Bankruptcy Predictions

Data Notice: This dataset provides academic research access with a 6-month data lag. For real-time data access, please visit sov.ai to subscribe. For market insights and additional subscription options, check out our newsletter at blog.sov.ai.

from datasets import load_dataset
df_bankrupt = load_dataset("sovai/bankruptcy", split="train").to_pandas().set_index(["ticker","date"])

Monthly corporate bankruptcy predictions arrive the 2nd of every month.

Tutorials are the best documentation — Corporate Bankruptcy Tutorial

Input Datasets	SEC Bankruptcies, Delistings, Market Data, Financial Statements
Models Used	CNN, LightGBM, RocketModel, AutoEncoder
Model Outputs	Calibrated Probabilities, Shapley Values

Description

The model predicts the likelihood of bankruptcies in the next 6-months for US publicly listed companies using advanced machine learning models.

With an accuracy of around 89% and ROC-AUC of 85%, these models represent a large improvement over traditional methods of bankruptcy prediction for equity selection.

Advanced modeling techniques used in this dataset:

• The Boosting Model: Utilizes LightGBM technology, integrating both fundamental and market data for accurate predictions.
• The Convolutional Model: Employs a Convolutional Neural Network (CNN) for efficient pattern recognition in market trends.
• The Rocket Model: Specializes in time series data, using random convolutional kernels for effective classification and forecasting.
• The Encoder Model: Combines LightGBM with CNN autoencoders, enhancing feature engineering for more precise predictions.
• The Fundamental Model: Focuses solely on fundamental data via LightGBM, without extra architectural layers, for straightforward financial analysis.

Data Access

Monthly Probabilities

Specific Tickers

import sovai as sov
df_bankrupt = sov.data('bankruptcy', tickers=["MSFT","TSLA","META"])

Specific Dates

import sovai as sov
df_bankrupt = sov.data('bankruptcy', start_date="2017-01-03", tickers=["MSFT"])

Latest Data

import sovai as sov
df_bankrupt = sov.data('bankruptcy')

All Data

import sovai as sov
df_bankrupt = sov.data('bankruptcy', full_history=True)

Daily Probabilities

import sovai as sov
df_bankrupt = sov.data('bankruptcy/daily', tickers=["MSFT","TSLA","META"])

The daily probabilities are experimental, and have a very short history of just a couple of months.

Feature Importance (Shapleys)

import sovai as sov
df_importance = sov.data('bankruptcy/shapleys', tickers=["MSFT","TSLA","META"])

Feature Importance (Shapley Values) calculates the contribution of each input variable (features) such as Debt, Assets, and Revenue to predict bankruptcy risk.

Reports

Sorting and Filtering

import sovai as sov
sov.report("bankruptcy", report_type="ranking")

Filter the outputs based on the top by Sector, Marketcap, and Revenue and bankruptcy risk. You can also change ranking to change to investigate the month on month change.

sov.report("bankruptcy", report_type="sector-change")

Plots

Bankruptcy Comparison

import sovai as sov
sov.plot('bankruptcy', chart_type='compare')

Timed Feature Importance

import sovai as sov
df = sov.plot("bankruptcy", chart_type="shapley", tickers=["TSLA"])

Total Feature Importance

import sovai as sov
sov.plot("bankruptcy", chart_type="stack", tickers=["DDD"])

Bankruptcy and Returns

import sovai as sov
df= sov.plot("bankruptcy", chart_type="line", tickers=["DDD"])

PCA Statistical Similarity

import sovai as sov
df= sov.plot("bankruptcy", chart_type="line", tickers=["DDD"])

Correlation Similarity

import sovai as sov
sov.plot("bankruptcy", chart_type="similar", tickers=["DDD"])

Trend Similarity

import sovai as sov
sov.plot("bankruptcy", chart_type="facet", tickers=["DDD"])

Model Performance

Confusion Matrix

import sovai as sov
sov.plot("bankruptcy", chart_type="confusion_global")

Threshold Plots

import sovai as sov
sov.plot("bankruptcy", chart_type="classification_global")

Lift Curve

import sovai as sov
sov.plot("bankruptcy", chart_type="lift_global")

Global Explainability

import sovai as sov
sov.plot("bankruptcy", chart_type="time_global")

Computations

Leverage advanced computational tools for deeper analysis:

• Distance Matrix:

  sov.compute('distance-matrix', on="attribute", df=dataframe)
  

  Assess the similarity between entities based on selected attributes.

• Percentile Calculation:

  sov.compute('percentile', on="attribute", df=dataframe)
  

  Calculate the relative standing of values within a dataset.

• Feature Mapping:

  sov.compute('map-accounting-features', df=dataframe)
  

  Map accounting features to standardized metrics.

• PCA Calculation:

  sov.compute('pca', df=dataframe)
  

  Perform principal component analysis for dimensionality reduction.

For more advanced applications, see the tutotrial.

Data Dictionary

Name	Description	Type	Example
ticker	Stock ticker symbol.	TEXT	"TSLA"
date	Record date.	DATE	2023-09-30
probability_light	LightGBM Boosting Model prediction.	FLOAT	1.46636
probability_convolution	CNN Model prediction for bankruptcies	FLOAT	0.135975
probability_rocket	Rocket Model prediction for time series classification	FLOAT	0.02514
probability_encoder	LightGBM and CNN autoencoders Model prediction.	FLOAT	0.587817
probability_fundamental	Prediction using accounting data only.	FLOAT	1.26148
probability	Average probability across models.	FLOAT	0.553823
sans_market	Fundamental prediction adjusted for market predictions.	FLOAT	-0.20488
volatility	Variability of model predictions.	FLOAT	0.62934
multiplier	Coefficient for model prediction calibration.	FLOAT	1.951868
version	Model/data record version.	INT	20240201

When sans_market is positive, it means that the fundamentals show a larger predicted bankruptcy than what the market predicts (stock might go down in medium term) , when sans_market is negative, the market might have overreacted, and predict a larger probability of bankruptcy than what the fundamentals suggest (stock might go up in medium term).

Use Cases

1. Bankruptcy Prediction Analysis: Offer insights into predicted corporate bankruptcies and identify key factors, clarifying main drivers across different cycles.
2. Variable Impact Breakdown: Analyze how each individual variable affects bankruptcy predictions, providing in-depth feature contribution insights.
3. Temporal Feature Distribution Analysis: Reveal how variables contribute to predictions over time, emphasizing key features in forecasting models.
4. Correlation Discovery: Identify stocks with similar bankruptcy probability trends, revealing correlated market behaviors.
5. Probability Shift Overview: Showcase changes in bankruptcy probabilities among correlated stocks, providing a comprehensive market perspective.
6. Sentiment Inversion Analysis: Convert bankruptcy predictions into positive sentiment indicators to gauge potential impacts on stock returns.
7. Behavioral Similarity Mapping: Locate stocks with similar behaviors to a selected reference, based on bankruptcy trends and PCA feature analysis.