Weather Trend Forecasting

Context

This was the PM Accelerator advanced data scientist technical assessment: forecast weather trends on the Global Weather Repository Kaggle dataset, with cleaning, EDA, multiple models, ensemble, anomalies, feature importance, environmental and geographical analysis, plus a short demo video and public repo.

Problem

The dataset is a panel of scrape snapshots—268 cities, 41 columns, 167k+ rows—not a single tidy time series. The brief requires using last_updated as the time index and delivering advanced analyses beyond a single chart.

Approach

1. Notebook 01 — Parse datetimes, drop one duplicate (location_name, last_updated), global IQR/QA documentation, Kyiv EDA (temperature + precipitation at observation time).
2. Notebook 02 — Kyiv temperature_celsius, 80/20 chronological split, naive / seasonal-365 / Ridge lags [1,2,3,7,14,28], ensemble of Ridge + naive; report lag-span diagnostic for seasonal naive.
3. Notebook 03 — Robust-z and Isolation Forest anomalies, Ridge lag coefficients, Kyiv weather–air-quality correlations, country mean temperatures with sparse-country filter.

Results (holdout)

Model	MAE (°C)	RMSE (°C)
Naive	2.509	3.753
Seasonal naive (365 rows)	5.993	7.528
Ridge	2.474	3.599
Ensemble (Ridge + naive)	2.474	3.653

Ridge is the best single model on RMSE; the ensemble ties MAE. The interesting negative result is seasonal naive—documented with observation-hour drift and weather variability, not a false calendar-alignment story.

Artifacts

• GitHub repository — notebooks, requirements.txt, report PDF, slides, figures
• Written report — methodology, caveats, limitations
• Demo video — 2-minute slide walkthrough on YouTube

What this demonstrates

End-to-end data science communication: honest baselines, time-respecting evaluation, advanced methods with interpreted failures, and reproducible artifacts suitable for reviewer scrutiny—not a notebook that only reports the best number.