Machine learning gives organizations a leg-up on predictive analytics
In 2008, the global financial services firm Lehman Brothers officially went bankrupt. It delivered a huge blow to the banking and financial community as, at the time, the firm was the fourth largest investment bank in the United States.
The move seemingly came out of the blue and sent shockwaves across the world, creating the largest financial crisis since the Great Depression. In the aftermath, firms like Merrill Lynch, AIG, Freddie Mac, Fannie Mae, and HBOS all came dangerously close to the same fate.
Since then, companies, lenders, and investors have done considerable research into how to use statistical models to proactively monitor and predict the likelihood of bankruptcy. Financial analysts will take data – like gross margin, debt to assets, and cash flow –and plug them into programs, process the information, and create parametric and non-parametric models for assessing risk.
But, what if a computer could do all the heavy lifting? Today, it’s possible with machine learning technology.
Traditional models for looking at bankruptcy risk
Traditional parametric models look at how different factors will shift a company’s standing in the marketplace. For instance, the number of locations recently opened or closed in a quarter or the market’s views of a new product. Meanwhile, non-parametric models use ratios, such as debt to equity, to determine how well an enterprise is expected to perform.
These models allow companies to intelligently determine if they should continue working with certain partners, vendors, or clients. Banks and other lending institutions can see if an organization is in good standing for a commercial loan. And, investors can decide if they are going to buy or sell stock in a company. But while useful, both models require extensive analysis to gauge their accuracy for predictions.
Let the machines go to work
Machine learning is an approach to artificial intelligence where a system learns or “trains” itself using an initial dataset, then improves from experience without explicit programming.
At Genpact, my team and I pulled together 11 years’ worth of financial data on 267 bankrupt companies and 585 healthy organizations in the United States, stretching across the IT, industrial and healthcare industries. We normalized this large dataset and plugged it in to train machine learning models to recognize warning signs for bankruptcy.
The models combed through the data and found a direct relationship between financial metrics and bankruptcy risk. We then went back in history and selected companies that actually filed for bankruptcy to see if the models held up. We found they were able to accurately predict bankruptcies two years well in advance of an event – which would have been nearly impossible by just manually reviewing the data and using traditional statistical methods.
Now, we are continuously adding more financial data from other companies and industries to further enrich these models.
Machine learning can serve virtually any industry and purpose, but especially in predictive analysis. For instance, in healthcare, a machine learning model can use a database of patient records to understand patients’ symptoms and identify illnesses early on. In shipping and logistics, machine learning models can analyze data on inventory, transit routes, truck location, etc. to accurately determine when a package will reach its destination. In auto insurance, insurers can use machine learning to review data on an applicant’s driving record, personal information, and other variables, compare them with other similar people and calculate the best insurance policy they could offer with the lowest amount of risk.
Equipped with the right data from the beginning, these systems can discover hidden patterns and uncover valuable information so you can always stay one step ahead of the game.