AI is ‘absolutely useless’ at forecasting inflation. This proven model is 12 times more accurate.

A new study finds ChatGPT performs poorly at forecasting inflation, with Joachim Klement calling it "absolutely useless" for macroeconomic forecasting. The article says a low-tech Cleveland Fed model is about 12 times more accurate than generative AI. The piece reinforces skepticism around AI’s near-term usefulness for economic prediction rather than suggesting a direct market catalyst.

Analysis

The actionable takeaway is not that AI is weak in macro, but that investors are overpaying for narrative-driven forecasting tools when the edge still sits in simple, domain-specific statistical models. That tends to punish companies trying to monetize generic “AI for finance” claims while favoring vendors whose value proposition is workflow automation, data cleaning, or distribution rather than prediction. In practice, the market should start distinguishing between copilots that summarize and systems that actually forecast; the former can still earn adoption, the latter face a higher credibility bar.

For inflation-sensitive assets, this is a reminder that the next big move will likely come from regime identification, not headline sentiment. If the market leans harder on human or model-based inflation signals that are systematically better than LLM output, breakevens and rate-volatility may become less whipsawed by viral AI takes and more anchored to labor, shelter, and supply metrics. That reduces the odds of a sustained “AI says disinflation” trade and increases the value of positioning around data-release convexity over the next 1-3 months.

The second-order effect is reputational: once investors internalize that general-purpose AI underperforms on noisy macro series, adoption budgets may shift away from broad inference products toward narrower vertical applications with measurable ROI. That is a headwind for the most hyped AI software names and a relative tailwind for incumbent data providers and fintech infrastructure that can package proprietary datasets into repeatable signals. The contrarian angle is that the disappointment itself may be overdone for public AI equities; the market could already be discounting model failure, so the real short may be the subset of names priced for autonomous decision-making rather than the entire AI complex.

Key Decisions for Investors

• Short the most narrative-heavy AI software names on any post-earnings strength over the next 2-6 weeks; prefer companies marketing predictive intelligence over workflow software, as the multiple compression risk is highest if buyers start demanding proof of forecast alpha.
• Pair trade: long quality data/market-infrastructure providers (e.g., MSCI, CME) vs short speculative AI application names; thesis is that monetization shifts toward trusted inputs and execution rails, with better downside protection if AI forecasting skepticism persists for 3-6 months.
• In rates, use dips to buy short-dated receiver convexity or call spreads on front-end yields ahead of CPI/PCE releases over the next 1-2 months; if the market stops overweighting AI-generated macro narratives, real data should drive sharper repricing and volatility should rise.
• Avoid paying up for firms pitching ‘AI inflation forecasting’ products until there is auditability and live out-of-sample proof; the risk/reward is poor because a single regime shift can invalidate the product thesis and compress ARR multiples quickly.