Bankruptcy Prediction using data from 10-K reports

We propose a multimodal approach designed to handle sparse data in the task of bankruptcy prediction.

⚠️ Key Challenges Recap

Challenge	Description
📊 Severe Class Imbalance	Next-year bankruptcies are rare (~0.45%), making supervised learning fragile
🧮 Tabular Data with Missing Values	Missing-by-design or structural NaNs must be handled appropriately
📝 Sparse Text Modalities	Two sources of text (e.g., MD&A and Auditor comments) are high-dimensional and sparse
🔄 Multimodal Fusion	Combining heterogeneous sources (tabular + multiple texts) is nontrivial
🧪 Limited Labeled Data	Labels may be unreliable, noisy, or unavailable for certain years or firms
📈 Need for Robust Generalization	Overfitting on the majority class is easy; detecting subtle minority patterns is hard

Time-based Data Split
Training samples (from 1999-12-31 to 2017-12-31)	29752
Validation samples (from 2018-01-01 to 2018-12-31)	2908
Test samples (from 2019-01-01 to 2021-10-31)	6381
Training bankrupt samples	126
Validation bankrupt samples	19
Test bankrupt samples	47

Financial Dataset Info
Number of features	9700
Mean of number of non-NaN of features per sample	210.37
Std of number of non-NaN features per sample	93.93

Usage:

1. git clone

2. Move /home/andreasideras/research_ideas/BankruptcyPredictionAD/data/ mda_auditor.csv, numerical.csv, subset_mda_auditor.csv and subset_numerical.csv to /data folder. Subset files are just smaller dense datasets for quick experimentation (use --subset True) (INTERNAL-REMOVE IT). You can use either a stratified or a time-split on the bankrupt cases setting --split from values {'stratified','time'}.

3. You can run any modality by giving the argument --modality a value from {financial, mda_auditor, multimodal}, for example:

   python train.py --model_type XGBoost --modality multimodal --n_estimators 100 --max_depth 10 --subset False --tfidf_dim 1000

   This will train the model and save the validation set performance and hyperparameters used in the hyperparameters folder. It saves a model only when there was an improvement on the validation performance based on the config.VAL_METRIC_TO_OPTIMIZE metric.

4. Test the (best) saved model on the test set:

   python test.py --model_type XGBoost --modality multimodal --subset False

To implement your own model, you need to extend the models.base_model.BaseModel class. If some sort of preprocessing is required (e.g., tf-idf extraction - check preprocessing.tfidf_preprocessor.TFIDFpreprocessor), create your own preproccessor by extending the preprocessing.base_preprocessor.BasePreprocessor class. If not, just use the preprocessing.pass_preprocessor.PASSpreprocessor.

Density Percentage of the Top 40 Densest Features
CashAndCashEquivalentsAtCarryingValue	92.3%
IncomeTaxExpenseBenefit	88.6%
StockholdersEquity	86.5%
Assets	86.2%
LiabilitiesAndStockholdersEquity	85.6%
NetIncomeLoss	84.7%
NetCashProvidedByUsedInFinancingActivities	81.4%
NetCashProvidedByUsedInInvestingActivities	81.4%
NetCashProvidedByUsedInOperatingActivities	81.2%
RetainedEarningsAccumulatedDeficit	80.3%
EarningsPerShareBasic	80.2%
WeightedAverageNumberOfSharesOutstandingBasic	79.6%
EarningsPerShareDiluted	79.5%
DeferredIncomeTaxExpenseBenefit	79.3%
WeightedAverageNumberOfDilutedSharesOutstanding	78.8%
ComprehensiveIncomeNetOfTax	76.6%
PropertyPlantAndEquipmentNet	76.4%
ShareBasedCompensation	75.4%
AccumulatedOtherComprehensiveIncomeLossNetOfTax	75.3%
CommonStockSharesAuthorized	72.8%
InterestExpense	71.1%
CommonStockValue	69.2%
AccumulatedDepreciationDepletionAndAmortizationPropertyPlantAndEquipment	69.0%
CommonStockSharesIssued	68.9%
CashAndCashEquivalentsPeriodIncreaseDecrease	68.3%
CommonStockParOrStatedValuePerShare	67.2%
PaymentsToAcquirePropertyPlantAndEquipment	67.1%
CurrentFederalTaxExpenseBenefit	66.6%
OperatingIncomeLoss	65.9%
Liabilities	65.8%
UnrecognizedTaxBenefits	65.0%
PropertyPlantAndEquipmentGross	64.7%
Goodwill	64.6%
DeferredFederalIncomeTaxExpenseBenefit	64.4%
CurrentStateAndLocalTaxExpenseBenefit	64.4%
CurrentIncomeTaxExpenseBenefit	63.8%
EffectiveIncomeTaxRateReconciliationAtFederalStatutoryIncomeTaxRate	61.7%
LiabilitiesCurrent	60.0%
AssetsCurrent	60.0%
CommonStockSharesOutstanding	59.8%