Feature Engineering for Time Series

Duration: 5 min

This module delves into the crucial process of feature engineering for time series data. We will explore various techniques to transform raw time series data into meaningful features that can significantly enhance the performance of forecasting models. Understanding and implementing effective feature engineering is vital for accurate and reliable time series forecasting.

Lag Features

Lag features are created by shifting the time series data back by a certain number of time steps. These features help capture the temporal dependencies and trends in the data, allowing models to better understand the underlying patterns. Lag features are essential for models like ARIMA and LSTM, which rely on past values to make predictions.

import pandas as pd

# Sample time series data
data = {'value': [10, 15, 20, 25, 30, 35, 40, 45, 50, 55]}
df = pd.DataFrame(data)

# Creating lag features
df['lag_1'] = df['value'].shift(1)
df['lag_2'] = df['value'].shift(2)

print(df)

Try it in Google Colab:

   value  lag_1  lag_2
0     10    NaN    NaN
1     15   10.0    NaN
2     20   15.0   10.0
3     25   20.0   15.0
4     30   25.0   20.0
5     35   30.0   25.0
6     40   35.0   30.0
7     45   40.0   35.0
8     50   45.0   40.0
9     55   50.0   45.0

Rolling Window Statistics

Rolling window statistics involve calculating summary statistics (such as mean, sum, max, min) over a sliding window of a specified size. These features can help capture trends and seasonality in the data, providing additional context for the forecasting models. Rolling window features are particularly useful for highlighting short-term patterns and anomalies.

import pandas as pd

# Sample time series data
data = {'value': [10, 15, 20, 25, 30, 35, 40, 45, 50, 55]}
df = pd.DataFrame(data)

# Creating rolling window features
df['rolling_mean'] = df['value'].rolling(window=3).mean()
df['rolling_sum'] = df['value'].rolling(window=3).sum()

print(df)

💡 Tip: When creating lag and rolling window features, ensure that the time series data is stationary or apply differencing to make it stationary. Non-stationary data can lead to misleading features and poor model performance.

Key Concepts

Check Your Understanding