Voices

How to educate an AI model: What financial advisors should know

As AI continues to permeate the financial services industry, understanding the mechanics of how generative artificial intelligence is trained will help advisors collaborate in helping them learn

It all starts with understanding its relationship to data. The two most prevalent methodologies AI uses to learn are supervised and unsupervised learning. Supervised learning is akin to a mentor-apprentice relationship. An AI model is provided with training data (input-output pairs). The model then learns the relationship between the input and output, with its primary objective to minimize the difference between its predictions and actual results. Imagine a CFO who wants to forecast quarterly sales for the next year. She has historical data of advertising spend (input) and actual sales figures (output) for the past 10 years. Using supervised learning, an AI model trains on this data to understand the relationship between advertising spend and sales. Then the CFO can input predicted advertising spends for each quarter, and the AI model will provide an estimated sales figure.

Eli Gill of Paro
Eli Gill, vice president of engineering, product, and AI at Paro

In unsupervised learning, AI is provided with data that doesn't have clear input-output pairs or labeled responses. Instead, the model looks for inherent structures, patterns or similarities within the data, often grouping similar data points together. Consider trying to identify anomalies in a company's extensive transaction records. Using unsupervised learning, an AI system would sift through this data and group transactions by similarity. Any transaction that doesn't fit into a recognized group would be flagged. This is invaluable for detecting fraudulent activities that a typical rule-based system may overlook.

The importance of high-quality training data
Just as human apprentices or trainees learn best from accurate and unbiased information, AI models require high-quality training data. High-quality data should be accurate, relevant, comprehensive and unbiased; only then can AI models serve the financial services and accounting sectors most effectively. Imagine training a model to identify fraudulent transactions. If the data only includes past patterns while ignoring new or less common methods of fraud, the AI system may be blind to novel threats.

Bias is another concern. For instance, if a lending institution's past data exhibits bias against a certain group of borrowers, an AI model trained on this data might perpetuate this bias, leading to unethical and potentially illegal lending practices.

Rigorous data preparation and model validation
Preparing data for AI is like preparing a financial statement: It requires precision, diligence and an understanding of the end goal. Data might need to be cleaned (removing inconsistencies or errors), normalized (scaling data to a standard format) or even augmented (enhancing data to improve training). 

Once an AI model is trained, it needs rigorous validation. Before trusting an AI model with stock price predictions, for example, one might test its forecasts against a set of unseen data to gauge its accuracy. Regular validation and retraining ensure that models remain relevant, and this holds especially true in finance and accounting.

Correcting AI myths
Among the myriad misconceptions surrounding AI, three are especially prevalent when it comes to training and trusting AI models. Here are the facts:

  • While AI is powerful, its predictions are based on patterns in past data. It cannot foresee a black swan event. 
  • AI aids human professionals, but it does not replace them. A tax software may suggest deductions, for instance, but a seasoned CPA would consider the nuances and intricacies that software may overlook.
  • More data doesn't necessarily lead to better results. Feeding an AI model copious amounts of irrelevant data can confuse it, leading to inaccurate predictions. 

The bottom line is that artificial intelligence can be an invaluable tool for financial services and accounting professionals, enhancing accuracy, efficiency and insights. But like any tool, its efficacy depends on the hands that wield it. 

For reprint and licensing requests for this article, click here.
Technology Artificial intelligence Machine learning Practice management Wealth management
MORE FROM FINANCIAL PLANNING