metrics calculated for each feature. The intuition is that by forcing the model to simultaneously fit multiple features well, the parameter space that satisfies all conditions shrinks significantly, leading to a more identifiable parameter. They also theoretically demonstrate that linearly dependent features do not contribute to this alleviation, emphasizing the importance of selecting diverse features.

Datasets

The paper uses two types of data for evaluation:

  • Synthetic Data: 10 synthetic time series datasets were generated using the Preis-Golke-Paul-Schneid (PGPS) model, a popular ABM for financial markets. These datasets were created by sampling parameters within specified ranges and represent various market dynamics.
  • Real Data: One real-world time series dataset from the Shenzhen Stock Exchange (stock code 000001.sz) was used.

Key Insights and Ideas

  • Non-identifiability is a fundamental challenge: The paper formally defines and quantifies non-identifiability, moving beyond anecdotal observations.
  • More features = better identifiability: The core theoretical insight is that increasing the number of independent time series features used for calibration exponentially reduces the non-identifiability of the model parameters.
  • Maximization is key for aggregation: The proposed maximization-based aggregation of individual feature objectives is shown to be the unique method that preserves the intersection property of non-identifiable sets, effectively shrinking the ambiguous parameter space.
  • Feature independence matters: Linearly dependent features offer little to no benefit in reducing non-identifiability, highlighting the need for careful feature selection.
  • Model-agnostic approach: The proposed calibration objective is designed to work with various ABM structures and discrepancy metrics, making it broadly applicable.

Key Results

  • Exponential reduction in non-identifiability: Empirical results show that as the number of features increases, the non-identifiable set shrinks exponentially, aligning with theoretical predictions. For example, using 6 features significantly reduced the non-identifiable set compared to using just one.
  • Improved simulation fidelity: The new calibration objective consistently led to better simulation performance across multiple metrics (Wasserstein distance and Mean Squared Error) compared to calibrating with single features. For instance, using 6 features resulted in the best simulation performance on 7 out of 10 synthetic datasets.
  • Robustness to noise: The method's effectiveness was shown to be robust across different levels of observation noise (epsilon values).
  • Successful real-world application: The approach successfully calibrated a real market dataset, demonstrating its practical applicability beyond synthetic benchmarks. The simulated mid-price using 3 features showed a 37.9% improvement in MSE compared to using only 1 feature.

Limitations and Future Work

  • Feature selection: While the paper suggests selecting independent features, the process of identifying and selecting the "best" set of features for a given market and model could be further automated.
  • Learning-based feature extraction: The current work relies on handcrafted features (e.g., from the Limit Order Book). Future work could explore using advanced representation learning techniques to automatically extract more meaningful and informative features.
  • Dynamic recalibration: The proposed method focuses on a static calibration. Developing adaptive recalibration methods that can dynamically adjust parameters over time to maintain high fidelity in changing market conditions is a promising direction.
  • Generalizability: While the approach is demonstrated on financial markets, its application to other complex systems generating multivariate time series data (e.g., weather forecasting, traffic flow modeling) warrants further investigation.

Summary updated

BibTeX
@article{tang2025alleviating,
  author = {Ke Tang and Chenkai Wang and Junji Ren and Peng Yang},
  title = {{Alleviating Nonidentifiability: A High-Fidelity Calibration Objective for Financial Market Simulation With Multivariate Time Series Data}},
  journal = {IEEE Trans. Comput. Soc. Syst},
  year = {2025},
  doi = {10.1109/tcss.2025.3574236},
}
Figure