Company Bankruptcy Prediction

Inspiration

The pandemic caused many businesses across the country to struggle to stay afloat. We wanted to predict the likelihood that a business would go bankrupt, based on historical business statistics.

What it does

It predicts whether a company is likely to go bankrupt based on features such as, net income, net value per share, debt ratio, their cashflow rate, and their inventory.

How we built it

We decided to create a logistic neural network, based on the frameworks of tensor and keras. We implemented a confusion matrix in order to analyze the accuracy of our model. We used a random 75% of our data set to train, and the other random 25% to validate our model. We utilized 4 layers. Our first layer was 16 nodes in length, had an activation of relu, and an input size of 5114 X 50. Our second layer, was 16 nodes in length, had an activation of relu. Our third layer was 16 nodes in length, had an activation of relu. Our final, and output layer was 1 node in size, and had an activation of sigmoid. we had 100 epochs, and a batch size of 80 for each epoch.

Challenges we ran into

Our original data set had 96 features, this caused the model to less accurate, as more features were utilized. In order to combat this, we uploaded our dataset into a software called Orange3. This allowed us to rank each feature from most influential to least influential. We then took the top 50 most influential features, and trained our model using that. Our original model had 300 epochs, however, when we ran our model, we noticed that our training loss was much lower than our validation loss. This indicated to us that our model was not learning from the training data, but instead memorizing it. In order to mitigate this, we lowered the number of epoch our model was trained on, from 300 to 100 features.

Accomplishments that we're proud of

We were proud that our model was able to predict the likelihood of of a business going bankrupt, with around 96% accuracy.

What we learned

We learned that the most influential feature was the degree of financial leverage. This had a negative correlation with whether or not a company will go bankrupt, meaning that the lower the leverage, the greater the risk of bankruptcy.