Problem Description

1
Problem DescriptionUsing Machine Learning to
Make Money at Horse Races

• Pos Draw Btn Horse Wgt Jockey Trainer Age SP
  Comments Raceid
• 1 4 Timocracy 10-0 S Drowne A B Haynes 5 4/9 f
  led after 1f, ridden 2f out, stayed on well and
  in command final furlong opened 4/5 touched 4/5
  800-1100 400-550 400-650 (x3) 400-750
  (x4) 500-1000 (x4) 300-600 200-400 (x2)
  372966
• 2 5 1¾ Bussell Along (IRE) 9-3 S Sanders Stef
  Higgins 4 8/1 held up early, headway on outside
  to chase leaders 4f out, effort and hung left
  from 2f out, went 2nd 1f out, no chance with
  winner opened 10/1 touched 10/1 372966
• Task Given a training set learn a function
  which predicts the winner/selects a horse to bet
  10 on from a given set of entries.
• Performance Measures
• Accuracy in picking the winner of a race (simple
  version)
• Return of placing a 10 bet on a horse in the
  race (advanced version solves the real problem
  trying to make money on the track)

Links http//www.racingpost.com
http//www.drf.com/ http//www.shrp.com/
http//socialmediaseo.net/2010/05/01/kentucky-derb
y-2010/
http//www.equibase.com/premium/eqbPDFChartPlus.cf
m?RACE11BorPPTIDCDCTRYUSADT05/01/2010DAY
DSTYLEEQB
2
Problem Description2

• This is an individual project
• In general, the problem is a ranking problem one
  approach is to learn a function that assigns a
  score to the horses in a race and pick the horse
  with the highest score. But it can also be viewed
  as a classification or prediction problem.
• The datasets will be very basic only containing
  a few attributes, but you are allowed to create
  additional attributes by creating statistics from
  datasets/by extracting information from other
  sources (e.g. percentage of races won by a
  jockey)
• Basically, the project tries to predict the
  future. Likely we will use races of a single race
  track, given you are true temporal sequence of
  race DS1(races in Jan./Feb.), DS2 (races in
  March/April),DS6(true testset---you are not
  allowed to peak into this one only Chun-sheng
  has access to this dataset) which serve as
  training sets, validation sets, test sets, and
  sources of new feature generation in the project.
• Student have freedom in what approaches to
  usethere are many of them adhoc approaches are
  welcome likely every student will use a
  different approach, and some will be quite
  complicated while others use simpler appraoches.
• The goal is to get something running students
  who use a well-tuned simple approach will get a
  better grades than students who use a very
  complicated, sophisticated approach which does
  not run at all.
• Deliverables You will demo your system, write a
  medium-sized report, and Chun-sheng will test
  your system with a test set of his own.
• You are allowed to use any software/tool in the
  project you just have to mention what you used
  in your report/
• In general, the submission deadline is We., March
  23, 11p, but the idea is you spent at most 5
  weeks on the project!

3
Data (Wolverhampton (UK))

• http//maps.google.com/maps?rlz1T4ADRA_enUS403US4
  03um1ieUTF-8qwolverhamptonracecoursefb1g
  lushqwolverhamptonracecoursecid0,0,404086676
  5308641574eiT0RdTfnEFYP98AaOqoGMCwsaXoilocal
  _resultctimageresnum2ved0CC4QnwIwAQ
• http//www.racingpost.com/horses2/cards/meeting_po
  pup.sd?crs_id513action_date2011-02-17selected_
  tabCOURSE_MAP
• http//www.wolverhampton-racecourse.co.uk/
• http//www.racingpost.com/horses2/cards/meeting_po
  pup.sd?crs_id513action_date2011-02-17selected_
  tabMEETING_INFO
• There is a chance that we still change the race
  track, but the data sets formats likely will not
  change.
• Data
• RaceID ID identify the race
• Pos. Finish position of the horse
• Draw The start stall that a horse has been
  allocated
• Dist. The distance a horse has finished behind
  the winner
• Horse The name of the horse
• Wt The weight that the horse carries
• Jockey The name of the Jockey
• Trainer The name of the trainer
• Age The age of the horse
• SP The official starting price of the horse
  (optional)
• RaceHeader Metadata of the race
• RaceDetail Metadata of the race

4
Problem Description3

• More on alternative approaches
• Treat it as a regression/prediction problem e.g.
  assign 1/n to the n-th finisher in the race (or
  use the percentage of the prize money allocated
  to the place the horse took in the race over the
  total price money) then learn a prediction
  function f if h1,,hn are the horses entered
  into a race, use a decision making policy that
  uses f(h1),,f(hn), possibly the prior odds
  oh1),,o(hn), e.g bet on horse i which has the
  maximum value for
• o(hi)/(f(hi)/?nf(hi))
• Treat the problem as a classification problem, in
  which the classification algorithm picks the
  horse to win/the horse to bet on
• Treat the problem as a ranking problem (you can
  learn a similar function as in the first approach
  except the objective function minimizes the
  number of rank violations(e.g. http//olivier.chap
  elle.cc/pub/err.pdfdescribes such a function
  e.g. look at http//www.cs.cornell.edu/people/tj/s
  vm_light/svm_rank.html which gives the code for a
  ranking support vector machine, this approach is
  also connected to a Yahoo! Contest Learning to
  Rank http//learningtorankchallenge.yahoo.com/in
  dex.php
• Honesty Rule Datasets a temporally ordered
  D1ltltDN (lt means prior to) when learning from
  dataset Di, you are only allowed to use knowledge
  from datasets D1,,Di but not from later datasets
  Di1,,Dn
• When using prior statistics, you will have to
  deal with a cold start problem there will be new
  horses, jockeys, usually, you should initialize
  those using average values or other values but
  not 0.
• Your approach should focus on horses which win,
  and not on horses that place second and third a
  lot, due to the performance measures used in the
  project (see first slide).

5
Problem Description 4

• You can use any tools and software packages, such
  as WEKA, regression packages
• Your system should have 3 (2) modules
• A Preprocessing module that formats the dataset
  and adds additional information to the dataset
  (optional)
• A Learning module which takes a dataset and
  creates a model (also reports some training
  statistics)
• A Testing modules that uses a model, picks
  horses, and creates a detailed report of using
  the model for the races in the test set e.g.
• Race 337 Bet on Sally and lost 10
• Race 338 Bet on Enforcer and lost 10
• Race 339 Bet on Caregiver and won 5
  (odds were ½)
• Race 340 Bet on Trailer and won 30
  (odds were 3/1)
• Race 341 Bet on Lateentry and lost
  10
• Total Won 5 total, 1 per race
• Comment There should be a way to deliver your
  model for testing (e.g. to Chun-sheng)
• I suggest you the non-horse picking parts of the
  systems first (e.g. Use choosing the horse with
  the second highest odds as your initial horse
  picking function), and then focus on learning
  good horse picking functions from the project.
• There is a transfer learning aspect of the
  project e.g. you could use the model for other
  race tracks in the UK and US.
• We might allow the use of some basic statistics,
  for jockeys and trainers (but not for horses) for
  the Wolverhampton Race Track such as
  http//www.racingpost.com/horses2/cards/meeting_po
  pup.sd?crs_id513action_date2011-02-17selected_
  tabMEETING_INFO