TS Feature Engineering in AI & Artificial Intelligence

The most basic step is extracting Calendar Features from your datetime index. Features like Day of Week, Is Weekend, or Is Holiday are powerful predictors for many industries. For example, electricity demand is significantly lower on weekends, and retail sales spike on holidays. By explicitly giving these features to your model, you allow it to 'learn' these recurring patterns without needing complex time-series logic.

To help a model understand the past, we use Lag Features. A 'Lag 1' feature for today is simply 'Yesterday's Price.' This provides the model with the most recent context. We also use Rolling Window Features, such as a 7-day average or 30-day standard deviation. These features summarize the recent history and help the model distinguish between a temporary spike and a long-term trend change.

Time is circular, but numbers are linear. To a standard model, Hour 0 and Hour 23 are as far apart as possible (0 vs 23). However, in reality, they are only one hour apart. Cyclical Encoding solves this by mapping time components to Sine and Cosine coordinates on a circle. This preserves the 'closeness' of December and January, or midnight and 1:00 AM, leading to significantly better model performance on periodic data.

Raw dates are hard for models to understand. To build a great forecast, you must turn '2024-05-15' into meaningful features like 'Is it a weekend?' or 'Is it a holiday?'.

We extract calendar features to capture seasonality. 'Month', 'DayOfWeek', and 'Hour' are the first pieces of the puzzle.

We also create 'Lag Features'. By shifting the data, we teach the model that 'today's value' depends on 'yesterday's value'.

Checkpoint: What is a 'Lag Feature'?

• →A completely new data point
• →A value from a previous time step used as an input feature

For cyclical features like Hour or Month, we use Sine and Cosine transforms. This tells the model that Hour 23 is actually very close to Hour 0.

Great features are the secret sauce of forecasting. They turn a simple model into an intelligent predictor of real-world behavior.

Checkpoint: Why do we use Sine/Cosine transforms for the 'Month' feature?

• →To make it faster
• →To show the model that December (12) is close to January (1)

Feature engineering mastered! You've learned to speak the language of time. Ready to measure your model's success with error metrics?

Build Real Lag Features. Finish extracting lagged values from a series — the most common time-series feature.