I joined Opera, a multinational big-data analytics firm, after graduating from IIT Delhi. It was an entry-level role as an analytics specialist and my first full-time job.
Opera offered some freedom to choose the team which you wanted to join. After some deliberatiom, I chose to join Lalit Bansal in the Global Markets division. The role was product-oriented. Our flagship product was a cloud-based application suite, Mobiuss, which enabled portfolio construction, valuation and risk management for mortgage markets.
Mobiuss
Mobiuss was actually a collection of web-based products. Mobiuss Resi catered to Residential Whole Loans, Mobiuss RMBS for residential mortgage backed securities,1 and Mobiuss CRE for Commercial Real Estate. Our clients were investment managers in big banks (e.g. Deutsche Bank) who were interested in valuating portfolios of financial instruments based on their risk profile.
A client would upload a portfolio of loans or mortgage backed securities, and they they would see our forecast of their performance under different macroeconomic scenarios. The user could also adjust the granularity of the data: whether it is loan level or aggregated. For each loan, our algorithms would predict the key attributes of default rate, delinquencies, loss severity, prepayment including several others. 2
My responsibilities were varied.
I was tasked with overseeing the data loading process. We had third party vendors which we sourced huge amounts of data from. We would load this data every month - a sequency of SQL queries that would run for more than 24 hours. I had to go over these queries to see if any optimizations were possible.
I worked with the design team for Mobiuss Resi and prepared mockups for its next iteration. At the time I thought this work of mine was pretty top-notch, but I later came to know that my mockups were not used in the final design.
For the launch of Mobiuss Resi, I wrote the user manual.
How does a hurricane affect the bond market?
The Global Markets team regularly gleaned insights from the data they collected and published studies on them.
This was 2013. A year before, Hurricane Sandy had ravaged the states of New York and New Jersey in the US. I did a study on the impact of the hurricane on bond markets.
Our team had created a surveillance dashboard to monitor the performance of the deals and loans from the neighbourhoods most affected by flooding. How we idenified the relevant loans and deals was interesting. Our loan-level data contained US postal zip codes. We superimposed all zip codes on the post-storm flooding survey map. We then filtered the deals which had exposure above certain threshold of loan value. Overall, we identified 45 deals which had roughly $6 billion of exposure.
These deals were then ran through our credit model, their performance forecasted with the help the same algorithm which was used in Mobiuss forecasts (which used Monte Carlo simulations).
The team used the data from Hurricane Katrina (which had hit Louisiana in 2005) as a benchmark. During Katrina, heavy flood damage displaced much of the population, resulting in a huge increase in delinquencies that remained above baseline levels for two years. In our forecast, the 60+ delinquency rates increased as expected. However the size of the increase was a fraction of what was experienced after Katrina in 2005. Delinquency rates increased by 10% and 15%. By contrast, the same figure for Katrina was at 1100%. So it was unlikely the effect of Sandy on bond markets would be as severe
Most of the research had already been done by our excellent researchers: my responsbility was purely to extract the actual data that had come in one year after the hurricane and compare how our forecasts had done.
It turned out that that the actual performance matched the forecasts pretty well. The spike in delinquencies was counteracted by a corresponding surge in loan prepayments, mainly driven by government disaster relief and insurance payouts. As a result, there were smaller peaks in foreclosures and subsequent liquidations that nevertheless produced some stress.
I reported my analysis to Bill Hunt, our head of research, who seemed pretty happy with the accuracy of our model.
Summary
I worked at Opera for six months - after which I had to leave to start my preparations for Oxford. Even so, in that brief stint, I learnt a lot. I learnt how to form assumptions and extract the relevant data for macroeconomic analyses, how web-based products work behind the scenes, and how large swaths of data can be used for forecasting.
A residential mortgage-backed security (RMBS) is a financial represents unwhich cash yields are paid to investors. These cash yields are supported by cash payments received from homeowners who pay interest and principal according to terms agreed to with their lenders. The cash flow received from individuals are pooled and paid out to investors with with waterfall priorities. This enable investors to become comfortable with the certainty of receipt of cash at any point in time. RMBS are used as a funding instrument created by the originator of the mortgage loan without cross-collateralizing individual loans and mortgages - because it would be impossible to receive permission from individual homeowners. ↩
This was very similar to the work I later did at Aurora. Just replace financial portfolios with energy power plants, financial instrument type by power plant type, volatility in house prices with volatility in power generation, portfolio valuation with electricity price. ↩