# Bankruptcy Predictions {% hint style="info" %} Monthly corporate bankruptcy predictions arrive the **2nd of every month***.* {% endhint %} {% hint style="success" %} Dataset contains 4762+ tickers, available from 1998-03-31 onwards. {% endhint %} `Tutorials` are the best documentation — [`Corporate Bankruptcy Tutorial`](https://colab.research.google.com/github/sovai-research/sovai-public/blob/main/notebooks/datasets/Bankruptcy%20Prediction.ipynb)


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

**Specific Dates** ```python import sovai as sov df_bankrupt = sov.data('bankruptcy', start_date="2017-01-03", tickers=["MSFT"]) ``` **Latest Data** ```python import sovai as sov df_bankrupt = sov.data('bankruptcy') ``` **All Data** ```python import sovai as sov df_bankrupt = sov.data('bankruptcy', full_history=True) ``` ### Daily Probabilities ```python 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) ```python 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 ```python 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. ```python sov.report("bankruptcy", report_type="sector-change") ``` ## Plots ### Bankruptcy Comparison ```python import sovai as sov sov.plot('bankruptcy', chart_type='compare') ```

### Timed Feature Importance ```python import sovai as sov df = sov.plot("bankruptcy", chart_type="shapley", tickers=["TSLA"]) ```

### Total Feature Importance ```python import sovai as sov sov.plot("bankruptcy", chart_type="stack", tickers=["DDD"]) ```

### Bankruptcy and Returns ```python import sovai as sov df= sov.plot("bankruptcy", chart_type="line", tickers=["DDD"]) ```

### **PCA Statistical Similarity** ```python import sovai as sov df= sov.plot("bankruptcy", chart_type="line", tickers=["DDD"]) ```

### Correlation Similarity ```python import sovai as sov sov.plot("bankruptcy", chart_type="similar", tickers=["DDD"]) ```

### Trend Similarity ```python import sovai as sov sov.plot("bankruptcy", chart_type="facet", tickers=["DDD"]) ```

## Model Performance ### **Confusion Matrix** ```python import sovai as sov sov.plot("bankruptcy", chart_type="confusion_global") ```

### **Threshold Plots** ```python import sovai as sov sov.plot("bankruptcy", chart_type="classification_global") ```

### **Lift Curve** ```python import sovai as sov sov.plot("bankruptcy", chart_type="lift_global") ```

### Global Explainability ```python import sovai as sov sov.plot("bankruptcy", chart_type="time_global") ```

## Computations Leverage advanced computational tools for deeper analysis: * **Distance Matrix:** ```python sov.compute('distance-matrix', on="attribute", df=dataframe) ``` Assess the similarity between entities based on selected attributes. * **Percentile Calculation:** ```python sov.compute('percentile', on="attribute", df=dataframe) ``` Calculate the relative standing of values within a dataset. * **Feature Mapping:** ```python sov.compute('map-accounting-features', df=dataframe) ``` Map accounting features to standardized metrics. * **PCA Calculation:** ```python 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

{% hint style="info" %} 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)**. {% endhint %} ## 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.