Dimensionality reduction
Time Series with PCA and SVD
3
components retain about 99.9% of variance in global temperature data
Two very different datasets show how much shared structure a panel of time series really has.
What I did
- Built country-by-year and station-by-year matrices from Berkeley Earth temperatures (about 234 countries, 1901 to 2012) and Canadian ozone readings (18 stations, 1995 to 2022).
- Implemented PCA from the covariance matrix and SVD variants, then plotted scree curves and RMSE against number of components.
- Global temperature is dominated by a few components; ozone needs many more, meaning weaker shared structure.
Stack
NumPyscikit-learnMatplotlib