P2P Lending MLP Digital Brain

This repository contains a Multilayer Perceptron (MLP) model, along with Logistic Regression and Random Forest baselines, designed to predict the risk of loan default in a peer-to-peer (P2P) lending context. The models are trained on a subset of the Lending Club dataset.

Model Architecture (MLP Digital Brain)

The MLP model uses the following architecture:

Loan Features → Dense(64) → ReLU → Dense(32) → ReLU → Sigmoid → P(Default)

Features

The models use the following features to predict loan risk:

loan_amnt: The amount of the loan applied for.
annual_inc: The annual income of the borrower.
dti: Debt-to-income ratio.
int_rate: The interest rate on the loan.
grade: The loan grade assigned by Lending Club (categorical).

Target Variable

The target variable is loan_status mapped to risk:

0: Fully Paid
1: Charged Off

Performance Metrics

The models were evaluated on a test set (20% of the data) with the following metrics. Note that the MLP model uses a custom threshold (0.20) to maximize F1-score on the validation set, while baselines use a default threshold of 0.50.

Validation Metrics

{
  "Logistic Regression": {
    "ROC-AUC": 0.734351,
    "PR-AUC": 0.392071,
    "Accuracy": 0.678738,
    "Precision": 0.323171,
    "Recall": 0.666667,
    "F1": 0.435318
  },
  "Random Forest": {
    "ROC-AUC": 0.676701,
    "PR-AUC": 0.332663,
    "Accuracy": 0.767523,
    "Precision": 0.359155,
    "Recall": 0.320755,
    "F1": 0.33887
  },
  "MLP Digital Brain": {
    "ROC-AUC": 0.734171,
    "PR-AUC": 0.375425,
    "Accuracy": 0.815421,
    "Precision": 0.555556,
    "Recall": 0.031447,
    "F1": 0.059524
  }
}

Test Metrics

{
  "Logistic Regression": {
    "ROC-AUC": 0.733922,
    "PR-AUC": 0.372299,
    "Accuracy": 0.674766,
    "Precision": 0.32304,
    "Recall": 0.683417,
    "F1": 0.43871,
    "Threshold": 0.5
  },
  "Random Forest": {
    "ROC-AUC": 0.701048,
    "PR-AUC": 0.320653,
    "Accuracy": 0.754206,
    "Precision": 0.346154,
    "Recall": 0.361809,
    "F1": 0.353808,
    "Threshold": 0.5
  },
  "MLP Digital Brain": {
    "ROC-AUC": 0.739213,
    "PR-AUC": 0.397156,
    "Accuracy": 0.679439,
    "Precision": 0.327751,
    "Recall": 0.688442,
    "F1": 0.444084,
    "Threshold": 0.2
  }
}

Usage

The model can be used to predict the probability of a loan being charged off and to classify the loan as 'Fully Paid' or 'Charged Off' based on the features.

from huggingface_hub import HfApi
import joblib
import pandas as pd

repo_id = "Nestifytech/p2p-MLP-lending-risk-model"
api = HfApi()

# Download artifacts
api.hf_hub_download(repo_id=repo_id, filename="artifacts/p2p_lending_mlp_digital_brain.joblib", local_dir=".")
api.hf_hub_download(repo_id=repo_id, filename="artifacts/p2p_lending_mlp_metadata.json", local_dir=".")

# Load the model and metadata
loaded_mlp = joblib.load("artifacts/p2p_lending_mlp_digital_brain.joblib")
with open("artifacts/p2p_lending_mlp_metadata.json", "r") as f:
    metadata = json.load(f)

best_threshold = metadata["threshold"]

def digital_brain_predict(loan_amnt, annual_inc, dti, int_rate, grade):
    row = pd.DataFrame([{        "loan_amnt": loan_amnt, "annual_inc": annual_inc,        "dti": dti, "int_rate": int_rate, "grade": grade,    }])
    p_default = float(loaded_mlp.predict_proba(row)[0, 1])
    label = "Charged Off" if p_default >= best_threshold else "Fully Paid"
    return {         "label": label,         "p_charged_off": p_default,         "p_fully_paid": 1 - p_default,         "threshold": best_threshold    }

# Example prediction
prediction = digital_brain_predict(10000, 60000, 15.0, 12.0, "C")
print(prediction)

Demo

A Gradio-based interactive demo for this model is available, allowing users to input loan details and get real-time risk predictions.

Try the Gradio Demo

Environment

Python version: 3.13.15
Pandas version: 3.0.6
NumPy version: 2.5.3

License

MIT License

Downloads last month: -