Cyclical

cyclical features¶

links:

• https://mlpills.substack.com/p/issue-89-encoding-cyclical-features

• https://skforecast.org/0.16.0/faq/cyclical-features-time-series.html

df['cyc_sncs'] = df['cyc_sin'] * df['cyc_cos']
# Add squared terms to capture non-linear relationships
df['cyc_sin2'] = df['cyc_sin'] ** 2
df['cyc_cos2'] = df['cyc_cos'] ** 2

encoding types:

• one-hot

• sin/cos

• spine

• rbf

pros and cons:

• nn: sin/cos is good

• tree-based models (Random Forest, XGBoost, LightGBM): split features one at a time - cannot do that for sin/cos together

• prophet: automatically models seasonality using Fourier transformations - no need to encode cyclical features