如何基于指定范围绘制31+状态的Matplotlib Vintage Curve？

Hey there! Let's fix the issues in your current code and walk through building a proper vintage curve step by step. Your current approach isn't handling the 31+ status correctly (since those are string values like "31+") and isn't organizing data into vintage cohorts—this is the key piece missing for this type of curve.

Step 1: Clean and Prepare Your Data #

First, we need to fix data types and filter for only the 31+ entries:

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns

# Load and rename columns for clarity
df = pd.read_csv('sample_data.csv')
df.columns = ['record_date', 'status']

# Parse dates correctly (matches your day/month/year format)
df['record_date'] = pd.to_datetime(df['record_date'], format='%d/%m/%Y')

# Convert status to numeric values (remove "+" signs first)
df['status_numeric'] = df['status'].str.replace('+', '').astype(int)

# Filter to keep only entries where status is 31+
df_31plus = df[df['status_numeric'] >= 31].copy()

Step 2: Define Vintage Cohorts #

Vintage curves track groups (cohorts) of accounts by their origination month/year. For this example, let's assume record_date is when the account hit 31+ delinquency, and we need to calculate how many months after origination this happened. If you have a separate origination date column, use that instead of generating dummy data below:

# Generate dummy origination dates (replace this with your actual origination date column!)
np.random.seed(42)
df_31plus['origination_date'] = df_31plus['record_date'] - pd.to_timedelta(np.random.randint(1, 12, size=len(df_31plus)), unit='M')

# Calculate months since origination
df_31plus['months_since_origination'] = (df_31plus['record_date'].dt.year - df_31plus['origination_date'].dt.year)*12 + (df_31plus['record_date'].dt.month - df_31plus['origination_date'].dt.month)

# Create vintage cohort labels (year-month of origination)
df_31plus['vintage'] = df_31plus['origination_date'].dt.to_period('M').astype(str)

Step 3: Aggregate Data for the Curve #

Next, we group data by vintage cohort and months since origination to count how many accounts hit 31+ status in each period:

# Aggregate counts of 31+ accounts per vintage and month since origination
vintage_summary = df_31plus.groupby(['vintage', 'months_since_origination']).size().reset_index(name='count_31plus')

# Optional: If you have total accounts per vintage, calculate percentage of delinquents
# total_per_vintage = df.groupby('vintage').size().reset_index(name='total_accounts')
# vintage_summary = vintage_summary.merge(total_per_vintage, on='vintage')
# vintage_summary['percent_31plus'] = (vintage_summary['count_31plus'] / vintage_summary['total_accounts']) * 100

Step 4: Plot the Vintage Curve #

Now we can visualize the curve—each line represents a different vintage cohort, showing how delinquency trends over time:

plt.figure(figsize=(12, 6))
sns.lineplot(
    data=vintage_summary,
    x='months_since_origination',
    y='count_31plus',  # Replace with 'percent_31plus' if you calculated percentages
    hue='vintage',
    marker='o'
)

plt.title('Vintage Curve: 31+ Delinquent Accounts by Origination Cohort')
plt.xlabel('Months Since Account Origination')
plt.ylabel('Number of 31+ Delinquent Accounts')
plt.legend(title='Vintage Cohort', bbox_to_anchor=(1.05, 1), loc='upper left')
plt.grid(True)
plt.show()

Key Fixes from Your Original Code #

• Status Handling: You tried comparing string values (like "31+") to integers—we converted status to numeric values first to fix this.
• Vintage Cohorts: You weren't grouping data by origination periods, which is essential for a vintage curve (it shows how different cohorts perform over time).
• Avoid Redundant Plots: Your loop was plotting multiple distplots unnecessarily—we aggregate data first, then plot once for clean results.

Notes for Your Specific Data #

• If your record_date is the origination date (not delinquency date), you'll need a separate column for when the account became 31+ to calculate months since origination.
• If each account has multiple status entries, first find the earliest date each account hit 31+ to avoid counting duplicates.

内容的提问来源于stack exchange，提问作者Raj