Over a million developers have joined DZone.

Prediction Competitions

DZone's Guide to

Prediction Competitions

· Big Data Zone
Free Resource

Effortlessly power IoT, predictive analytics, and machine learning applications with an elastic, resilient data infrastructure. Learn how with Mesosphere DC/OS.

Com­pe­ti­tions have a long his­tory in fore­cast­ing and pre­dic­tion, and have been instru­men­tal in forc­ing research atten­tion on meth­ods that work well in prac­tice. In the fore­cast­ing com­mu­nity, the M com­pe­ti­tion and M3 com­pe­ti­tion have been par­tic­u­larly influ­en­tial. The data min­ing com­mu­nity have the annual KDD cup which has gen­er­ated atten­tion on a wide range of pre­dic­tion prob­lems and asso­ci­ated meth­ods. Recent KDD cups are hosted on kag­gle.

In my research group meet­ing today, we dis­cussed our (lim­ited) expe­ri­ences in com­pet­ing in some Kag­gle com­pe­ti­tions, and we reviewed the fol­low­ing two papers which describe two pre­dic­tion competitions:

  1. Athana­sopou­los and Hyn­d­man (IJF 2011). The value of feed­back in fore­cast­ing com­pe­ti­tions. [preprint ver­sion]
  2. Roy et al (2013). The Microsoft Aca­d­e­mic Search Dataset and KDD Cup 2013.

Some points of discussion:

  • The old style of com­pe­ti­tion where par­tic­i­pants make a sin­gle sub­mis­sion and the results are com­piled by the orga­niz­ers is much less effec­tive than com­pe­ti­tions involv­ing feed­back and a leader­board (such as those hosted on kag­gle). The feed­back seems to encour­age par­tic­i­pants to do bet­ter, and the results often improve sub­stan­tially dur­ing the competition.
  • Too many sub­mis­sions results in over-​​fitting to the test data. There­fore the final scores need to be based on a dif­fer­ent test data set than the data used to score the sub­mis­sions dur­ing the com­pe­ti­tion. Kag­gle does not do this, although they par­tially address the prob­lem by com­put­ing the leader­board scores on a sub­set of the final test set.
  • The met­ric used in the com­pe­ti­tion is impor­tant, and this is some­times not thought through care­fully enough by com­pe­ti­tion organizers.
  • There are sev­eral com­pe­ti­tion plat­forms avail­able now includ­ing Kag­gle, Crow­d­An­a­lytix and Tunedit.
  • The best com­pe­ti­tions are focused on spe­cific domains and prob­lems. For exam­ple, the GEF­com 2014 com­pe­ti­tions are about spe­cific prob­lems in energy forecasting.
  • Com­pe­ti­tions are great for advanc­ing knowl­edge of what works, but they do not lead to data sci­en­tists being well paid as many peo­ple com­pete but few are rewarded.
  • The IJF likes to pub­lish papers from win­ners of pre­dic­tion com­pe­ti­tions because of the exten­sive empir­i­cal eval­u­a­tion pro­vided by the com­pe­ti­tion. How­ever, a con­di­tion of pub­li­ca­tion is that the code and meth­ods are fully revealed, and win­ners are not always happy to comply.
  • The IJF will only pub­lish com­pe­ti­tion results if they present new infor­ma­tion about pre­dic­tion meth­ods, or tackle new pre­dic­tion prob­lems, or mea­sure pre­dic­tive accu­racy in new ways. Just run­ning another com­pe­ti­tion like the pre­vi­ous ones is not enough. It still has to involve gen­uine research results.
  • I would love to see some seri­ous research about pre­dic­tion com­pe­ti­tions, but that would prob­a­bly require a com­pany like kag­gle to make their data pub­lic. See Frank Diebold’s com­ments on this too.
  • A nice side effect of some com­pe­ti­tions is that they cre­ate a bench­mark data set with well tested bench­mark meth­ods. This has worked well for the M3 data, for exam­ple, and new time series fore­cast­ing algo­rithms can be eas­ily tested against these pub­lished results. How­ever, over-​​study of a sin­gle bench­mark data set means that meth­ods are prob­a­bly over-​​fitting to the pub­lished test data. There­fore, a wider range of bench­marks is desirable.
  • Pre­dic­tion com­pe­ti­tions are a fun way to hone your skills in fore­cast­ing and pre­dic­tion, and every stu­dent in this field is encour­aged to com­pete in a few com­pe­ti­tions. I can guar­an­tee you will learn a great deal about the chal­lenges of pre­dict­ing real data — some­thing you don’t always learn in classes or via textbooks.

Learn to design and build better data-rich applications with this free eBook from O’Reilly. Brought to you by Mesosphere DC/OS.


Published at DZone with permission of Rob J Hyndman, DZone MVB. See the original article here.

Opinions expressed by DZone contributors are their own.


Dev Resources & Solutions Straight to Your Inbox

Thanks for subscribing!

Awesome! Check your inbox to verify your email so you can start receiving the latest in tech news and resources.


{{ parent.title || parent.header.title}}

{{ parent.tldr }}

{{ parent.urlSource.name }}