Original Summary

I built a pipeline that pulls NHTSA&#x27;s consumer complaints and official recall file every day and scores how unusual each automaker&#x27;s complaints are, per component, against its own past year.<p>A few things I learned along the way: - A plain z-score flags 4–10x too many weeks on small, noisy counts. A negative-binomial tail probability fixes that. - NHTSA&#x27;s per-vehicle recall endpoint silently misses about 8% of recalls because of model-name mismatches, so I use the official bulk file instead. - Recalls cause complaints, so a naive backtest flatters itself. Excluding weeks right after a recall, spikes are followed by a recall of the same component about 1.5x as often as a normal week. Useful, but far from a crystal ball.<p>There&#x27;s a free sample on the site and a free tier on the API. Happy to answer questions about the data or the method.

  • 情报分类：商业与市场研究
  • 分类依据：内容涉及商业、投资或市场动态
  • 信息来源：Hacker News 新项目
  • 发布时间：2026/10/14 11:10:04