• Zeth0s@lemmy.world
    link
    fedilink
    arrow-up
    10
    arrow-down
    1
    ·
    edit-2
    1 year ago

    You know there are people in bank and credit institutions that have been doing this for centuries? Probably millennia… EU explicitly requires that some of this is done by what you call AI (i.e. mathematical models) because they are fairer than humans and safer for customers and society

    Check basel III for an intro on the topic

    • Aceticon@lemmy.world
      link
      fedilink
      arrow-up
      9
      ·
      edit-2
      1 year ago

      Just a small correction because I worked around that area (not for loans but for investment), it’s all Algorithms rather than AI.

      Algorithms are basically mathematical formulas turned into code, whilst AI is a totally different beast that can produce quite different results on slightly different inputs and it’s not really made by turning mathematical models into code but rather it’s trained with real world data containing inputs and outputs and “somehow” finds the patterns in that data and can predict the correct outputs if given fresh, never seen before inputs.

      AI is probably used for fraud detection (and I expect nowadays it’s likely used in algorithmic trading to try and predict market movements) but unless a lot has changed since I was in the business, it’s not used for valuations.

      • Zeth0s@lemmy.world
        link
        fedilink
        arrow-up
        6
        arrow-down
        1
        ·
        edit-2
        1 year ago

        It was just to give an idea that what OP mentioned is already an established thing, fairer than alternatives.

        Most of the time trivial linear logistic regression is used in this context. Nowadays decision tree ensambles are pretty heavily used, which are ML. Simply they perform better with fewer data than neural networks on structured tabular data.

        What you refer to as AI is probably methods based on deep learning. The truth is that they work exactly as any other algorithm that you are referring to. They are used for regression and classification, same way as a standard linear regression. The difference is that the models are non linear, and their complexity is so that a lot of data are needed to train them.

        But conceptually one can absolutely create a credit score with deep neural networks. It is just an overkill, for performances that are likely worst than a random forest on relatively small training datasets

        Neural networks-based methods are indeed used in fraud detection