Company Bankruptcy Prediction!

This project is a collaboration between the Pacmann Data School's Introduction to Machine Learning (IML) and Machine Learning Process (MLP) courses.

General Introductions

Financial risk prediction is an important topic in financial analysis because it can assist businesses in reducing financial distress and taking appropriate actions in the future. Bankruptcy prediction has been a topic of discussion for nearly a century, and it remains one of the hottest topics in economics. The goal of predicting financial institutions' distress is to create a predictive model that combines various econometric measures and allows one to forecast a firm's financial situation at least 1 year in the future. Predicting corporate bankruptcy is critical for both investors and businesses.

Introduction to Machine Learning (IML)

The IML project is primarily concerned with machine learning modeling experiments. The 2 notebooks folder contains the source code for the modeling and testing scenarios that were completed. See: for a detailed modeling

https://docs.google.com/document/d/1wBAJakg23FQ721Ep7cdm6Ab7pohvkP0rBosVS8d8siY/edit?usp=sharing

Machine Learning Process (MLP).

The MLP project focuses on end-to-end machine learning workflows, beginning with understanding business problems, deriving business metrics, machine learning metrics, and creating fire services for the backend and front end services until they are ready for server deployment. Details about the report can be found at:

https://docs.google.com/document/d/10AAaHIiPcAMV1gkTiMz54yJ8PeRf4Cv_hlIvaSW_DGQ/edit?usp=sharing

How to run API and Streamlit (Front End) Services

1. Check that you have Docker installed. If not, search for "How to install Docker on Windows/Ubuntu/Os that you used."

2. Make sure you have Git installed as well. If not, search for "How to install git on Windows/Ubuntu/Os that you used."

3. Make a clone of this repository : git clone https://github.com/bimapriambodo/Company_Bankruptcy_Prediction-.git

4. Navigate to the cloned folder's directory. Open a CMD terminal and compile the docker file as follows: docker-compose up Wait until everything is completed. If your internet connection is fast, the process will be as well.

5. To test Api, open a browser and enter the following address : http://0.0.0.0:8080/docs

6. To try Streamlit, launch a browser and type the following address : http://172.18.0.3:8501/

Congratulations, you are now free to use the application !!!

Input API Data

Input Api uses json format like this:

{
  "Attr21": 0.0,
  "Attr13": 0.0,
  "Attr27": 0.0,
  "Attr34": 0.0,
  "Attr24": 0.0,
  "Attr35": 0.0,
  "Attr51": 0.0,
  "Attr6": 0.0,
  "Attr56": 0.0,
  "Attr49": 0.0,
  "Attr40": 0.0,
  "Attr55": 0.0,
  "Attr50": 0.0,
  "Attr58": 0.0,
  "Attr43": 0.0
}

---

VARIABLES	DATA TYPE
Attr21	FLOAT
Attr13	FLOAT
Attr27	FLOAT
Attr34	FLOAT
Attr24	FLOAT
Attr35	FLOAT
Attr51	FLOAT
Attr6	FLOAT
Attr56	FLOAT
Attr49	FLOAT
Attr40	FLOAT
Attr55	FLOAT
Attr50	FLOAT
Attr58	FLOAT
Attr43	FLOAT

---

Future Works

Many scenarios have not yet been tried and would be quite interesting iterative combinations to investigate as material for future works, for example,

1. We do not remove outliers from the data in this experiment because we believe outliers are a natural phenomenon and provide information. However, what can be tested in future works if the outliers are removed?

2. Standard scalers are not the only option for standardizing data. There is a min/max scaler as well as other methods. What do you think if you compare it with other methods? Is there any other influence? It's possible.

3. Not only are machine learning methods used in this trial, but many are widely used, such as neural networks, Cat-boost, light GBM, and others. Can try to be compared, it appears to be interesting.

4. We do not perform feature engineering or variable reduction based on existing variables. Then, for example, try dimensional reduction using PCA, how does it work?

5. Deploy to Cloud Services !

We hope that this trial will provide an overview of highly influential iterative processes in data science. Matt, stay hungry and foolish!

Reference

Chang, Hanxu. 2019. The Application of Machine Learning Models in Company Bankruptcy Prediction. ICSEB: International Conference on Software and e-Business. Tokyo: Japan

Zieba,Maciej. 2016. Ensemble Boosted Trees with Synthetic Features Generation in Application to Bankruptcy Prediction. Expert Systems with Applications: An International Journal

Thi Kha Nguyen. 2018. PREDICTING BANKRUPTCY USING MACHINE LEARNING ALGORITHMS. JOURNAL OF SCIENCE AND TECHNOLOGY, NO. 12(133). THE UNIVERSITY OF DANANG

https://chirag-sehra.medium.com/decision-trees-explained-easily-28f23241248

https://towardsdatascience.com/understanding-random-forest-58381e0602d2#:~:text=The%20random%20forest%20is%20a,that%20of%20any%20individual%20tree.

https://towardsdatascience.com/machine-learning-basics-with-the-k-nearest-neighbors-alalgorithm-6a6e71d01761

https://www.ibm.com/id-en/topics/knn#:~:text=The%20k%2Dnearest%20neighbors%20algorithm%2C%20also%20known%20as%20KNN%20or,of%20an%20individual%20data%20point.

https://ai.plainenglish.io/xgboost-in-a-nutshell-211e170e8b48

https://towardsdatascience.com/curse-of-dimensionality-a-curse-to-machine-learning-c122ee33bfeb