Time Series Feature Engineering Techniques in Python

In business analytics, time is a core component for analyzing sales, revenue, profit, and growth, and for making forecasts; handling time‑sensitive data requires careful feature engineering.

The article begins with a brief introduction to time‑series concepts, emphasizing that observations are captured at regular intervals and each point depends on its predecessors.

Using a sample dataset named Train_SU63ISt.csv that contains a datetime column and a count column, the data is loaded and the datetime column is converted to a proper datetime type:

<code>import pandas as pd
 data = pd.read_csv('Train_SU63ISt.csv')
 data['Datetime'] = pd.to_datetime(data['Datetime'], format='%d-%m-%Y %H:%M')
 data.dtypes</code>

Several date‑related features are then derived, such as year, month, day, day of week (numeric and name):

<code>data['year'] = data['Datetime'].dt.year
 data['month'] = data['Datetime'].dt.month
 data['day'] = data['Datetime'].dt.day
 data['dayofweek_num'] = data['Datetime'].dt.dayofweek
 data['dayofweek_name'] = data['Datetime'].dt.weekday_name
 data.head()</code>

Time‑related features like hour and minute are extracted similarly:

<code>data['Hour'] = data['Datetime'].dt.hour
 data['minute'] = data['Datetime'].dt.minute
 data.head()</code>

Lag features are created by shifting the target variable, allowing the model to use past values as predictors:

<code>data['lag_1'] = data['Count'].shift(1)
 data['lag_2'] = data['Count'].shift(2)
 data['lag_3'] = data['Count'].shift(3)
 data['lag_4'] = data['Count'].shift(4)
 data['lag_5'] = data['Count'].shift(5)
 data['lag_6'] = data['Count'].shift(6)
 data['lag_7'] = data['Count'].shift(7)
 data = data[['Datetime', 'lag_1', 'lag_2', 'lag_3', 'lag_4', 'lag_5', 'lag_6', 'lag_7', 'Count']]
 data.head(10)</code>

Rolling‑window statistics (e.g., a 7‑day moving average) are computed to capture recent trends:

<code>data['rolling_mean'] = data['Count'].rolling(window=7).mean()
 data = data[['Datetime', 'rolling_mean', 'Count']]
 data.head(10)</code>

Expanding‑window features, which consider all data up to the current point, are generated with the expanding() method:

<code>data['expanding_mean'] = data['Count'].expanding(2).mean()
 data = data[['Datetime', 'Count', 'expanding_mean']]
 data.head(10)</code>

The article also discusses domain‑specific features, such as incorporating holidays or weather information, and stresses the importance of understanding the problem context to design effective features.

For model validation, a time‑ordered split is recommended: the last three months of the 25‑month dataset are held out as a validation set, while the earlier data is used for training, ensuring that future information is never leaked into the model.

Finally, the article concludes that mastering these feature‑engineering techniques transforms a time‑series forecasting problem into a supervised learning task, enabling the use of regression models like linear regression or random forests with greater confidence.