Network Graph Analysis for Trading Fraud Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for detecting market manipulation, such as pump-and-dump schemes and order spoofing, suffer from high false-positive rates due to their reliance on limited anomaly indicators and inability to distinguish between artificial and natural changes in security transaction behavior, particularly in low-liquidity and low-information securities like penny stocks.
Innovation Solution
A server system that generates a network graph from transaction datasets, transforms it into a time-dependent graph using similarity scores weighted by exponential decay, and performs network analysis to identify outlier clusters and potential fraud alerts, thereby improving detection accuracy and reducing false positives.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional techniques analyze independent account activity with limited anomaly indicators, then the detection process is simple, but the false-positive rate is high
Solution Approach 1:
The patent segments the analysis by creating separate network graphs for different time periods (e.g., pre-event, event, post-event windows) and different security types. This allows complex market manipulation detection to be broken into manageable temporal segments, improving detection accuracy without overwhelming computational complexity.
Solution Approach 2:
The patent adds temporal dimensionality by transforming static network graphs into time-dependent network graphs that evolve across different time windows. This dimensional transformation enables the system to capture dynamic trading patterns and distinguish between legitimate and manipulative activity more effectively.
2Reliability
If the system analyzes all security transactions to detect market manipulation, then detection coverage is comprehensive, but computational resources are consumed excessively
Solution Approach 1:
The patent extracts and focuses analysis on specific time windows around events of interest (e.g., earnings announcements, regulatory filings) rather than processing all transactions continuously. This extraction approach maintains comprehensive detection coverage for critical periods while significantly reducing computational resource consumption during normal periods.
Solution Approach 2:
The patent applies partial analysis by focusing computational resources on specific subsets of data (certain security types, specific time windows, particular account relationships) rather than uniformly analyzing all transactions. This partial action approach maintains effective detection where needed while conserving computational resources.
3Measurement precision
If the system uses time-dependent network graphs with exponential decay weighting, then temporal patterns are captured accurately, but processing time increases
Solution Approach 1:
The patent implements periodic analysis by dividing the time series into discrete windows (pre-event, event, post-event) and processing each window separately with exponential decay weighting. This periodic approach captures temporal patterns accurately within each window while avoiding the computational burden of continuous real-time processing across the entire time series.
4Reliability
If the system analyzes co-traded securities and cohorts, then detection of market manipulation improves, but the complexity of data relationships increases
Solution Approach 1:
The patent merges multiple analysis dimensions by combining network graphs that represent different relationships (co-traded securities, cohorts, account interactions) into a unified time-dependent network graph. This merging approach handles complex data relationships systematically, improving manipulation detection accuracy while providing a structured framework to manage the complexity.
Data Source
AI summary
A system, apparatus, method, and non-transitory computer readable medium for performing co-trading changepoint detection may include a server caused to, receive a transaction dataset, the transaction dataset including a plurality of transactions for analysis, each transaction of the plurality of transactions associated with a user account involved in the transaction, a transaction object involved in the transaction, and transaction timestamp information corresponding to a time of occurrence of the transaction, the user account being one of a plurality of user accounts, generate a first network graph based on the transaction dataset, the first network graph including object nodes and user account nodes representing each of the transactions of the transaction dataset, transform the first network graph into at least one time-dependent transformed graph, and generate at least one potential fraud alert based on results of the network analysis.