From: John Conover <john@email.johncon.com>
Subject: Re: (no subject)
Date: 3 Aug 2001 15:52:56 -0000
So, how much confidence should one put in market forecasting models
that use regression methodologies?
For a typical industrial market, the average daily marginal increment
is about 0.0004, and the deviation is about 0.02, (note that if those
two numbers are equal, there is no risk, and the market is perfectly
predictable-what these numbers say is that managing risk is the name
of the game in industrial markets; not managing market growth.) Those
numbers have been constant for the ten millenia of civilization.
We know from the previously mentioned graphs, that the deviation adds
root-mean-square, so at the end of a calendar quarter of 60 business
days, the deviation for the quarter will be 0.02 * sqrt (60) = 0.15,
or about 15%. The average adds linearly, so the quarterly average
would be 0.0004 * 60 = 0.024, or a little over 2%.
The probability, from information theory, that the future quarter
would be about like last quarter would be ((0.024 / 0.15) + 1) / 2, or
about 0.58, or just under 60%.
So, when we set around in staff meetings, pondering the future based
on the last quarter's pro forma, (or investing in a stock, based on a
company's quarterly results,) only 6 times out of 10, on the average,
will the conjectures materialize.
Its sobering thought-that is only 10% better than using a tossed coin
as a decision mechanism.
John
BTW, what fraction of the value of the company should be placed at
See:
http://www.johncon.com/ntropix/tsinvest.html#derivation
for a derivation.
John Conover writes:
> Cisco did almost everything right. Almost. They had the correct
> numbers. Where they went wrong was in their interpretation of the
> numbers; they were using regression analysis as a forecasting model.
>
> If you divide the average by the standard deviation of the marginal
> growth in a market, add one, and divide that by two, that gives the
> probability that a market is going to increase-according to entropic
> economics and information theory. The average is the growth of the
> market, and the deviation is a metric of the risk, (that's the way one
> handles risk-reward.)
>
> But one then has to evaluate the accuracy of the calculated
> probability of a market increasing, too; the probability has to be
> multiplied by 1 - 1 / sqrt (n), where n is the interval length used in
> the regression study.
>
> As fate would have it, it makes regression analysis inapplicable in
> business models.
>
> The probability that a market is going to increase is almost never
> higher than 55%, measured on daily information. If a calendar quarter
> is used as the interval for the regression, (that's about 60 business
> days,) one has 0.55 * 1 - 1 / sqrt (60) = 0.55 * 0.87 = 0.48, or about
> 48%.
>
> It is not wise to bet on less than 50% odds-but that's what Cisco did.
>
> John
>
> BTW, what should they have done? They should not have used regression
> analysis. Since industrial markets are fractal, the 1 / sqrt (n) stuff
> works on years, too. The Internet boom began, in earnest, in about
> 1995. By 1999 they should have been throttling manufacturing, since 1
> / sqrt (4) = 0.5, and 1995 + 4 = 1999. 2000 was when Cisco's problems
> started. In short, they should have exploited the dynamics of the
> data, instead of attempting to make sense out of it by smoothing.
>
> See:
>
> http://www.johncon.com/john/correspondence/981014184454.18095.html
> http://www.johncon.com/john/correspondence/981014210544.18525.html
> http://www.johncon.com/john/correspondence/981014222823.18931.html
> http://www.johncon.com/john/correspondence/981014233807.19309.html
>
> for industrial market particulars, and:
>
> http://www.johncon.com/john/correspondence/990215192020.29398.html
> http://www.johncon.com/john/correspondence/990905134341.23530.html
>
> for the US GDP. And,
>
> http://www.johncon.com/ndustrix/FAQs.html#linux
>
> is a series of internal e-mail where such concepts were used to
> develop a strategic framework for a company. The company faired much
> better than Cisco through the tecno-bust of late 2000.
>
> John Conover writes:
> >
> > Attached is a very well written article on the demise of Cisco.
> >
> > Interestingly, IP addresses from Cisco are often found in the
> > logs of http://www.johncon.com/ndustrix/.
> >
> > John
> >
> > http://www.cio.com/archive/080101/cisco_content.html
--
John Conover, john@email.johncon.com, http://www.johncon.com/