So this morning I was discussing my yesterday’s post on astrology and vector length with good friend and esteemed colleague Baada. Some interesting fundaes came out of it. Since Baada has given up blogging (and he’s newly married now so can’t expect him to blog) I’m presenting the stuff here.
So basically we believe that astrology started off as some kind of multinomial regression. Some of ancestors observed some people, and tried to predict their behaviour based on the position of their stars at the time of their birth. Maybe it started off as some arbit project. Maybe if blogs existed then, we could say that it started off as a funda session leading up to a blog post.
So a bunch of people a few millenia ago started off on this random project to predict behaviour based on position of stars at the time of people’s birth. They used a set of their friends as the calibration data, and used them to fix the parameters. Then they found a bunch of acquaintances who then became the test data. I’m sure that these guys managed to predict behaviour pretty well based on the stars – else the concept wouldn’t have caught on.
Actually it could have gone two ways – either it fit an extraordinary proportion of people in which case it would be successful; or it didn’t fit a large enough proportion of people in which case it would have died out. Our hunch is that there must have been several models of astrology, and that natural selection and success rates picked out one as the winner – none of the other models would have survived since they failed to predict as well on the initial data set.
So Indian astrology as we know it started off as a multinomial regression model and was the winner in a tournament of several such models, and has continued to flourish to this day. Some problem we find with the concept:
- correlation-causation: what the initial multinomial regression found is that certain patterns in the position of stars at the time of one’s birth is heavily correlated with one’s behaviour. The mistake that the modelers and their patrons made was the common one of associating correlation with causation. They assumed that the position of stars at one’s birth CAUSED one’s behaviour. They probably didn’t do much of a rigorous analysis to test this out
- re-calibration: another problem with the model is that it hasn’t been continuously recalibrated. We continue to use the same parameters as we did several millenia ago. Despite copious quantities of new data points being available, no one has bothered to re-calibrate the model. Times have changed and people have changed but the model hasn’t kept up with either. Now, I think the original information of the model has been lost so no one can recalibrate even if he/she chooses to
Coming back to my earlier post, one can also say that Western astrology is weaker than Indian astrology since the former uses a one-factor regression as against the multinomial regression used by the latter; hence the former is much weaker at predicting.
Do you ever write in normal English that regular people can understand?
astrology is a multinomial logit regression wherein one can only find probability based on various star positions at a point in time. it would be interesting to find the range of the time as a function of time. secondly the legendary error term in this case is quite important given insufficient understanding and the problem of unknown unknowns. probably a test on data points would help tell the nature of its distribution & expected mean
the core premise of this post- that some form of regression was performed to fit the data to a model- IMHO is flawed. It fails to take into account how much people would want a prediction for their future. If at all some kind of fitting was performed, it was done in some Monte Carlo kind of way- small perturbations to the current parameter sets were explored to see if they better fit the recent data. Now, these ‘fits’ were performed by a consortium of people who probably did not compare notes/ take efforts to have a common stable build. So, many flavours existed which were not differentiable to the common man… who just went to the astrologer with the best PR. I can also imagine crossing over occurring in the model- like in a genetic algorithm- where some ‘famous’ parameter values of one of the models get adopted by another. If treated as a genetic algorithm (which you allude to when you mention natural selection), the selection pressure is not stringent and hence allows for a lot of neutral drift… which allows for suboptimal solutions to exist in the population of solutions. Another factor to be considered is that the next generation of astrologers who take a currently ‘well performing’ model may not be competent enough, thereby allowing it to be superceded by other models. This feature adds to the neutral drift property of selection.
Also, many predictions/applications of astrology are in benign areas where a failure of the model cannot be noticed. For example, most weddings just worked in indian society, irrespective of astrology.
another feature of multinomial regression that comes to mind is IIA- which basically states that addition of another random noise kind of parameter to the model will not affect the other parameters. This feature would be particularly useful to the R&D astrologers, who could include new parameters and further fine tune the model.
One problem with using regression is the presence of correlation in the dataset of free variables- ie the planetary positions are not independent of each other.
OK. enough rambling. sorry for not writing in some cogent manner- i kind of just jotted down a subset of issues that came to mind when reading the blog. maybe one can blog about the train of thoughts that led to this comment- because i didnt edit the comment at all… so there should be a narrative… i hope its non-linear 🙂
whoa! finally got down to reading your entier comment. good stuff only. I had only written about the intial tournament and picking winner in that, but your funda of Genetic Algorithms is a killer! amazing stuff
good point, that the correlation between planetary positions might put spanner in the regression – but how ‘strongly’ correlated are the planetary positions among themselves? It’s cyclical, probably, but not necessarily correlated, because outside of the cycle, each body will have it’s own path with it’s own cycle.
An astrologer is a BSer who gets paid for making inconsistent statements and non-falsifiable claims. Astrology has survived as it offers “false hope” to people, which is a therapeutic cure for most of the mind problems.
Hi,
Given that you have an interest in astrology, thought I will let you know of a couple of good references on astrology:
1) Jagannatha Hora software:
Does not do any predictions but does all the calculations – as predictions are dependent on interpretation.
Written by PVR Narasimha Rao, IIT Chennai Computer Science Grad, AIR top 10 given away for free
http://www.vedicastrologer.org/jh/
A good book for beginners exists by same author
Otherwise you can also refer books by James Braha
2) For slightly advanced topics, a great book is
“Art and practice of ancient indian astrology: Nine intimate sessions between teacher and student” by James Braha
strong.
Really good one
1) The first issue is that of co-integration. While the position of planets is correlated to the events in life, one doesnt cause the other and thus any pooja etc. wouldnt have any effect (thats how I logically think but dont believe and have a feeling that may be I dont know enough as of now)
2) Re-caliberation is missing these days totally. Vedic astrology is a very very complex data set to model. you have 9 planets(variables) each of which can have 28 positions (values). But then it is expected if this has to model human future.
Thats my long term desire to one day successfully model it up