Reliability And Outcome Bias Issues In AI-Driven Forecasting Practices

Amid all the hype around the economic potential of AI technologies, there is a growing risk of data analysis overkill in many applications. That risk is particularly high for the forecasting and decision-making models being proposed in social contexts such as economic policy, financial investment, and corporate decisions. Common research practices in those areas keep focusing on incidents of statistical discoveries. They omit the substantial reliability issues stemming from the nature of the data that offers very limited 'learning potential' for the machine learning (ML) algorithms. In this paper, I focus on the use of ML algorithms applied to such forecasting problems. I illustrate the reliability issues with a detailed example that builds a stock investment strategy by using the XGBoost algorithm on a large data set. The example demonstrates how easy it is to discover seemingly interesting random patterns when we fit over-parameterized models on historical data. The results also offer practical methods to investigate the statistical flukes and the reliability issues that are concealed by complex algorithms of artificial intelligence being blended with natural human ignorance, as seen in popular practice.