Is betting just guessing (do odds reflect the true probability)?

This competition is for those who entered the Data Science Melbourne 2015 Datathon.

You will already have the data for all games upto the semi-finals and finals. The task is to use this historical data to rank order the punters on their profit for the final 3 games of the tournament (which is why we didn't give you this data).

We provide the list of Account_IDs to make predictions for, along with some limited features for the final 3 games that you may make use of.

The objective is to determine if betting is just guessing, or if past performance can be indicative of future performance. We expect this to be very hard, and will be impressed if anyone can come up with an algorithm that is better than a random number generator!

Evaluation

We are treating this as a binary classification problem - did the account make a profit or not. The evaluation metric is the AUC.

An AUC of 0.5 is random guessing and 1 is a prefect solution.

1st submission - 0.60778 (+0.00, +0.00%)

1. basic features
2. random forest
3. weighted profit formula

2nd submission - 0.62711 (+0.01933, +3.18%)

1. new features (BL ratio, cancel ratio etc.)
2. average profit formula

3rd submission - 0.63243 (+0.00532, +0.85%)

1. new features (difference between L and B)
2. xgboost

4th submission - 0.64118 (+0.00875, +1.384%)

1. new feature (invest amount)
2. blended models

5th submission - 0.62621/0.64088

1. X New benchmark (past history by game)
2. X Log transformation
3. X K-means (transactional features & customized imputation)
4. X Feature selection
5. X Multi-rounds
6. X New Calculation

6th submission - 0.63708

1. X Event Counts / Bag of Event
2. O Subset modeling
3. O Invest weigeted calculation

7th submission - 0.64421

1. X Meta features
   • xgboost (gbm, rf)
   • h2o (gbm, rf, nb, glm, dl)
   • spfia (svm, glm)
   • tsne cluster
   • k means cluster
   • fm
   • knn
2. O New customers 0.43/-5

8th submission - 0.63971

1. New Feature
2. Meta bagged modeling
3. O Separate models (new/existing customers)

9th submission - 0.

1. Past value (cumsum)
2. Factorization Machines (http://www.csie.ntu.edu.tw/~r01922136/libffm/)
3. Regression + Classification
4. python lasagne

Ref

1. https://github.com/Gzsiceberg/kaggle-avito
2. entropy based features
3. Bad features: win_hist / DL metafeatures