Machine learning for robo-advisors: testing for neurons specialization

  • Received October 1, 2019;
    Accepted November 18, 2019;
    Published December 9, 2019
  • Author(s)
  • DOI
  • Article Info
    Volume 16 2019, Issue #4, pp. 205-214
  • Cited by
    2 articles

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 International License

The rise of robo-advisor wealth management services, which constitute a key element of fintech revolution, unveils the question whether they can dominate human-based advice, namely how to address the client’s behavioral biases in an automated way. One approach to it would be the application of machine learning tools during client profiling. However, trained neural network is often considered as a black box, which may raise concerns from the customers and regulators in terms of model validity, transparency, and related risks. In order to address these issues and shed more light on how neurons work, especially to figure out how they perform computation at intermediate layers, this paper visualizes and estimates the neurons’ sensitivity to different input parameters. Before it, the comprehensive review of the most popular optimization algorithms is presented and based on them respective data set is generated to train convolutional neural network. It was found that selected hidden units to some extent are not only specializing in the reaction to such features as, for example, risk, return or risk-aversion level but also they are learning more complex concepts like Sharpe ratio. These findings should help to understand robo-advisor mechanics deeper, which finally will provide more room to improve and significantly innovate the automated wealth management process and make it more transparent.

view full abstract hide full abstract
    • Figure 1. Optimization algorithm
    • Figure 2. ANN architecture
    • Figure 3. Testing for neurons specialization: response to risk-aversion level
    • Table A1. Neurons elasticities (in percentages) to 1% change in the feature