3. Feature Engineering¶
Business Feature¶
Business Features are categorical or numeric variables that reflect domain-specific business behaviors, enhancing the model's interpretability and forecasting accuracy. In this project, key Business Features include:
| Business Feature | Description | Purpose / Impact |
|---|---|---|
| Wholesale | Sales from wholesale customers often exhibit high volume but unstable patterns. These transactions may influence retail customer purchasing behavior. | Captures volatility from wholesale segment; accounts for its effect on more stable retail sales |
| Visits | Represents the number of customer visits per day or time unit. Reflects store performance and customer behavior over time. | Helps model demand fluctuation based on shopper volume; supports sales variation and performance analysis |
Time Feature¶
Time features are extracted from the date column (ds) to describe calendar-based cycles, seasonality and trends. These features enable the model to capture recurring temporal patterns, such as:
| Time Feature | Description |
|---|---|
| Day of the week | Monday, Tuesday, etc. |
| Month of the year | January, February, etc. |
| Week of the year | Week 1, Week 2, etc. |
| Holidays | Tet, Christmas, Black Friday, etc. |
| Fourier coefficients | Encoded periodic patterns in time series using sine and cosine transforms |
Lag Feature¶
Lag features are sales values shifted back by a specific time period, derived from past values of the target variable (sales) or related variables. In this project, lag features play a key role in enabling the model to learn autoregressive relationships — a crucial factor in time series forecasting. However, the value of a lag feature depends on the point in the forecasting horizon you are predicting. For example, sales on day D can be used as:
- 1-day lag when forecasting D+1,
- 2-day lag when forecasting D+2,
- 7-day lag when forecasting D+7,
- 14-day lag when forecasting D+14,
- 28-day lag when forecasting D+28.
Auto statistic Feature (TSFresh)¶
Time series features are automatically extracted using the TSFresh library, which generates a wide variety of statistical and signal-based features, including:
| Auto statistic Feature (TSFresh) | Description |
|---|---|
| Energy and Entropy | Measure signal strength and randomness of the time series. |
| Mean, Variance, Skewness, Kurtosis | Describe central tendency, dispersion, and shape of distribution. |
| Autocorrelation and Partial Autocorrelation | Measure relationships between current and past values. |