{"id":99,"date":"2008-02-10T00:32:14","date_gmt":"2008-02-09T23:32:14","guid":{"rendered":"http:\/\/www.pauljorion.com\/blog_en\/?p=99"},"modified":"2008-02-10T16:54:14","modified_gmt":"2008-02-10T15:54:14","slug":"agents-using-financial-models-and-the-%e2%80%9chuman-cognitive-cocktail%e2%80%9d","status":"publish","type":"post","link":"https:\/\/www.pauljorion.com\/blog_en\/2008\/02\/10\/agents-using-financial-models-and-the-%e2%80%9chuman-cognitive-cocktail%e2%80%9d\/","title":{"rendered":"Agents using financial models and the \u201chuman cognitive cocktail\u201d"},"content":{"rendered":"

I\u2019m working on the paper to be given at the Human Complex Systems\u2019 one-day conference on March 8th<\/a>. I don\u2019t want to divulge prematurely any scoop but at the same time I\u2019d like to share some of my puzzlement as I go, and as if thinking aloud.<\/p>\n

Rube Goldberg machine<\/a><\/p>\n

To recap: I\u2019m toying with the concept of a large \u2013 as complex as need be \u2013 model of the subprime crisis. The easy parts are those which are already modeled of that Rube Goldberg machine (see diagram), like the cash flow structure (of variable complexity) of the financial instruments at the epicenter of the crisis, i.e. Asset-Backed Securities <\/em>(complex); Collateralized Debt Obligations <\/em>(complex); Asset-Backed Commercial Paper <\/em>(simple) and Credit-Default Swaps <\/em>(simple). <\/p>\n

But of course as soon as I\u2019ve claimed that these financial instruments are truly modeled, a number of caveats come to my mind:<\/p>\n

1) Some assumptions in these models are deemed \u201csubjective.\u201d <\/p>\n

2) Many of these models make wild and unwarranted assumptions about the feasibility of forecasting the future.<\/p>\n

3) However accurate the models might be, those human agents who make decisions from them<\/p>\n

— a. understand them in most cases only partially;
\n— b. make errors when using them \u2013 even when they fully understand them.<\/p><\/blockquote>\n

Let me review this in that order: <\/p>\n

1) \u201cSubjective\u201d assumptions: that means that there\u2019s a range of possible values and one is assigned as the outcome of the data having been processed by the \u201chuman cognitive cocktail\u201d \u2013 see below.<\/p>\n

2) Knowing the future: about some of these models that feel confident at the validity of their wild forecasts, there is \u201cindustry consensus\u201d in finance about that feasibility; to give just one example: the forward yield curve <\/strong>is erroneously assumed to hold information about the future (spot) yield curve<\/strong>. What exactly does \u201cindustry consensus\u201d mean in this instance? That people use the model because they either a) mistakenly believe it to be \u201cscientific,\u201d i.e. true?; b) do suspect it might not be true but use it anyway because of its \u201cinvestment-grade\u201d industry standard rating? Does it matter for any practical purpose if either a or b is the case? <\/p>\n

3) Agent\u2019s understanding of the model used. <\/p>\n

— a. Partial understanding of the model only: it probably just means that agents\u2019 decision-making is loosely linked to the model\u2019s outcome if understanding is minimal and tightly linked to it if it is fair; the decision is a function of the model\u2019s results plus minus a huge or a small epsilon (*) respectively.<\/p>\n

— b. Error when using the model: just epsilon.<\/p><\/blockquote>\n

Returning to the \u201chuman cognitive cocktail,\u201d I call \u201chuman cognitive cocktail,\u201d the methods used by human agents in puzzle-solving (the following list is likely to be revised and refined \u2013 its order is arbitrary): <\/p>\n

1. \u201cFuzzy logic type\u201d of probability calculus.
\n2. \u201cExpert-system-like\u201d combination of logical rules.
\n3. \u201cLogistic regression type\u201d of multi-factorial pseudo-quantitative processing.
\n4. \u201cMulti-layered perceptron type\u201d of pure (= unconscious) intuition.<\/p><\/blockquote>\n

(To be continued\u2026)<\/p>\n

(*) Greek letter traditionally used as a symbol for the error factor.<\/p>\n","protected":false},"excerpt":{"rendered":"

I\u2019m working on the paper to be given at the Human Complex Systems\u2019 one-day conference on March 8th<\/a>. I don\u2019t want to divulge prematurely any scoop but at the same time I\u2019d like to share some of my puzzlement as I go, and as if thinking aloud.<\/p>\n