Add N-HiTS-comparison traffic example under gists/timeseries/nhits

[OVVO-Financial]

Summary

Adds gists/timeseries/nhits/ — a third entry in the examples gallery, with a genuine head-to-head against the N-HiTS deep-learning forecaster (AAAI 2023).

Marco Peixeiro's All About N-HiTS forecasts the hourly Interstate 94 traffic series — same dataset, same [:-120]/[-120:] split, same MAE metric this example uses. NNS.ARMA (seasonal periods from nns_seas, MAE objective, no neural network or training loop) lands MAE 236, beating the article's results:

Model	MAE	Beats naive baseline?
NNS.ARMA.optim	236	✅
Naive seasonal (weekly, K=168)	249	—
N-HiTS	266	❌
N-BEATS	worse than N-HiTS	❌

The article's own conclusion is that neither N-HiTS nor N-BEATS beats the naive seasonal baseline on this series; NNS is the only method that does.

Contents

• nhits_traffic.py — runnable script (writes the figure for the README).
• README.md — gallery-style framing, the head-to-head table sourced to the article, the forecast plot, and a "how it works" walkthrough.
• nhits_traffic_forecast.png — generated figure (actual vs. forecast).

Notes

• Reproduced the MAE (236.16791666666677) exactly when generating the figure.
• The README fairly notes the article's caveat that this is a small, repetitive sample (part of why the baseline is strong).
• gists/ is excluded from ruff; pure examples/docs, no package or test changes.

🤖 Generated with Claude Code