Feature Engineering

🧭 Purpose

This page explains how Sake prepares raw data for machine learning through cleaning, transformation, encoding, scaling, and dimensionality reduction.

Feature engineering converts a source dataset into a model-ready feature matrix while preserving analytical meaning.

🧱 Feature Engineering Workflow

Step	Purpose
Column selection	Identify predictors, target, identifiers, and metadata.
Numeric coercion	Convert financial values to numeric form.
Missing-value handling	Impute, flag, or remove incomplete observations.
Categorical encoding	Convert categories to numeric features.
Scaling	Standardize numeric magnitude when models require it.
Dimensionality reduction	Reduce feature width while preserving useful structure.
Feature validation	Confirm outputs remain aligned to source records.

🎯 Separate Features and Target

target_column = "target"

df_features = df_accounts.drop(columns=[target_column])
target = df_accounts[target_column]

Preserve identifiers separately when they should not be used as predictors:

identifier_columns = ["treasury_account_symbol", "account_name"]

existing_identifiers = [
    column for column in identifier_columns
    if column in df_features.columns
]

df_identifiers = df_features[existing_identifiers].copy()
df_features = df_features.drop(columns=existing_identifiers)

🔢 Numeric Feature Preparation

Select numeric columns:

numeric_features = df_features.select_dtypes(include="number").columns.tolist()

Review numeric missing values:

df_features[numeric_features].isna().sum()

🧹 Missing-Value Handling

Simple numeric imputation:

from sklearn.impute import SimpleImputer
import pandas as pd

imputer = SimpleImputer(strategy="median")

df_numeric = pd.DataFrame(
    imputer.fit_transform(df_features[numeric_features]),
    columns=numeric_features,
    index=df_features.index
)

Common strategies:

Strategy	Use
Mean	Symmetric numeric distributions.
Median	Skewed financial data.
Most frequent	Categorical fields.
Constant	Explicit missing category or zero-fill when justified.

🧾 Categorical Encoding

Select categorical columns:

categorical_features = df_features.select_dtypes(include=["object", "category"]).columns.tolist()

One-hot encode categorical columns:

df_categorical = pd.get_dummies(
    df_features[categorical_features],
    drop_first=False,
    dummy_na=True
)

⚖️ Scaling

Some algorithms benefit from scaling, especially distance-based and regularized models.

from sklearn.preprocessing import StandardScaler
import pandas as pd

scaler = StandardScaler()

df_scaled = pd.DataFrame(
    scaler.fit_transform(df_numeric),
    columns=df_numeric.columns,
    index=df_numeric.index
)

Use scaling for:

Model Family	Scaling Importance
Logistic Regression	Often beneficial.
Support Vector Machine	Usually important.
K-Nearest Neighbors	Important.
Linear models with regularization	Important.
Tree-based models	Usually not required.

🧮 Combine Feature Blocks

df_model_features = pd.concat(
    [df_scaled, df_categorical],
    axis=1
)

📉 PCA Example

Principal Component Analysis can reduce dimensionality while preserving variance.

from sklearn.decomposition import PCA
import pandas as pd

pca = PCA(n_components=0.95, random_state=42)

pca_values = pca.fit_transform(df_scaled)

df_pca = pd.DataFrame(
    pca_values,
    columns=[f"pc_{index + 1}" for index in range(pca_values.shape[1])],
    index=df_scaled.index
)

📉 Truncated SVD Example

Truncated SVD is useful for sparse or high-dimensional encoded data.

from sklearn.decomposition import TruncatedSVD
import pandas as pd

svd = TruncatedSVD(n_components=10, random_state=42)

svd_values = svd.fit_transform(df_model_features)

df_svd = pd.DataFrame(
    svd_values,
    columns=[f"svd_{index + 1}" for index in range(svd_values.shape[1])],
    index=df_model_features.index
)

🏛️ Budget Execution Feature Ideas

Feature	Purpose
Obligation rate	Obligations divided by budgetary resources.
Outlay rate	Outlays divided by obligations or resources.
Unobligated balance ratio	Balance divided by resources.
Recovery ratio	Recoveries divided by obligations.
Account category flags	Encoded account type or availability.
Fiscal period indicators	Time or reporting-period context.
Log-transformed values	Reduce skew in high-dollar financial fields.

✅ Feature Engineering Checklist

Before modeling:

• Target column is separated.
• Identifiers are preserved but excluded from predictors unless justified.
• Numeric fields are converted.
• Missing values are handled.
• Categorical fields are encoded.
• Scaling is applied where appropriate.
• Feature matrix rows align to the target.
• Derived budget ratios are reviewed for divide-by-zero issues.
• Feature names remain interpretable.