{ "id": 284924, "title": "What exactly are we predicting...", "url": "/competitions/g-research-crypto-forecasting/discussion/284924", "author_name": "", "post_date": "2021-11-02T22:18:49.977556700Z", "votes": 27, "comment_count": 11, "views": 0, "content": "

So what's the prediction time frame.

\n\n

Do crypto coins even have a close of trading. Or are we just predicting a price at the end of a 24 hr period after the 9hrs of CPU/GPU time. So there's always a gap of at most, 9 hrs between the data we have to do prediction and the prediction start… There's much that I'm confused about as far as what we are actually predicting here.

", "messages": [ { "id": "1568601", "postDate": "11/02/2021 22:18:49", "content": "

So what's the prediction time frame.

\n\n

Do crypto coins even have a close of trading. Or are we just predicting a price at the end of a 24 hr period after the 9hrs of CPU/GPU time. So there's always a gap of at most, 9 hrs between the data we have to do prediction and the prediction start… There's much that I'm confused about as far as what we are actually predicting here.

", "rawMarkdown": "So what's the prediction time frame. \n- The price of the 14 crypto coins at the end of the 3 month window? \n- The daily close of the 14 crypto assets for the next 90 days?\n- the daily close of the 14 crypto assets for the next day only?\n- the price range of the crypto assets over the next 24 hrs, 90 days or ?\n\nDo crypto coins even have a close of trading. Or are we just predicting a price at the end of a 24 hr period after the 9hrs of CPU/GPU time. So there's always a gap of at most, 9 hrs between the data we have to do prediction and the prediction start... There's much that I'm confused about as far as what we are actually predicting here.", "votes": null }, { "id": "1568607", "postDate": "11/02/2021 22:23:48", "content": "

price movement over the next 24 hrs, week, month, 90 days. and at what granularity seconds, milliseconds, minutes, microseconds, hours, …

", "rawMarkdown": "price movement over the next 24 hrs, week, month, 90 days. and at what granularity seconds, milliseconds, minutes, microseconds, hours, ...", "votes": null }, { "id": "1568608", "postDate": "11/02/2021 22:25:48", "content": "

and how far in the future does the prediction start the contest limits it to 9hrs of cpu and gpu time so at most it would need to be 9hrs in advance of the latest data but if the model only takes 1 hr would it be 1hr in advance of the latest data…

\n

and is this going to be a continuous model prediction feeding data in real time to it and having it generate predictions for whatever the offset will be…

", "rawMarkdown": "and how far in the future does the prediction start the contest limits it to 9hrs of cpu and gpu time so at most it would need to be 9hrs in advance of the latest data but if the model only takes 1 hr would it be 1hr in advance of the latest data...\n \nand is this going to be a continuous model prediction feeding data in real time to it and having it generate predictions for whatever the offset will be...", "votes": null }, { "id": "1568642", "postDate": "11/02/2021 23:43:31", "content": "

LOL I was going to make the same question. Are we predicting an hour into the future?, a day, a week? Or predicting the price is going up, down? Predicting Elon Musk will tweet and fuck all models?

", "rawMarkdown": "LOL I was going to make the same question. Are we predicting an hour into the future?, a day, a week? Or predicting the price is going up, down? Predicting Elon Musk will tweet and fuck all models?", "votes": null }, { "id": "1568691", "postDate": "11/03/2021 01:31:13", "content": "

From my initial understanding, at a simple level you are attempting to predict the log returns over a 15 minute period. This it then adjusted for market signal and the weight of the asset (as indicated by the weight asset file). Then you predictions are compared to the actual value using a Pearson correlation.

\n

See formulas here

", "rawMarkdown": "From my initial understanding, at a simple level you are attempting to predict the log returns over a 15 minute period. This it then adjusted for market signal and the weight of the asset (as indicated by the weight asset file). Then you predictions are compared to the actual value using a Pearson correlation. \n\n[See formulas here](https://www.kaggle.com/cstein06/tutorial-to-the-g-research-crypto-competition/notebook#Building-your-prediction-model)", "votes": null }, { "id": "1568717", "postDate": "11/03/2021 02:20:27", "content": "

@proselotis ok if it's the log return over a 15 minute period how far in advance are we talking. The models will take some time to predict/infer the return. I don't recall if the 9hr cputime/gputime is training effort or inferencing. but if it's just training then, the model sits and accepts as input the last 15 minutes of activities and it is intended to predict the next 15 minutes of activity from that. OR is the model taking the last 24 hours of activity and predicting the next 15 or is the model taking more or less data in to predict the next 15 minutes of activity

", "rawMarkdown": "proselotis ok if it's the log return over a 15 minute period how far in advance are we talking. The models will take some time to predict/infer the return. I don't recall if the 9hr cputime/gputime is training effort or inferencing. but if it's just training then, the model sits and accepts as input the last 15 minutes of activities and it is intended to predict the next 15 minutes of activity from that. OR is the model taking the last 24 hours of activity and predicting the next 15 or is the model taking more or less data in to predict the next 15 minutes of activity", "votes": null }, { "id": "1568720", "postDate": "11/03/2021 02:23:19", "content": "

… there must be some sort of cut off we use to predict the next 15 minutes of log return for the cryptocoins
\nand that will take some time to input, and infer from so there's got to be some lag in the data in and the prediction out and this will be a function of the inferencing time which will be a function of the data stream read time

", "rawMarkdown": "... there must be some sort of cut off we use to predict the next 15 minutes of log return for the cryptocoins\nand that will take some time to input, and infer from so there's got to be some lag in the data in and the prediction out and this will be a function of the inferencing time which will be a function of the data stream read time", "votes": null }, { "id": "1569310", "postDate": "11/03/2021 10:51:23", "content": "

I'm mainly confused about the forecast horizon: Are the features given, the values at the present, and the target is the {log returns} over the next 15 minutes? OR, are the features given for the future, i.e 15 minutes from now?

", "rawMarkdown": "I'm mainly confused about the forecast horizon: Are the features given, the values at the present, and the target is the {log returns} over the next 15 minutes? OR, are the features given for the future, i.e 15 minutes from now?", "votes": null }, { "id": "1569612", "postDate": "11/03/2021 15:30:25", "content": "

But the central question is what are we using (being fed) to predict that log return over a 15 minute period. 15 minutes of prior history, 1hr, 1day, 1week… worth of pricing history for the coin? the more data we get to chew on the longer it's going to take to predict a given 15 minute interval. And let's say we get the prediction of a 15 minute window for a coin which interval are we actually predicting. Let's say we have a prediction for log return for a coin at T0 minutes, Is the prediction for the next T1-15 minutes or should it be for the next T16-30, T31-45, T45-60 etc. This will bear on how we process and train the model

", "rawMarkdown": "But the central question is what are we using (being fed) to predict that log return over a 15 minute period. 15 minutes of prior history, 1hr, 1day, 1week... worth of pricing history for the coin? the more data we get to chew on the longer it's going to take to predict a given 15 minute interval. And let's say we get the prediction of a 15 minute window for a coin which interval are we actually predicting. Let's say we have a prediction for log return for a coin at T0 minutes, Is the prediction for the next T1-15 minutes or should it be for the next T16-30, T31-45, T45-60 etc. This will bear on how we process and train the model", "votes": null }, { "id": "1572493", "postDate": "11/05/2021 17:56:35", "content": "

the 9 hours may be just the time it takes to load data, create features, generate and train a model..

\n

but once model has been created, it may only take milliseconds (or microseconds) for each individual test data to be input and a target output generated.. (its the training of the initial model that can take most of the time)

\n

[I'm an experienced coder though my is ML relatively new, so maybe a more experienced person can elaborate/correct ]

", "rawMarkdown": "the 9 hours may be just the time it takes to load data, create features, generate and train a model.. \n\nbut once model has been created, it may only take milliseconds (or microseconds) for each individual test data to be input and a target output generated.. (its the training of the initial model that can take most of the time)\n\n[I'm an experienced coder though my is ML relatively new, so maybe a more experienced person can elaborate/correct ]", "votes": null }, { "id": "1575478", "postDate": "11/08/2021 12:25:13", "content": "

So it's essential to get the length of time window clear.

", "rawMarkdown": "So it's essential to get the length of time window clear.", "votes": null }, { "id": "1575529", "postDate": "11/08/2021 12:56:11", "content": "

i am also confused in that what is the target prediction period .can any one from competition please clarify this .This is no where mention in overview or else where.

", "rawMarkdown": "i am also confused in that what is the target prediction period .can any one from competition please clarify this .This is no where mention in overview or else where.", "votes": null } ], "comments": [ { "id": 1568607, "author_name": "rlucc001", "author_url": "", "post_date": "11/02/2021 22:23:48", "content": "

price movement over the next 24 hrs, week, month, 90 days. and at what granularity seconds, milliseconds, minutes, microseconds, hours, …

", "votes": null, "replies": [] }, { "id": 1568608, "author_name": "rlucc001", "author_url": "", "post_date": "11/02/2021 22:25:48", "content": "

and how far in the future does the prediction start the contest limits it to 9hrs of cpu and gpu time so at most it would need to be 9hrs in advance of the latest data but if the model only takes 1 hr would it be 1hr in advance of the latest data…

\n

and is this going to be a continuous model prediction feeding data in real time to it and having it generate predictions for whatever the offset will be…

", "votes": null, "replies": [ { "id": 1572493, "author_name": "regipublic", "author_url": "", "post_date": "11/05/2021 17:56:35", "content": "

the 9 hours may be just the time it takes to load data, create features, generate and train a model..

\n

but once model has been created, it may only take milliseconds (or microseconds) for each individual test data to be input and a target output generated.. (its the training of the initial model that can take most of the time)

\n

[I'm an experienced coder though my is ML relatively new, so maybe a more experienced person can elaborate/correct ]

", "votes": null, "replies": [] } ] }, { "id": 1568642, "author_name": "gyangarcia", "author_url": "", "post_date": "11/02/2021 23:43:31", "content": "

LOL I was going to make the same question. Are we predicting an hour into the future?, a day, a week? Or predicting the price is going up, down? Predicting Elon Musk will tweet and fuck all models?

", "votes": null, "replies": [] }, { "id": 1568691, "author_name": "proselotis", "author_url": "", "post_date": "11/03/2021 01:31:13", "content": "

From my initial understanding, at a simple level you are attempting to predict the log returns over a 15 minute period. This it then adjusted for market signal and the weight of the asset (as indicated by the weight asset file). Then you predictions are compared to the actual value using a Pearson correlation.

\n

See formulas here

", "votes": null, "replies": [] }, { "id": 1568717, "author_name": "rlucc001", "author_url": "", "post_date": "11/03/2021 02:20:27", "content": "

@proselotis ok if it's the log return over a 15 minute period how far in advance are we talking. The models will take some time to predict/infer the return. I don't recall if the 9hr cputime/gputime is training effort or inferencing. but if it's just training then, the model sits and accepts as input the last 15 minutes of activities and it is intended to predict the next 15 minutes of activity from that. OR is the model taking the last 24 hours of activity and predicting the next 15 or is the model taking more or less data in to predict the next 15 minutes of activity

", "votes": null, "replies": [] }, { "id": 1568720, "author_name": "rlucc001", "author_url": "", "post_date": "11/03/2021 02:23:19", "content": "

… there must be some sort of cut off we use to predict the next 15 minutes of log return for the cryptocoins
\nand that will take some time to input, and infer from so there's got to be some lag in the data in and the prediction out and this will be a function of the inferencing time which will be a function of the data stream read time

", "votes": null, "replies": [] }, { "id": 1569310, "author_name": "danofer", "author_url": "", "post_date": "11/03/2021 10:51:23", "content": "

I'm mainly confused about the forecast horizon: Are the features given, the values at the present, and the target is the {log returns} over the next 15 minutes? OR, are the features given for the future, i.e 15 minutes from now?

", "votes": null, "replies": [] }, { "id": 1569612, "author_name": "rlucc001", "author_url": "", "post_date": "11/03/2021 15:30:25", "content": "

But the central question is what are we using (being fed) to predict that log return over a 15 minute period. 15 minutes of prior history, 1hr, 1day, 1week… worth of pricing history for the coin? the more data we get to chew on the longer it's going to take to predict a given 15 minute interval. And let's say we get the prediction of a 15 minute window for a coin which interval are we actually predicting. Let's say we have a prediction for log return for a coin at T0 minutes, Is the prediction for the next T1-15 minutes or should it be for the next T16-30, T31-45, T45-60 etc. This will bear on how we process and train the model

", "votes": null, "replies": [] }, { "id": 1575478, "author_name": "complacement", "author_url": "", "post_date": "11/08/2021 12:25:13", "content": "

So it's essential to get the length of time window clear.

", "votes": null, "replies": [] }, { "id": 1575529, "author_name": "rajgbabu", "author_url": "", "post_date": "11/08/2021 12:56:11", "content": "

i am also confused in that what is the target prediction period .can any one from competition please clarify this .This is no where mention in overview or else where.

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1568601": "So what's the prediction time frame. \n- The price of the 14 crypto coins at the end of the 3 month window? \n- The daily close of the 14 crypto assets for the next 90 days?\n- the daily close of the 14 crypto assets for the next day only?\n- the price range of the crypto assets over the next 24 hrs, 90 days or ?\n\nDo crypto coins even have a close of trading. Or are we just predicting a price at the end of a 24 hr period after the 9hrs of CPU/GPU time. So there's always a gap of at most, 9 hrs between the data we have to do prediction and the prediction start... There's much that I'm confused about as far as what we are actually predicting here.", "1568607": "price movement over the next 24 hrs, week, month, 90 days. and at what granularity seconds, milliseconds, minutes, microseconds, hours, ...", "1568608": "and how far in the future does the prediction start the contest limits it to 9hrs of cpu and gpu time so at most it would need to be 9hrs in advance of the latest data but if the model only takes 1 hr would it be 1hr in advance of the latest data...\n \nand is this going to be a continuous model prediction feeding data in real time to it and having it generate predictions for whatever the offset will be...", "1568642": "LOL I was going to make the same question. Are we predicting an hour into the future?, a day, a week? Or predicting the price is going up, down? Predicting Elon Musk will tweet and fuck all models?", "1568691": "From my initial understanding, at a simple level you are attempting to predict the log returns over a 15 minute period. This it then adjusted for market signal and the weight of the asset (as indicated by the weight asset file). Then you predictions are compared to the actual value using a Pearson correlation. \n\n[See formulas here](https://www.kaggle.com/cstein06/tutorial-to-the-g-research-crypto-competition/notebook#Building-your-prediction-model)", "1568717": "proselotis ok if it's the log return over a 15 minute period how far in advance are we talking. The models will take some time to predict/infer the return. I don't recall if the 9hr cputime/gputime is training effort or inferencing. but if it's just training then, the model sits and accepts as input the last 15 minutes of activities and it is intended to predict the next 15 minutes of activity from that. OR is the model taking the last 24 hours of activity and predicting the next 15 or is the model taking more or less data in to predict the next 15 minutes of activity", "1568720": "... there must be some sort of cut off we use to predict the next 15 minutes of log return for the cryptocoins\nand that will take some time to input, and infer from so there's got to be some lag in the data in and the prediction out and this will be a function of the inferencing time which will be a function of the data stream read time", "1569310": "I'm mainly confused about the forecast horizon: Are the features given, the values at the present, and the target is the {log returns} over the next 15 minutes? OR, are the features given for the future, i.e 15 minutes from now?", "1569612": "But the central question is what are we using (being fed) to predict that log return over a 15 minute period. 15 minutes of prior history, 1hr, 1day, 1week... worth of pricing history for the coin? the more data we get to chew on the longer it's going to take to predict a given 15 minute interval. And let's say we get the prediction of a 15 minute window for a coin which interval are we actually predicting. Let's say we have a prediction for log return for a coin at T0 minutes, Is the prediction for the next T1-15 minutes or should it be for the next T16-30, T31-45, T45-60 etc. This will bear on how we process and train the model", "1572493": "the 9 hours may be just the time it takes to load data, create features, generate and train a model.. \n\nbut once model has been created, it may only take milliseconds (or microseconds) for each individual test data to be input and a target output generated.. (its the training of the initial model that can take most of the time)\n\n[I'm an experienced coder though my is ML relatively new, so maybe a more experienced person can elaborate/correct ]", "1575478": "So it's essential to get the length of time window clear.", "1575529": "i am also confused in that what is the target prediction period .can any one from competition please clarify this .This is no where mention in overview or else where." }, "source": "meta" }