To be right 99% of time.
An amazing article by Joe Nocera that demystify how VaR mislead everyone to take more risk. RISK Mismanagement - What Led to the Financial Meltdown - NYTimes.com
I am a great fan of Nicholas Taleb. I am not an expert on Statistics or Modeling but I find it appalling how data can be misrepresented for vested interests.
One of the most interesting aspects was that VaR gave a 99% risk prediction for the next day.
99% may sound reassuring but the remaining 1% percent failure means 3.5 days in a year. I first encountered 99% folly when I saw my Internet SLA had 99% up time. It means that 3.5 days in a year, I could not do business.
The cost of calculating the confidence level or delivering the service is an exponential curve.
Actually I find it better to get 95% assurance because our minds is tuned to build backups for failure.
The other aspect of this is to get 99% is by using less data to get the desired result. For example, if I get a result which is very different from the desired result using 1000 data points, I would rather use the 100 data points to give me the desired results. This can take variations such as
I am in favor of using data analysis to make decision but make sure the person who makes the judgment is aware of the shortcoming of the data and its analysis.
The gravest mistakes the "quants" have made is not clearly stated the limitations but only dealt with the strengths.
I am a great fan of Nicholas Taleb. I am not an expert on Statistics or Modeling but I find it appalling how data can be misrepresented for vested interests.
One of the most interesting aspects was that VaR gave a 99% risk prediction for the next day.
99% may sound reassuring but the remaining 1% percent failure means 3.5 days in a year. I first encountered 99% folly when I saw my Internet SLA had 99% up time. It means that 3.5 days in a year, I could not do business.
The cost of calculating the confidence level or delivering the service is an exponential curve.
- 95% may cost $1 Million,
- 99% will cost $10 Million,
- 99.9% will cost $100 Million.
Actually I find it better to get 95% assurance because our minds is tuned to build backups for failure.
The other aspect of this is to get 99% is by using less data to get the desired result. For example, if I get a result which is very different from the desired result using 1000 data points, I would rather use the 100 data points to give me the desired results. This can take variations such as
- Making analysis in a smaller time frame with limited data rather than wait to gather more data
- Making analysis use smaller historical data and yet claiming to be 99% correct.
I am in favor of using data analysis to make decision but make sure the person who makes the judgment is aware of the shortcoming of the data and its analysis.
The gravest mistakes the "quants" have made is not clearly stated the limitations but only dealt with the strengths.
Comments