Loading ...
Sorry, an error occurred while loading the content.

The Fourth Quarter

Expand Messages
  • John Carter
    One of my favourite authors at the moment is Nassim Nicholas Taleb. So he mostly speaks about finance, because that s what he knows, but underneath he is
    Message 1 of 1 , Sep 18 4:38 PM
      One of my favourite authors at the moment is Nassim Nicholas Taleb.

      So he mostly speaks about finance, because that's what he knows, but
      underneath he is mostly talking about statistics and risk. And that is
      what we all deal with.

      His previous book "The Black Swan" in some senses wasn't very
      useful... it can be abbreviated as "We are really Bad at prediction,
      much worse than you believe."

      His latest essay is actually quite handy. It provides a map of where
      we are going to be startlingly Bad at predicting.


      Many Agilisto's will say, "Yip, he is right, thats why our practices

      Others may look at Taleb's essay and get an "Aha!" moment and finally
      realize why some of the Agile practices work.

      He suggests you divide problems on the basis of moments of a random variable.

      If your decision is a "yes/no" choice, it is simple. Will the project
      be finished by the 5th of December 2008?

      If your question is based on the value of a random variable, it is more
      complex. What will be the completion date of the project?

      If your question is based on a higher moment of a random variable, it is very
      complex. What will be the ROI of a project?

      Then look at the nature of the randomness... Is it fat tailed, or well behaved?

      For non-statistical types a probability distribution can be fat or
      thin tailed. The one you learned about in the stats course you have
      mostly forgotten was a thin tailed one. (Gauss / Normal distribution).

      Odds on if you did any stats course they went on for hours about thin
      tailed distributions, because they can do the mathematics for them.

      Unfortunately most real world distributions are fat tailed.

      If you have a 1000 guys in the company, the average weight of
      employees is simply not going to shift by much if you employ the
      fattest guy in the world. (Fat guys come from a thin tail probability

      If you look at a 1000 random project case studies, the average project
      overrun is going to massively shift if you add the worlds largest
      project overrun.

      ie. Things like food requirements for project workers are random
      variables from what Taleb calls "mediocristan".

      Things like time to completion are from "extremistan".

      So if divide your problems in to quadrants like this....

      Simple Payoffs | Complex (Higher Moment) payoffs

      Thin tail distribution Predictable | Less predictable
      Fat tail distribution Less predictable | You're utterly stuffed.

      Exercise for the Reader...

      1) Catalogue the random variables in your work situation and
      categorise them as from mediocristan or extremistan.

      eg. Time to complete an item of work - Extremistan

      Programmer Productivity - Very high variance, but probably

      Security Risks - Extremistan. (No valid distribution on attack models, motivations etc.)

      Exchange Rate fluctuations - Extremistan

      Programmer Defect rates - Not sure. Maybe mediocristan for simple monothreaded programs.
      Extremistan for concurrent.

      Probably "Programmer attribute XXX" for what ever XXX is
      truncated. Truncated in one sense because we don't hire / or we
      fire / or we "transfer" the worst.

      Truncated in another sense because our biological meatware is
      from mediocristan.


      2) Catalogue (crudely) the predictions you are required to make by moment of the random variables involved.

      Will the project be completed by christmas?
      Do we need an extra hard drive?
      Is the CPU fast enough?

      How long will the project take?
      How much disk space will we use?

      Number of page hits per day is correlated to number of registered
      users. Look at our records and tell me what is the ratio?

      Number of page hits per day is correlated to number of registered
      users. What is the error bars on that number?

      We have N registered users, what will be the worst case response time?

      3) Which predictions have you made in your last project proposal that
      were in the "utterly stuffed" quadrant?

      Note the gotcha, we are appalling bad at estimating what the worst
      case response time will be, but sort of OK at estimating "Is the CPU
      fast enough". How does that work?

      Easy. If the CPU was very very slightly too slow, it wasn't fast
      enough. If the response time was longer than the age of the
      universe... the CPU still wasn't fast enough. ie. Whoever asked the
      question, probably was asking the wrong question.

      John Carter Phone : (64)(3) 358 6639
      Tait Electronics Fax : (64)(3) 359 4632
      PO Box 1645 Christchurch Email : john.carter@...
      New Zealand
    Your message has been successfully submitted and would be delivered to recipients shortly.