KM Kanav Mehta Quantitative AI Research

AI / NLP · Macro

FOMC sentiment, from research repo to product.

The FinBERT-FOMC research line, productized into an installable package and CLI: transcript extraction, LLM-assisted labeling, fine-tuning, embeddings, novelty scores, and market regressions, wrapped in an interactive dashboard.

Back to Projects Toolkit ↗ Research ↗
FOMC meeting explorer in dark mode with the same panels FOMC meeting explorer in light mode with paragraph labels and confidence scores
The meeting explorer, running locally; the capture follows the site theme.

Pipeline

Every step, one CLI.

  • Extraction, training-data builds, LLM-assisted labeling, fine-tuning, inference, embeddings, and factor construction, each a subcommand with checkpoints, orchestrated by run-all.
  • Hawk/dove lexicon baseline and per-speaker profiles anchor the model-based scores.
  • Factor series regress against SPX returns under walk-forward evaluation, with no in-sample shortcuts.

Dashboard

Read the meetings, not just the scores.

  • Overview: sentiment, labor, inflation, and novelty time series across meetings.
  • Meeting explorer: every paragraph with its labels and confidence; hawkish red, dovish green.
  • Novelty explorer: per-meeting and per-paragraph novelty against the previous meeting, from embeddings.
  • Speakers: hawk/dove profiles and per-meeting trends.