Reddit's 'pulse' on US inflation: forecasting with large language models

We show that large language models (LLMs) can transform Reddit discussions into timely predictors of US inflation. Using inflation-related submissions and time-local comments from major economics-focused subreddits, we construct monthly narrative indicators that capture perceived price dynamics. Signals are generated by fine-tuning pre-trained models (BERT-, Qwen-, LLaMA-, and Gemma-type architectures) for labels produced by human annotators and ChatGPT and benchmarked against a non-fine-tuned LLaMA-70B model. Forecasting and nowcasting are implemented in pseudo-real time with strictly backward-looking transformations, recursive expanding windows, and explicit data-availability constraints. In a recursive pseudo out-of-sample evaluation with horizons up to 18 months, Reddit-LLM models and MSE-weighted forecast combinations improve point and density forecasts of headline CPI and core PCE relative to standard benchmarks, including autoregressive models augmented with Michigan survey expectations and inflation swaps. In real-time nowcasting, Reddit signals constructed using information available early in the month improve nowcasts and perform competitively with the Cleveland Fed Inflation Nowcast. Importantly, much of the predictive content can be captured with fine-tuned small language models (SLMs), which often deliver performances close to those of much larger LLMs at a fraction of the computational cost, supporting scalable and resource-efficient deployment.