Periodic Reporting for period 1 - HiddenTipChains (Hidden Cascade Modelling for the Identification of Hidden Tip Chains in Stock Markets)
Reporting period: 2024-07-01 to 2026-06-30
Summary of the context and overall objectives of the project
Information and other forms of influence spread through social and interaction networks. Most existing cascade models assume that it is possible to observe who has received the information, when they received it, and how it subsequently spread. In many real-world settings this information is hidden. Instead, only indirect and potentially noisy signals of the underlying process are available.
HiddenTipChains addressed this challenge by developing methods for analyzing hidden information cascades. The main application was financial-market surveillance, where regulators can observe transactions and some social connections between investors, but they generally cannot directly observe whether an investor has received private information or through which pathway that information traveled. Distinguishing genuinely informative patterns from coincidental trading behavior is therefore a difficult but important problem for market integrity and investor protection.
The project combined topological data analysis, machine learning, graph neural networks, and probabilistic cascade modeling. It pursued three complementary objectives. First, it aimed to develop interpretable topological methods for identifying investors whose trading behavior is unusually consistent with access to non-public information. Second, it aimed to use machine-learning methods to understand how network structure and social connections are related to hidden information propagation. Third, it aimed to estimate the probabilities governing information transfer even when the true informed status of individual agents is not directly observable.
A central principle of the project was robust validation. Methods were first tested using synthetic cascade data for which the true underlying process was known before being applied to empirical financial-market data. Particular attention was given to reducing false-positive identifications, modeling noisy observations, understanding the effect of complex network structure, and developing methods that remain reliable when the observed network itself contains missing or uncertain links.
The project contributes to trustworthy data-driven market surveillance by providing tools for prioritizing patterns for expert investigation rather than making automatic determinations of wrongdoing. The methodological results are also relevant beyond finance. Similar hidden-cascade problems arise in epidemiology, cybersecurity, misinformation, and other networked systems where the underlying state of individual agents cannot be observed directly.
HiddenTipChains addressed this challenge by developing methods for analyzing hidden information cascades. The main application was financial-market surveillance, where regulators can observe transactions and some social connections between investors, but they generally cannot directly observe whether an investor has received private information or through which pathway that information traveled. Distinguishing genuinely informative patterns from coincidental trading behavior is therefore a difficult but important problem for market integrity and investor protection.
The project combined topological data analysis, machine learning, graph neural networks, and probabilistic cascade modeling. It pursued three complementary objectives. First, it aimed to develop interpretable topological methods for identifying investors whose trading behavior is unusually consistent with access to non-public information. Second, it aimed to use machine-learning methods to understand how network structure and social connections are related to hidden information propagation. Third, it aimed to estimate the probabilities governing information transfer even when the true informed status of individual agents is not directly observable.
A central principle of the project was robust validation. Methods were first tested using synthetic cascade data for which the true underlying process was known before being applied to empirical financial-market data. Particular attention was given to reducing false-positive identifications, modeling noisy observations, understanding the effect of complex network structure, and developing methods that remain reliable when the observed network itself contains missing or uncertain links.
The project contributes to trustworthy data-driven market surveillance by providing tools for prioritizing patterns for expert investigation rather than making automatic determinations of wrongdoing. The methodological results are also relevant beyond finance. Similar hidden-cascade problems arise in epidemiology, cybersecurity, misinformation, and other networked systems where the underlying state of individual agents cannot be observed directly.
Work performed from the beginning of the project to the end of the period covered by the report and main results achieved so far
The project developed a set of complementary methods for learning hidden information cascades when the true state of each individual is not directly observable. The scientific work progressed from identifying suspicious behavioral patterns to estimating the parameters governing information spread and finally to improving graph-based inference when the observed network is incomplete or noisy.
For the first objective, a topological data analysis method based on the Mapper algorithm was developed to identify investors whose trading behavior is unusually consistent with access to non-public information. The method combines expert-informed behavioral indicators with topological clustering and statistical validation. A conservative selection strategy was used to minimize false positives, which is particularly important in market-surveillance applications. In synthetic experiments with known ground truth, the method achieved a precision of 0.981. It was subsequently applied to 293,750 real transactions involving 1,586 investors and 119 companies. The identified investors showed stronger concentration of profitable trading around company announcements, greater abnormal performance relative to their own normal trading periods, and stronger connectivity to company insiders.
The second major line of work addressed the more general Hidden Cascade problem: the true informed or infected status of individuals is unknown, and only noisy behavioral symptoms can be observed. A likelihood-free machine learning method, Distribution Classification, was developed to estimate the parameters of the hidden spreading process by comparing distributions generated from observed and simulated cascades. The framework explicitly allows positive symptoms, negative symptoms, false positives, and unobserved carriers. It was validated on tree-like, highly connected, and empirical investor networks. In the financial application, estimated information-propagation probabilities were generally higher before company announcements than during normal periods.
The project then extended hidden-cascade inference from a single global spreading parameter to heterogeneous node-specific parameters. Symptom-aware cascade features were developed to summarize repeated noisy cascade observations, and a gated graph neural network was designed to estimate individual spreading probabilities. Dynamic Message Passing was adapted as a benchmark. The experiments showed that message-passing methods perform extremely well on tree-like networks, while the gated graph-neural-network approach is substantially more robust on heterogeneous networks containing many loops and alternative propagation paths.
Finally, the project investigated uncertainty in the network itself. A topology-aware Gaussian graph repair method was developed to improve graph neural networks when links are missing or spurious. The method adds sparse feature-consistent connections and reweights observed links using local topological information. Experiments on several benchmark networks showed improved robustness under both edge addition and edge deletion.
Together, these results established a coherent framework for studying hidden cascades under three important sources of uncertainty: hidden agent states, noisy behavioral observations, and imperfect network structure. The scientific work produced four closely connected research outputs and provides a foundation for future applications in financial surveillance as well as other networked systems such as epidemiology, cybersecurity, and misinformation analysis.
For the first objective, a topological data analysis method based on the Mapper algorithm was developed to identify investors whose trading behavior is unusually consistent with access to non-public information. The method combines expert-informed behavioral indicators with topological clustering and statistical validation. A conservative selection strategy was used to minimize false positives, which is particularly important in market-surveillance applications. In synthetic experiments with known ground truth, the method achieved a precision of 0.981. It was subsequently applied to 293,750 real transactions involving 1,586 investors and 119 companies. The identified investors showed stronger concentration of profitable trading around company announcements, greater abnormal performance relative to their own normal trading periods, and stronger connectivity to company insiders.
The second major line of work addressed the more general Hidden Cascade problem: the true informed or infected status of individuals is unknown, and only noisy behavioral symptoms can be observed. A likelihood-free machine learning method, Distribution Classification, was developed to estimate the parameters of the hidden spreading process by comparing distributions generated from observed and simulated cascades. The framework explicitly allows positive symptoms, negative symptoms, false positives, and unobserved carriers. It was validated on tree-like, highly connected, and empirical investor networks. In the financial application, estimated information-propagation probabilities were generally higher before company announcements than during normal periods.
The project then extended hidden-cascade inference from a single global spreading parameter to heterogeneous node-specific parameters. Symptom-aware cascade features were developed to summarize repeated noisy cascade observations, and a gated graph neural network was designed to estimate individual spreading probabilities. Dynamic Message Passing was adapted as a benchmark. The experiments showed that message-passing methods perform extremely well on tree-like networks, while the gated graph-neural-network approach is substantially more robust on heterogeneous networks containing many loops and alternative propagation paths.
Finally, the project investigated uncertainty in the network itself. A topology-aware Gaussian graph repair method was developed to improve graph neural networks when links are missing or spurious. The method adds sparse feature-consistent connections and reweights observed links using local topological information. Experiments on several benchmark networks showed improved robustness under both edge addition and edge deletion.
Together, these results established a coherent framework for studying hidden cascades under three important sources of uncertainty: hidden agent states, noisy behavioral observations, and imperfect network structure. The scientific work produced four closely connected research outputs and provides a foundation for future applications in financial surveillance as well as other networked systems such as epidemiology, cybersecurity, and misinformation analysis.
Progress beyond the state of the art and expected potential impact (including the socio-economic impact and the wider societal implications of the project so far)
HiddenTipChains introduced and developed the framework of Hidden Cascade Models (HCM) for situations in which the true states of agents are not directly observable and can only be inferred through indirect and potentially noisy observations. Existing information-cascade approaches commonly assume that activation states, activation times, or sufficiently informative cascade outcomes are observable. The project addressed this gap by formulating the hidden-cascade problem and developing complementary Topological Data Analysis (TDA), Machine Learning (ML), and probabilistic approaches for identifying and analyzing hidden information cascades in investor networks.
For the identification of suspicious trading, the project developed an expert-informed TDA Mapper approach that goes beyond conventional clustering by combining behavioral information with topological structure. The method was designed to identify investors whose trading behavior is unusually consistent with access to non-public information while giving particular attention to the reduction of false-positive identifications. It was first validated through Monte Carlo experiments with known ground truth and subsequently applied to empirical investor-level transaction data. The resulting framework provides an interpretable and conservative approach for profiling potentially opportunistic agents and supporting further human investigation in market-surveillance settings.
Building on the HCM framework introduced in the project, the research then developed a formal hidden-cascade observation model in which the true carrier status of an agent remains latent and only noisy symptoms of information reception are observed. These symptoms may include positive signals, negative signals, absence of observable behavior, false positives, and unobserved carriers. Distribution Classification was developed as a likelihood-free machine learning approach for estimating the parameters governing the underlying spreading process by comparing the distributions of observed and simulated cascade data. The method was validated on different synthetic network structures and on an empirical insider network. In the financial-market application, estimated information-transfer probabilities were generally higher during pre-announcement periods than during non-announcement periods, providing evidence consistent with increased information flow before public company announcements.
A further advance was the extension from a single or company-level spreading parameter to heterogeneous node-specific information-transfer parameters. Symptom-aware cascade features were developed to summarize repeated noisy observations, and a gated graph neural network was designed to infer node-level spreading probabilities. Dynamic Message Passing (DMP) was adapted to the same hidden-observation setting and used as a benchmark. The comparison clarified an important methodological issue already identified in the original proposal: DMP is highly effective on tree-like networks but becomes less reliable in networks containing many loops and alternative propagation pathways. The gated GNN approach provided a more robust alternative for heterogeneous and highly interconnected graphs.
The project also went beyond the original work plan by addressing uncertainty in the observed social network itself. In real applications, the available network may contain missing links, spurious links, or connections that do not correspond to actual information-transfer pathways. To address this problem, a topology-aware Gaussian graph repair method was developed to improve graph neural network inference under imperfect graph structure. The method introduces sparse feature-consistent links and reweights observed connections using local structural information, improving robustness under both missing-edge and noisy-edge settings.
Together, these results established HCM as a broader methodological framework for analyzing information propagation when both agent states and observations are incomplete or uncertain. The project progressed from the identification of suspicious agents to estimation of information-transfer probabilities to heterogeneous node-level inference and finally to robustness under uncertain network structure. This provides a substantially broader scientific framework than originally envisaged while remaining aligned with the central objective of HiddenTipChains: understanding how private information may propagate through investor networks when the underlying cascade itself is hidden.
The immediate potential impact is in financial-market surveillance, where the developed methods can support regulators and surveillance professionals by prioritizing unusual patterns for further expert investigation rather than making automatic determinations of wrongdoing. Further uptake will require continued validation on authorized financial datasets and further development of calibrated methods for identifying the most important information pathways. Beyond financial markets, the HCM framework has potential applications in other domains where the true states of agents and the transmission process are only indirectly observable, including epidemiology, cybersecurity, and information diffusion.
For the identification of suspicious trading, the project developed an expert-informed TDA Mapper approach that goes beyond conventional clustering by combining behavioral information with topological structure. The method was designed to identify investors whose trading behavior is unusually consistent with access to non-public information while giving particular attention to the reduction of false-positive identifications. It was first validated through Monte Carlo experiments with known ground truth and subsequently applied to empirical investor-level transaction data. The resulting framework provides an interpretable and conservative approach for profiling potentially opportunistic agents and supporting further human investigation in market-surveillance settings.
Building on the HCM framework introduced in the project, the research then developed a formal hidden-cascade observation model in which the true carrier status of an agent remains latent and only noisy symptoms of information reception are observed. These symptoms may include positive signals, negative signals, absence of observable behavior, false positives, and unobserved carriers. Distribution Classification was developed as a likelihood-free machine learning approach for estimating the parameters governing the underlying spreading process by comparing the distributions of observed and simulated cascade data. The method was validated on different synthetic network structures and on an empirical insider network. In the financial-market application, estimated information-transfer probabilities were generally higher during pre-announcement periods than during non-announcement periods, providing evidence consistent with increased information flow before public company announcements.
A further advance was the extension from a single or company-level spreading parameter to heterogeneous node-specific information-transfer parameters. Symptom-aware cascade features were developed to summarize repeated noisy observations, and a gated graph neural network was designed to infer node-level spreading probabilities. Dynamic Message Passing (DMP) was adapted to the same hidden-observation setting and used as a benchmark. The comparison clarified an important methodological issue already identified in the original proposal: DMP is highly effective on tree-like networks but becomes less reliable in networks containing many loops and alternative propagation pathways. The gated GNN approach provided a more robust alternative for heterogeneous and highly interconnected graphs.
The project also went beyond the original work plan by addressing uncertainty in the observed social network itself. In real applications, the available network may contain missing links, spurious links, or connections that do not correspond to actual information-transfer pathways. To address this problem, a topology-aware Gaussian graph repair method was developed to improve graph neural network inference under imperfect graph structure. The method introduces sparse feature-consistent links and reweights observed connections using local structural information, improving robustness under both missing-edge and noisy-edge settings.
Together, these results established HCM as a broader methodological framework for analyzing information propagation when both agent states and observations are incomplete or uncertain. The project progressed from the identification of suspicious agents to estimation of information-transfer probabilities to heterogeneous node-level inference and finally to robustness under uncertain network structure. This provides a substantially broader scientific framework than originally envisaged while remaining aligned with the central objective of HiddenTipChains: understanding how private information may propagate through investor networks when the underlying cascade itself is hidden.
The immediate potential impact is in financial-market surveillance, where the developed methods can support regulators and surveillance professionals by prioritizing unusual patterns for further expert investigation rather than making automatic determinations of wrongdoing. Further uptake will require continued validation on authorized financial datasets and further development of calibrated methods for identifying the most important information pathways. Beyond financial markets, the HCM framework has potential applications in other domains where the true states of agents and the transmission process are only indirectly observable, including epidemiology, cybersecurity, and information diffusion.