Sometimes doing uni work can be practical

Some of you may recall I said I had to do a stats assignment this
semester for the master's, so I decided to do it on motogp (cos I was
planning on changing jobs and hence didn't want to use work data and
besides doing it on work stuff is boring)

So for anyone who cares(1), the model which best predicts the
championship rankings for motoGP turning out to be a straight forward
linear model (none of the fancy stuff added much of any value).

The variables considered included:
Prev Yr Championship pts
2yrs Prev Championship pts
No. of Motogp podiums
No. of junior class (125/250) podiums
No. of cylinders
Vee vs Inline
Dummy variables for:
Honda, Yamaha and Ducati
Rookie Year
Low Performing team
Non GP Championship winner (ie WSBK etc)

These whittled down in the parsimonious model to:

Final Championship Pts = 34.5707 + 0.4723(Last yrs points) + 0.1983(yr
before last's pts) + 28.0384(Dummy for rookie yr) + 57.3952 (Dummy for
Honda) -41.7039(dummy for low performer)

For those not familiar with stats a dummy(2) is a binary yes/no variable
- so if the rider is in their rookie year then it would be 1 for yes and
0 if it's not their rookie year.

Or to put it another way the best predictor of how you'll go next year
is how you went this year, but if you're on a Honda you have an
advantage (yes I know that's intuitively obvious but now we have
statistical "proof"(3)), and the interesting thing is a rookie is likely
to go better in their first year (keenness factor) although I'm thinking
that result may be a bit warped by Bayliss' bad trot in his second year
with ducati.

I'll tot up the podium finishes for this year and add to the data set
and predict the 05 outcome for laughs, it'll be interesting to see how
good the model actually turns out to be.

JL
1 I've tried to translate this into layman's terms, for those serious
maths heads out there I can send you the data and regression stats if
you want
2 I just KNOW Postman Pat or Clem are going to have a field day with
that one !
3 You PhD maths types bugger off, I know it's not a proof per se, it's
statistically significant etc etc <yawn>

 

and people say I have too much time on my hands..

TB

 

Maaaate... I don't have nearly enough time on my hands 'cos I'm doing
stuff like this

JL
(it's a uni assignment, you don't often get a choice to at least do it
on something mildly interesting)

 

I built a brake system for a Formula SAE car for my final year thesis...

TB

 

Hmm the difference between doing engineering and business

JL

 

Stuck the updated podium data into the model and re-ran the regression
and got slightly different outcome, 02 ranking got dropped and overall
motogp podiums got put back in so now there's a slightly controversial
result - it's predicting that 05 will finally be Biaggi's year !!

New Model top 10
Rider Predicted 05

BIAGGI Max 246.56
ROSSI Valentino 243.98
GIBERNAU Sete 225.79
BARROS Alex 195.35
TAMADA Makoto 157.93
HAYDEN Nicky 145.64
BAYLISS Troy 128.51
MELANDRI Marco 127.45
CAPIROSSI Loris 123.41
CHECA Carlos 119.59

Submitted model (the one I posted earlier) gives the more predictable
result of a Rossi win (although interestingly it's suggesting that Checa
will beat his new team mate)

Rider Pred 05
Submittedmodel
ROSSI Valentino 248.54
BIAGGI Max 237.08
GIBERNAU Sete 223.46
BARROS Alex 210.34
TAMADA Makoto 162.81
HAYDEN Nicky 147.22
MELANDRI Marco 127.38
BAYLISS Troy 125.49
CHECA Carlos 117.79
CAPIROSSI Loris 111.44

Don't worry too much about the actual numbers the standard error is
about 15 points per variable so the outcome numbers could be
significantly farther apart (plus or minus 40 or 50 points) it's meant
to rank them rather than predict a specific number outcome.

Personally, intuitively, I think both rankings are wrong, I reckon
Bayliss is going to come out higher up the rankings than predicted, and
Checa is going to crash his arse out of contention. Capirossi will
remain consistent and come in about 5th overall.

JL

 

That's cool.

Did the variables: "No. of junior class podiums" and "Non GP
Championship winner" exhibit a high degree of multicollinearity?

How do you objectively define a low performing team?


Jules

 

You could have at least put SPOILER in the subject line ehehe.

I haven;t had a chance to watch the tape of next years events yet so I'll look away now.

 

Could I have a copy please ? I'll just setup my online betting account
so I can put some money down ;-)

JL

 

No, none at all - because I didn't include winning the 250's or 125's as
a non GP champ - it was BSB WSBK etc only. I figured the high quantity
of jnr class podiums was a good enough descriptor, and I don't think it
really matters too much in the scheme of things for GP success as to
whether they won the 250s or came second.

Well, that's the difference between a Pass and a HD isn't it ? When
I get the marks back I'll tell you whether I succeeded in being
objective <grin>

My argument was around relative team budget size and gross turnover over
of the underlying factories - you can see that Yam, Duc, Suzuki and
Honda have big numbers there, and Aprilia, Kwaka, Harris WCM, Proton etc
don't. I also bought up the number of factory teams in other racing
series (ie Ducati are somewhat small but have a strong racing heritage
and lineage and a big presence in WSBK etc). Dunno if it'll fly

If the lecturer knew more about bikes & GP he might catch me out on the
relative size of Kawasaki Heavy Industries vs Kawasaki Motorcycles,and
the Aprilia dominance in 125/250s but I think I'm safe there <grin>

JL

 

the Aprilia dominance in 125/250s but I think I'm safe there <grin>

Until the professor Googles your name / topic looking for plagarism ;-)

 

google is soooooo last year.

We use www.turnitin.com these days.

Don't even think about incorporating a little snippet you found on the web.


Moike

 

DOH !

Well there's no plagariasm at least,now THAT would be a drama ! If he
thinks to google up this and finds me mentioning the weakness in my
argument, well hey, good luck to him

JL

 

You can also select your own topics. I designed a racebike chassis
around my IT490 engine for my thesis.
It was a great excuse to learn all about bike handling and chassis
design.

Mark

 

How does it go predicting this years results or the 2003 results?
That might provide some validataion without having to wait a year. It
does seem that the usual suspects are up front though so it can't be
too far off.

Mark

 

My third year dynamics thesis was on a suspension system based on the Hossack/ telelever
system.

TB

 

If that is the measure of it's validity then designing a formula is easy.

last_season_result = next_season_result.

Nev..
'03 ZX12R

 

Not as easy as it sounds Nev, if you derive a descriptor of behaviour
based on this year's results and then substitute last years (and N lags
before that if required) you aren't necessarily going to 100% predict
this year's.

FWIW, last season's = this season's IS a significant predictor of
outcome, but if you think about it the rankings in 03 don't 100% match
the 04 results - which is what he was suggesting.

JL

 

It's usefulness to predict 04 wasn't too bad, you can't go much further
back though - think about how often riders and teams come into and leave
GP or at least have done in the last few years - you end up with
insufficient degrees of freedom to be able to draw stat sig conclusions,
given a full field is about 25 and there's been a turnover of at least 6
or 7 a year, go back more than 2 years and 2/3rds of your field isn't
there.

Realistically it's a snapshot of a couple of years, there's too much
volatility to expect a long term consistent model (IMHO, and not being a
serious stats/maths head, it's just a subject in a course work Masters)

JL

 

But the general structure of the end of season points table is about the
same. You wouldn't expect the end of season results to match perfectly
because minor fluctuations in the results are inevitable, and while your
complex formula would have taken propensity to crash into consideration (being
a function of past years championship points) it can't predict severity of
crashes which will lead to even greater fluctuations in results... etc

Nev..
'03 ZX12R