Datetime¶

• https://medium.com/data-science-at-microsoft/introduction-to-feature-engineering-for-time-series-forecasting-620aa55fcab0

• https://medium.com/@karanbhutani477/feature-engineering-for-time-series-data-a-deep-yet-intuitive-guide-b544aeb26ec2

features:

• month/day/hour

• weekday/weekend/holiday

• lag (shift): use previous data

• windows (rolling average): mean for the last day/week etc.

• expanding

timeseries feature engineering¶

• https://www.kaggle.com/code/pinardogan/time-series-using-lightgbm-with-explanations/notebook

• https://www.kaggle.com/code/zhikchen/forecasting-prediction-using-lightgbm-model

note:

• lag

• rolling

• ewma

• lasso feature selection

• customized evaluation function feval

Feature types: | Feature Type | Description | Effect on Accuracy | |--------------------|------------------------------------------|-------------------------------| | Lag Features | Values from previous time steps | High for short-term trends | | Rolling Statistics | Moving averages or standard deviations | Medium for identifying trends | | Time-based | Cyclical encodings (e.g., sine/cosine for time) | High for seasonal patterns |

trend decomposition¶

from statsmodels.tsa.seasonal import seasonal_decompose
decompose_result = seasonal_decompose(decom_data, period=5, model='additive')

observed = decompose_result.observed
trend    = decompose_result.trend
seasonal = decompose_result.seasonal
residual = decompose_result.resid

datetime feature¶

Usually the day of year, day of week and hour will cover all the season patterns. Each will encoded into the sine and cosine features.

sin/cos features¶

# sin(2 * pi * x / max(x))
# cos(2 * pi * x / max(x))

# hour
df['h_sin'] = np.sin(np.pi * df['hour'] / 12)
df['h_cos'] = np.cos(np.pi * df['hour'] / 12)
# month
df['m_sin'] = np.sin(np.pi * df['month'] / 6)
df['m_cos'] = np.cos(np.pi * df['month'] / 6)
# day of year
df['y_sin'] = np.sin(np.pi * df['day_of_year'] / 183)
df['y_cos'] = np.cos(np.pi * df['day_of_year'] / 183)

radial basis function¶

sin/cos vs radial basis function (rbf): https://developer.nvidia.com/blog/three-approaches-to-encoding-time-information-as-features-for-ml-models/

diff:

• sin/cos: cyclical or periodic data, such as time series data

• rbf: non-linear relationships and capturing complex patterns in data, such as images

• rbf can introduce more complexity to the model, which may lead to overfitting if not carefully tuned

• rbf can be computationally more expensive than sine/cosine features, especially for large datasets

lag and lag diff/auto correlation¶

lag features: https://www.kaggle.com/code/ryanholbrook/time-series-as-features

df['x_lag1'] = df.groupby(['id'])['x'].shift(1)
df['x_lag1_diff'] = df['x'] - df['x'].groupby(['id']).shift(1)
df['x_lag1_autocorr'] = df['x'].corrwith(df['x'].shift(1))

percentage diff¶

df['pct_diff_x_y'] = (df['x'] - df['y']) / df['y'] * 100

diff first/median¶

df['x_diff_first'] = df['x'] - df['x'].groupby(['id']).transform('first') #median

imbalance¶

df['imb_x_y'] = (df['x'] - df['y']).divide(df['x'] + df['y'], fill_value=np.nan)

rolling¶

mean/var/std/min/max/med/skew/kurt/cumsum/range

df['x_mean'] = df.groupby(['id'])['x'].agg('mean') # std etc
df['x_cumsum'] = df.groupby(['id'])['x'].cumsum()
df['x_range'] = df.groupby(['id'])['x'].max(axis=1) - df.groupby(['id'])['x'].min(axis=1)