Core Concepts
- Lag Features: Use past values as predictors for future points.
- Rolling Windows / Aggregations: Compute moving averages, sums, or other statistics over fixed intervals.
- Time-Based Features: Include hour of day, day of week, or calendar effects; can be per SIM, per cell tower, etc.
Data Quality
- Handle missing timestamps and irregular events appropriately.
- Impute missing values or resample data to regular intervals if needed.
Statistical Properties
- Stationarity: Check if mean, variance, and autocorrelation are constant over time. Use tests like ADF.
- Trend and Seasonality: Identify and potentially remove to stabilize the series.
- ACF / PACF Analysis: Understand autocorrelation structure; helps determine relevant lag features.
Practical Feature Engineering
- Combine lags, rolling statistics, and time-based features to create predictive input matrices.
- Adjust features for irregular timestamps and missing data.
Common Interview Questions
- “How do you check if a time series is stationary?”
- You typically examine whether its statistical properties (mean, variance, autocorrelation) remain constant over time. You can visualise it, get summary stats, do ADF tests.