Bank Term Deposit Subscription Prediction

A machine learning project to predict whether a client will subscribe to a term deposit based on data from a Portuguese bank's direct marketing campaign.

Webapp Deployment

Click to use the webapp

Dataset

Source: Direct marketing campaign of a Portuguese banking institution.
Size: 45,211 entries, 17 original attributes.
Target variable: y (binary: “yes” if subscribed, “no” otherwise).

Original attributes

Attribute	Type	Description
age	numeric	Client age
job	categorical	Type of job (admin., blue‑collar, entrepreneur, etc.)
marital	categorical	Marital status (married, single, divorced)
education	categorical	Education level (primary, secondary, tertiary, unknown)
default	categorical	Has credit in default? (yes, no)
balance	numeric	Average yearly balance (euros)
housing	categorical	Has housing loan? (yes, no)
loan	categorical	Has personal loan? (yes, no)
contact	categorical	Communication type (unknown, telephone, cellular)
day	numeric	Last contact day of month (1–31)
month	categorical	Last contact month (jan, feb, …, dec)
duration	numeric	Last contact duration (seconds)
campaign	numeric	Number of contacts during this campaign
pdays	numeric	Days since last contact from a previous campaign (-1 if never contacted)
previous	numeric	Number of contacts before this campaign
poutcome	categorical	Outcome of previous campaign (unknown, other, failure, success)
y	binary	Subscribed term deposit? (yes, no)

Cleaning

• No completely empty columns were present.
• “unknown” values were converted to proper nulls.
• 3 rows with duration equal to 0 were removed.
• No duplicate rows were found.
• Class imbalance: only 5,289 “yes” subscriptions (≈11.6%).
• A flag prev_contacted (0 = never contacted, 1 = contacted before) was derived.
• Missing values in job and education were dropped; missing values in contact and poutcome were replaced with the string “missing”.
• Categorical columns were converted to category dtype.

Exploratory Data Analysis (EDA)

Categorical features (subscription rate “yes” highlights)

• job: retired and student show highest subscription rates.
• marital: singles subscribe most.
• education: tertiary education yields the highest yes rate.
• default: clients without default subscribe more.
• housing: those without a housing loan subscribe more.
• loan: those without a personal loan subscribe more.
• contact: missing contact type has the lowest yes rate.
• month: December and March have the highest subscription rates.
• poutcome: previous success strongly predicts subscription.

Numerical features (comparing yes vs. no)

• age: higher maximum and standard deviation for yes; median similar.
• balance: larger IQR, higher median, higher maximum for yes.
• day: distribution nearly identical across target classes.
• duration: much larger IQR, higher median, higher maximum for yes.
• campaign: distributions similar.
• pdays: larger values and IQR for yes (clients previously contacted longer ago).
• previous: larger values and IQR for yes.
• prev_contacted: larger values, higher median, higher max for yes.

Correlation heatmap

• duration had the strongest linear correlation with the target (+0.4).
• Increase in duration tended to increase subscription likelihood regardless of age.
• prev_contacted correlated very highly with pdays (+0.87).

Feature Engineering & Preprocessing

All categorical features were transformed into binary indicators based on the EDA insights. The original categorical columns were dropped after encoding.

1. Job: flag for student or retired → job_student_or_retired_feat.
2. Marital: flag for single → marital_single_feat (later dropped, as it did not pass final selection).
3. Education: flag for tertiary → education_tertiary_feat (dropped later).
4. Contact: flag for cellular or telephone → contact_cellular_telephone_feat.
5. Month: flag for months dec, oct, sep, mar → month_dec_oct_sep_mar_feat.
6. Poutcome: flag for success → poutcome_success_feat.
7. Housing, Loan, Default: binary encoding (yes=1, no=0) → housing_feat, loan_feat, default_feat (only housing_feat kept in final selection).

Numerical features were binned and then turned into binary indicators.

• Duration: binned into quartiles; flag for Q3/Q4 → duration_bin_Q3_Q4_feat (dropped later in favour of raw duration).
• Age: binned and inspected but not retained.
• Previous: flag for any previous contact (>0) → previous_any_feat (not retained).
• Day, Campaign: binned but distributions showed negligible effect, so they were dropped.
• Pdays: replaced -1 with 999 to create pdays_replaced_feat.

Log-transformed versions of balance and duration were created but not used in the final model.

Feature selection

Correlation with target (y) was computed for all engineered features. Features with absolute correlation > 0.1 were considered candidates. Among those, pairs with inter‑feature correlation > 0.7 were examined; the feature with lower target correlation in each pair was removed to reduce multicollinearity.

Final selected features:

• duration
• poutcome_success_feat
• month_dec_oct_sep_mar_feat
• pdays_replaced_feat
• contact_cellular_telephone_feat
• housing_feat
• job_student_or_retired_feat

The final dataset was saved as bank_final.csv.

Scaling

Continuous features (duration, pdays_replaced_feat) were scaled using RobustScaler. The scaler object was saved as scaler.pkl.

Modeling

Data split: 85% training, 15% test (stratified). A dummy classifier served as baseline.

Model	Accuracy	Class 1 Recall	ROC AUC
Dummy (baseline)	0.50	0.50	0.50
Logistic Regression	0.83	0.82	0.896
Decision Tree	0.72	0.91	0.851
Random Forest	0.81	0.86	0.906
XGBoost	0.79	0.88	0.895

Random Forest was selected for deployment due to its highest ROC AUC and strong recall on the minority class.

All trained models were saved for later use.

Deployment (Streamlit App)

A Streamlit application was built to serve the model with two modes:

Single Prediction

• User provides: poutcome, month, pdays, contact, housing, job, duration.
• Feature engineering steps are applied:
  • poutcome_success_feat = 1 if poutcome is “success” else 0.
  • month_dec_oct_sep_mar_feat = 1 if month is in [dec, oct, sep, mar] else 0.
  • pdays_replaced_feat = 999 if pdays == -1 else the actual value.
  • contact_cellular_telephone_feat = 1 if contact is cellular or telephone else 0.
  • housing_feat = 1 if housing is “yes” else 0.
  • job_student_or_retired_feat = 1 if job is student or retired else 0.
• The seven features are assembled into an array, then duration and pdays_replaced_feat are scaled using the loaded scaler.pkl.
• The Random Forest model predicts the subscription outcome, which is displayed to the user.

Batch Prediction

• User uploads a CSV file containing the required columns: poutcome, month, pdays, contact, housing, job, duration.
• The same feature engineering logic is applied to the entire dataset.
• Scaled features are fed to the model, and a new column y_predicted is added.
• The first 10 rows of the resulting dataframe are shown, and a download button provides the full dataset with predictions.

Requirements

• pandas
• numpy
• matplotlib
• seaborn
• scikit-learn
• xgboost
• lightgbm
• joblib
• streamlit