From Prediction to Precision: The Evolving World of AI Insights

Predictive AI is a subset of artificial intelligence that uses past data and machine learning to guess what might happen in the future. Using AI to make forecasts is now more accessible than ever, but businesses need to be extremely careful in designing these models. General AI best practices are just as relevant for predictive models, and they can make the difference between getting a jump on the market or losing a fortune.

Predictive AI works by looking at past data, identifying patterns, and then making forecasts based on current trends. By looking at patterns and connections in big sets of data, these AI models can predict things like customer habits, market trends, equipment breakdowns, and even health issues.

Predictive AI has already been extensively utilized across research, academia, and the corporate and financial sectors, each leveraging its predictive capabilities to drive innovation and efficiency. Examples of this are in our everyday life as well, without us even noticing it – personalized content recommendations on your favorite streaming platforms such as Netflix as well as navigation apps that suggest the best routes based on historic and real-time traffic data for instance.

As predictive AI becomes democratized, however, companies new to the technology need to take precautions. Even with the advances in Generative AI, widescale use of AI Predictive modeling needs to be implemented with care.

In 2021, Zillow, an online real estate marketplace, was forced to cut 25% of its workforce because of an error in the AI algorithm used to predict home prices. The predictive model was used to enable the company to make cash offers on properties based on current versus predicted future values.

Zillow said the error in the predictive model had led to unintentionally purchasing homes at higher prices than the predicted estimates of future selling prices, resulting in a $304 million inventory write-down in Q3 2021.

With recent advances in generative and predictive AI becoming evident and the technology becoming more accessible to non-technical users, interest in AI modeling has expanded beyond just Fortune 500 companies to businesses of all sizes. However, as the barriers to entry lower, potential pitfalls arise for those looking to adopt predictive AI. So, what are some of the necessary steps and considerations required to develop AI-powered forecasting?

Ensuring the accuracy and trustworthiness of AI-generated data and predictions is crucial, especially given the growing reliance on artificial intelligence. One of the foundational steps in this process is ensuring data quality from the outset. High-quality, clean, and relevant data is essential for training AI models. This is why auditing and reviewing your existing systems and data prior to building a model is crucial.

A common problem in every organization with data is the existence of duplicates and incomplete entries. Oftentimes, data also exists, irreconciled, in multiple systems. These simple errors can badly skew a predictive model and can rapidly make a forecast useless.

Implementing transparent AI systems allows stakeholders to understand and trace the decision-making process and is vital for building trust and ensuring that the AI’s outputs can be scrutinized and validated. The workings of Generative AI models in particular are notorious for their murkiness, and knowledge of the system prompts that direct the model is necessary to ensure that the predictive outputs make sense.

Regular validation and testing of AI models against known benchmarks and real-world scenarios is also crucial to ensure their accuracy and this process will help identify any discrepancies and the necessary adjustments required to improve the model’s performance.

Identifying and mitigating biases in data and algorithms is essential to prevent skewed results that could lead to unfair or inaccurate outcomes. Whilst the heavy lifting can be done by AI when analyzing large historical datasets and synthetic data, human experts are still relied on to review and interpret AI outputs, providing a critical check, especially in applications where accuracy is paramount.

In 2019, a predictive AI model used by a major healthcare system was found to be biased in its predictions of patient health risks. The AI model was designed to identify patients who would benefit from extra medical care, but it systematically underestimated the health risks of patients with lower socioeconomic status.

This bias occurred because the model used healthcare costs as a proxy for health needs. Patients from wealthier backgrounds typically incurred higher healthcare costs, leading the AI to prioritize them over less affluent patients.

Consequently, patients from lower socioeconomic backgrounds, who might have had significant health needs but lower healthcare expenditures, were less likely to be flagged for additional care. This example highlights the importance of carefully selecting and validating the variables used in predictive AI models to avoid unintended biases.