A machine learning approach to support decision in insider trading detection

Consob - Scuola Normale Superiore di Pisa

FinTech n. 11 - November 2022 [PDF]
 

Abstract
The identification of potential cases of market abuse is a complex and demanding activity due to the enormous volume of data to be processed and the multiplicity of factors to be taken into account when assessing the operational behaviour of each investor. The objective of the analysis is to assess the possible continuity/discontinuity of the behaviour and its magnitude, in absolute and relative terms. The study, developed on anonymous data sets, describes the characteristics of two proofs of concept which, by employing artificial intelligence methods of the unsupervised machine learning type, could usefully support, once the experimentation of the prototypes is completed, the preliminary analyses for the identification of subjects suspected of insider trading behaviour, which could be followed by the investigation activity aimed at gathering additional useful elements to hypothesise single cases of abuse.
The first model uses a clustering analysis method. This method makes it possible to identify those groups of investors whose trading activity in the vicinity of a price-sensitive event is not only carried out in a rewarding direction, but is also characterised by operational discontinuity, both with respect to their previous trading history and the typical operations of the group to which they belong. In particular, the clustering analysis first elaborates on the trading pattern of each investor based on selected quantitative parameters (net balance in purchases and sales, trading concentration, and exposure). By processing these parameters, the k-means clustering methodology identifies homogeneous investors groups regarding a specific time horizon. Finally, by examining the evolution over time of the position taken by each investor, the analysis distinguishes those subjects who, in the vicinity of a price-sensitive event, are found to have modified their trading behaviour (so-called discontinuous investors).
The second model aims to identify small groups of investors who act in rewarding directionality and in a synchronised manner in the vicinity of a price-sensitive event (so-called insider rings). The methodology used, called Statistically Validated Networks - after characterising the trading activity of each investor in three possible states (buying, selling, buying-selling) - constructs a network of investors characterised by synchronous activity in terms of trading states and timing. Starting from a statistically validated network of investors, homogeneous groups of individuals with similar activity who traded in rewarding directionality with respect to a price-sensitive event are identified.

Authors
This paper, which is the result of a collaboration between Consob and the Scuola Normale Superiore di Pisa, was edited by:
- Piero Mazzarisi - SNS Pisa, Dipartimento di Economia Politica e Statistica, Università di Siena (piero.mazzarisi@sns.it);
- Adele Ravagnani - SNS Pisa (adele.ravagnani@sns.it);
- Paola Deriu - Consob, Markets Division (p.deriu@consob.it);
- Fabrizio Lillo - SNS Pisa, Dipartimento di Matematica, Università di Bologna (fabrizio.lillo@sns.it);
- Francesca Medda - Consob, General Direction (f.medda@consob.it);
- Antonio Russo - Consob, Markets Division (a.russo@consob.it).

The authors are the only responsible for errors and imprecisions. The opinions expressed in the paper are the authors' personal views and are in no way binding on CONSOB.

Keywords: Insider trading, Market abuse, Unsupervised learning, Statistically validated networks