The data showed warning indicators in the months prior to the 2008 financial crisis. Subprime mortgage delinquency rates had been increasing. The markets for credit default swaps were under strain. In important markets, the expansion of housing prices had stopped. In the forensic accounting that follows a collapse, economists discovered the majority of these signals after the fact. This led to the inquiry, “Why didn’t the models catch this in advance?” Since then, it has been spearheading a subtle but important change in the way governments anticipate the economy.
The majority of nations have decided to use machine learning and AI-assisted analysis on datasets that are bigger and more complicated than what conventional econometric models were designed to handle. These days, early warning systems run by central banks and finance ministries routinely monitor for the nonlinear patterns that precede instability by ingesting macro-financial data, such as bond spreads, credit flows, employment statistics, and trade volumes. This strategy is appealing because of its quickness. What has happened in a market over the last 48 hours cannot be seen by a traditional model that is updated quarterly using data from the national statistics agency. Real-time feed processing by AI systems is possible.
Because it pulls from sources that formal economics had mostly disregarded, the sentiment analysis portion is very intriguing. Although sluggish, consumer surveys are helpful. Search trend data, hiring announcements, and social media posts all travel more quickly and can forecast changes in consumer behavior before official statistics do. The Federal Reserve and a number of European governments have been experimenting with these inputs as leading indicators; they are not meant to replace traditional data, but rather to provide an extra signal in situations where the standard data stream is too sluggish to identify rapidly changing stress.
The use of policy becomes most immediately impactful when it comes to targeted stimulus. One of the recurring complaints about government spending during recessions is that large stimulus packages benefit industries that don’t require them while falling short of providing sufficient support for those who do. Theoretically, machine learning models trained on sector vulnerability indices, credit histories, and business survival data can determine which industries and small firms are most likely to collapse during a downturn—before the failure occurs—and provide help appropriately. Several nations have started experimenting with this strategy, with differing outcomes. It makes sense. The timeliness and quality of the underlying data have a significant impact on the execution.
Among the organizations pointing out the discrepancies between the promise and the present situation is the OECD. There is a real issue with algorithmic bias: models that were trained on past economic data mirror past trends, which might not adequately characterize new crisis kinds. For example, models trained mostly on the 2008 credit-driven crisis were not adequately calibrated for the 2020 pandemic, which was a demand shock with a structure very different from the latter. What the models can reliably output is further limited by data quality issues, such as incomplete records, conflicting definitions among national statistics agencies, and private data that governments can access in certain jurisdictions but not in others.
Another issue that isn’t emphasized as often as it ought to be is accountability. Who is in charge of the subsequent choice when an AI-assisted model instructs a finance ministry to get ready for a contraction that doesn’t occur or fails to identify one that does? In ways that institutions are still figuring out, the transition from human expert judgment to algorithmic recommendation shifts accountability. There is currently a lack of uniformity across the governments using these models in terms of transparency on their operation, the inputs they employ, and the locations of their confidence intervals.
None of this implies that the method is flawed; the speed and pattern-recognition improvements over conventional models are genuine and significant. However, there is a feeling that the gap between what the technology can accomplish in a controlled test environment and what it consistently produces in the chaotic, fast-paced context of an actual economic crisis is still greater than the enthusiasm surrounding it occasionally recognizes.
