The Future of Retail Financial Services: What Policy Mix for a Balanced Digital Transformation? by Sylvain Bouyon

Author:Sylvain Bouyon [Bouyon, Sylvain]
Language: eng
Format: epub
ISBN: 9781786604798
Google: PC3lAQAACAAJ
Goodreads: 34774555
Publisher: Centre for European Policy Studies
Published: 2017-08-31T09:24:13+00:00

2.‘First-order’ supervision versus ‘second-order’ supervision

One key question remains regarding the way supervisors can take action to enforce specific rules or ensure that some practices are in line with the enacted core principles. In the rapid development of big data, supervisors have to cope with two severe constraints: technical skills and resources. In order to deal with these two constraints, supervisors need to have sufficient in-house skills to understand the inner workings of the supervised processes and sufficient resources to supervise properly by taking action, if needed. To a certain extent, the constraints are intertwined.

Given that an increasing number of processes are complex algorithms that notably structure machine learning methods, it is generally too costly in terms of time and resources for the supervisors to understand the related coding and to ask for an adjustment of the algorithm, if necessary. Furthermore, such practices are likely to appear too invasive in many cases given that entire business models could be markedly affected as a result. Therefore, as highlighted by Wagner (2016), one possibility is to occasionally introduce case-by-case filters in order to modify the prima facie responses of the system. In this context, a distinction needs to be made between ‘first-order’ supervision and ‘second-order’ supervision. The former implies that supervisors require the business to change the coding of the algorithm itself in order to comply with the regulation. Within the latter supervisory framework, data inputs or outputs of an algorithm has to be limited without actually changing the algorithm itself.

The privileged approach is that supervisors by default take actions that are in line with a second-order supervisory framework: some of the data inputs or outputs of the algorithms that are unwanted will have to be removed (especially to address risks of discrimination). The decision to remove data should conform to the GDPR regarding the legitimacy of the purpose for which the data is processed and the adequacy and relevance of the data used for that purpose (see Recital 39 of this regulation).20 Such an approach will obviously imply that a proper input-outcome analysis is conducted before taking action.

For example, in order to limit the impact of certain behaviours on the pricing of health insurance, supervisors can instruct the insurer not to use the related behavioural data. A similar supervisory approach can be adopted regarding loans when some providers assess the creditworthiness of a specific consumer by using the financial situation of the users included in his Facebook network (some of these practices are likely to result in discriminatory selection).21 As regards data outputs, supervisors can, for instance, require one provider to limit individual online search results by filtering out certain products that might not be adequate to specific consumers.

In this context, the algorithm itself does not need to be changed in depth, rather its results simply need to be limited. This approach can help address the issues related to both the collection of data (in terms of privacy concerns) and the use of this data, without excessive intervention.22

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.