The point 3 in rules is not very clear. Can you please elaborate on what cannot be done?
\r\n3. Participants must not attempt to reconstruct full time series data out of the trade examples and previous 10 trades. Doing so in order to obtain the trade price of the test data will be considered cheating and such entries\r\n will be disqualified and not considered for a prize.
\r\nRegards
", "messages": [ { "id": "8303", "postDate": "01/28/2012 04:39:21", "content": "The point 3 in rules is not very clear. Can you please elaborate on what cannot be done?
\r\n3. Participants must not attempt to reconstruct full time series data out of the trade examples and previous 10 trades. Doing so in order to obtain the trade price of the test data will be considered cheating and such entries\r\n will be disqualified and not considered for a prize.
\r\nRegards
", "rawMarkdown": "", "votes": null }, { "id": "8308", "postDate": "01/28/2012 14:04:49", "content": "Definitely needs clarification. We have time series data but are not allowed to use that fact? Why not?
", "rawMarkdown": "", "votes": null }, { "id": "8312", "postDate": "01/28/2012 17:16:31", "content": "We shouldn't try to reconstruct whole trends out of the data we're given. It means that examples have probably overlapping periods of the same bond. So in theory you don't need to predict anything just find a particular row in test/train data in which the\r\n trade price is a part of 10 lagging periods.
\r\nBut it can be quite tricky and it would need clarification. We could reverse engineer test data to get trade prices not to predict anything but to tune model parameters outside of the "official" source code. Nobody could prove that someone used the reverse-engineered\r\n data to choose the best possible parameters - it could be a matter of luck or expert knowledge.
", "rawMarkdown": "", "votes": null }, { "id": "8323", "postDate": "01/28/2012 22:49:57", "content": "This is a bit tricky. What we mean is that you should not use the 10 previous trades to construct a time series of longer than 10 trades. The only data that should be used for each trade is the information in the row. Specifically you may not string 10 trade\r\n pieces together to get more than 10 previous trades and certainly not to attempt to get future trades.
", "rawMarkdown": "", "votes": null }, { "id": "8327", "postDate": "01/29/2012 01:44:47", "content": "Thank you for the reply. Somewhat clearer now:
\r\nAre you concerned about data outside what is given to us? In the sense that one could identify the bond using the 10 datapoints and then get its correct price from somewhere else?
\r\n", "rawMarkdown": "", "votes": null }, { "id": "8331", "postDate": "01/29/2012 10:05:32", "content": "
[quote=DanGlaser;8323]This is a bit tricky. What we mean is that you should not use the 10 previous trades to construct a time series of longer than 10 trades. The only data that should be used for each trade is the information in the row. Specifically you\r\n may not string 10 trade pieces together to get more than 10 previous trades and certainly not to attempt to get future trades.[/quote]Did you mean that\r\nall derived parameters of our models must come from an one particular (predicted) row of the data? And competition is a contest of a tricks for implicit introduction of properties of the reconstructed sequence by selection of "right" models, \r\n cross validation tuned for rediscovering of found trends (or may be just predicted values in other rows) etc. Clarify please what is acceptable and what is not. How we can generalize our models (using more than one row in dataset) and do not violate your restriction.
\r\nHow exactly dataset was sampled? How splitting on the train/test parts was done? How splitting on the private/public leaderboards parts was done?
", "rawMarkdown": "", "votes": null }, { "id": "8369", "postDate": "01/31/2012 14:32:11", "content": "What I'm saying is that you should not try to reconstruct a series of trades longer than 11 (the current trade and the previous 10). This is the main point.
\r\nTrades were taken from a multi-month period over many bonds of many issuers. Not all trades were selected, but there is likely some overlap. Trades were randomly sorted into training and test.
", "rawMarkdown": "", "votes": null }, { "id": "8376", "postDate": "01/31/2012 18:57:31", "content": "[quote=DanGlaser;8369]What I'm saying is that you should not try to reconstruct a series of trades longer than 11 (the current trade and the previous 10). This is the main point.[/quote]What is "reconstruct a series of trades longer than 11"? Give please\r\n strict definition (or change your view on the issue).
\r\nSomeone can build k-NN model based on similarities of 5 consecutive trades (for example). Some of the NN's will give perfect predictions in case of the overlaps. He does not reconstruct 12 or more points of the sequence but exploits presence of the overlaps.
\r\nOther can build a model based on trajectory matrix. He is not reconstruct the long series directly but can have access to a major properties of it.
\r\nThird guy will build model that will holds a Lyapunov surface with attraction points related to the fragment positions in the reconstructed sequence. This may happens without explicit target to do this (and violate the restriction) just in result of stochastic\r\n training and without visible signs that sequence was reconstructed by the surface.
\r\nFragments of a long sequence carry information about the whole. Position of a fragment in the sequence changes properties of the fragment. I am sorry but statement "reconstruct a series of trades longer than 11" is not enough for me to break these dependencies,\r\n understanding your concern and what is allowed and what is not.
", "rawMarkdown": "", "votes": null }, { "id": "8377", "postDate": "01/31/2012 19:05:56", "content": "\r\n
You bring up a valid point and a possible flaw with the competition. Hold for further notice.\r\n
", "rawMarkdown": "", "votes": null } ], "comments": [ { "id": 8308, "author_name": "matlab", "author_url": "", "post_date": "01/28/2012 14:04:49", "content": "Definitely needs clarification. We have time series data but are not allowed to use that fact? Why not?
", "votes": null, "replies": [] }, { "id": 8312, "author_name": "paweljankiewicz", "author_url": "", "post_date": "01/28/2012 17:16:31", "content": "We shouldn't try to reconstruct whole trends out of the data we're given. It means that examples have probably overlapping periods of the same bond. So in theory you don't need to predict anything just find a particular row in test/train data in which the\r\n trade price is a part of 10 lagging periods.
\r\nBut it can be quite tricky and it would need clarification. We could reverse engineer test data to get trade prices not to predict anything but to tune model parameters outside of the "official" source code. Nobody could prove that someone used the reverse-engineered\r\n data to choose the best possible parameters - it could be a matter of luck or expert knowledge.
", "votes": null, "replies": [] }, { "id": 8323, "author_name": "danglaser", "author_url": "", "post_date": "01/28/2012 22:49:57", "content": "This is a bit tricky. What we mean is that you should not use the 10 previous trades to construct a time series of longer than 10 trades. The only data that should be used for each trade is the information in the row. Specifically you may not string 10 trade\r\n pieces together to get more than 10 previous trades and certainly not to attempt to get future trades.
", "votes": null, "replies": [] }, { "id": 8327, "author_name": "sankalp", "author_url": "", "post_date": "01/29/2012 01:44:47", "content": "Thank you for the reply. Somewhat clearer now:
\r\nAre you concerned about data outside what is given to us? In the sense that one could identify the bond using the 10 datapoints and then get its correct price from somewhere else?
\r\n", "votes": null, "replies": [] }, { "id": 8331, "author_name": "alegro", "author_url": "", "post_date": "01/29/2012 10:05:32", "content": "
[quote=DanGlaser;8323]This is a bit tricky. What we mean is that you should not use the 10 previous trades to construct a time series of longer than 10 trades. The only data that should be used for each trade is the information in the row. Specifically you\r\n may not string 10 trade pieces together to get more than 10 previous trades and certainly not to attempt to get future trades.[/quote]Did you mean that\r\nall derived parameters of our models must come from an one particular (predicted) row of the data? And competition is a contest of a tricks for implicit introduction of properties of the reconstructed sequence by selection of "right" models, \r\n cross validation tuned for rediscovering of found trends (or may be just predicted values in other rows) etc. Clarify please what is acceptable and what is not. How we can generalize our models (using more than one row in dataset) and do not violate your restriction.
\r\nHow exactly dataset was sampled? How splitting on the train/test parts was done? How splitting on the private/public leaderboards parts was done?
", "votes": null, "replies": [] }, { "id": 8369, "author_name": "danglaser", "author_url": "", "post_date": "01/31/2012 14:32:11", "content": "What I'm saying is that you should not try to reconstruct a series of trades longer than 11 (the current trade and the previous 10). This is the main point.
\r\nTrades were taken from a multi-month period over many bonds of many issuers. Not all trades were selected, but there is likely some overlap. Trades were randomly sorted into training and test.
", "votes": null, "replies": [] }, { "id": 8376, "author_name": "alegro", "author_url": "", "post_date": "01/31/2012 18:57:31", "content": "[quote=DanGlaser;8369]What I'm saying is that you should not try to reconstruct a series of trades longer than 11 (the current trade and the previous 10). This is the main point.[/quote]What is "reconstruct a series of trades longer than 11"? Give please\r\n strict definition (or change your view on the issue).
\r\nSomeone can build k-NN model based on similarities of 5 consecutive trades (for example). Some of the NN's will give perfect predictions in case of the overlaps. He does not reconstruct 12 or more points of the sequence but exploits presence of the overlaps.
\r\nOther can build a model based on trajectory matrix. He is not reconstruct the long series directly but can have access to a major properties of it.
\r\nThird guy will build model that will holds a Lyapunov surface with attraction points related to the fragment positions in the reconstructed sequence. This may happens without explicit target to do this (and violate the restriction) just in result of stochastic\r\n training and without visible signs that sequence was reconstructed by the surface.
\r\nFragments of a long sequence carry information about the whole. Position of a fragment in the sequence changes properties of the fragment. I am sorry but statement "reconstruct a series of trades longer than 11" is not enough for me to break these dependencies,\r\n understanding your concern and what is allowed and what is not.
", "votes": null, "replies": [] }, { "id": 8377, "author_name": "ineligible", "author_url": "", "post_date": "01/31/2012 19:05:56", "content": "\r\n
You bring up a valid point and a possible flaw with the competition. Hold for further notice.\r\n
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "8303": "", "8308": "", "8312": "", "8323": "", "8327": "", "8331": "", "8369": "", "8376": "", "8377": "" }, "source": "meta" }