I am a noob at kaggle so please pardon my silly question. This problem which seems as time series has data of one year. And if we see the plots then there is a peak time in consumption 2-3 months for most of the building or meter types. Since there is no repeating data, how can someone apply walk forward methodology here? To me it seems that the problem can be solved better by taking it as regression problem. Please enlighten me. Thanks
", "messages": [ { "id": "653623", "postDate": "10/20/2019 18:12:40", "content": "I am a noob at kaggle so please pardon my silly question. This problem which seems as time series has data of one year. And if we see the plots then there is a peak time in consumption 2-3 months for most of the building or meter types. Since there is no repeating data, how can someone apply walk forward methodology here? To me it seems that the problem can be solved better by taking it as regression problem. Please enlighten me. Thanks
", "rawMarkdown": "I am a noob at kaggle so please pardon my silly question. This problem which seems as time series has data of one year. And if we see the plots then there is a peak time in consumption 2-3 months for most of the building or meter types. Since there is no repeating data, how can someone apply walk forward methodology here? To me it seems that the problem can be solved better by taking it as regression problem. Please enlighten me. Thanks", "votes": null }, { "id": "653646", "postDate": "10/20/2019 18:44:57", "content": "One thing you have to keep in mind is that the classical time series models exploit patterns in the data. We do not have enough data to understand the intrinsic patterns of each building. Moreover forecasting methods are usually best for short term forecasts because they use a bootstrap process of using previous predictions to make future predictions. The errors will accumulate and the method will necessarily be unreliable after a certain amount of time. General regression models will thus usually outperform standard time series models in this competition. You could however mix both up, by using a classical time series models for the short term (e.g January and February 2017) and a regression model for the long term.
", "rawMarkdown": "One thing you have to keep in mind is that the classical time series models exploit patterns in the data. We do not have enough data to understand the intrinsic patterns of each building. Moreover forecasting methods are usually best for short term forecasts because they use a bootstrap process of using previous predictions to make future predictions. The errors will accumulate and the method will necessarily be unreliable after a certain amount of time. General regression models will thus usually outperform standard time series models in this competition. You could however mix both up, by using a classical time series models for the short term (e.g January and February 2017) and a regression model for the long term.", "votes": null }, { "id": "653834", "postDate": "10/21/2019 03:02:48", "content": "Thanks @maxhalford, please can you tell us based on your experience and the available data, which is maybe the best single model that we can use in this contest. I am thinking about XGBoost and LightGBM, is there any other model that we can took it in consideration.\nThanks
", "rawMarkdown": "Thanks @maxhalford, please can you tell us based on your experience and the available data, which is maybe the best single model that we can use in this contest. I am thinking about XGBoost and LightGBM, is there any other model that we can took it in consideration.\nThanks", "votes": null }, { "id": "653874", "postDate": "10/21/2019 05:11:49", "content": "There is no such thing as a \"best model\" :). You should be spending time thinking about the features, not the model.
", "rawMarkdown": "There is no such thing as a \"best model\" :). You should be spending time thinking about the features, not the model.", "votes": null }, { "id": "653912", "postDate": "10/21/2019 06:57:04", "content": "Thanks a lot sir for your inputs Can you please give your 2 cents on CV? I know Stratified KFold won't work. So should I keep 1 month data (Last month) as a validation set for cv? Or any other approach?
", "rawMarkdown": "Thanks a lot sir for your inputs Can you please give your 2 cents on CV? I know Stratified KFold won't work. So should I keep 1 month data (Last month) as a validation set for cv? Or any other approach?", "votes": null }, { "id": "653923", "postDate": "10/21/2019 07:11:37", "content": "For the moment I've only tried standard k-fold CV with k set to 3. I'll be putting some effort into designing a better CV strategy this week.
", "rawMarkdown": "For the moment I've only tried standard k-fold CV with k set to 3. I'll be putting some effort into designing a better CV strategy this week.", "votes": null }, { "id": "653988", "postDate": "10/21/2019 09:14:40", "content": "Yes, actually in my view a standard KFold won't work as sit will leak the future data points and thus its score won't be reliable enough.
", "rawMarkdown": "Yes, actually in my view a standard KFold won't work as sit will leak the future data points and thus its score won't be reliable enough.", "votes": null }, { "id": "655253", "postDate": "10/22/2019 21:14:10", "content": "To streamline this though process it is useful to know the 3 categories in which missing data can be classified into:\nMissing Completely at Random (MCAR)\nMissing at Random (MAR)\nMissing Not at Random (MNAR)
\n\ndeeper into each one of them and look for potential treatments that can help us circumvent the problems that missing data creates.
\n\nSome ideas:\nTime series imputation\n - Non-time-series specific method\n- mean imputation\nmedian - imputation\nmode - imputation\n- calcucate the appropriate measure and replace NAs with the values. Appropriate for stationary time series, for example, white noise data\n- Random sample imputation\nreplace missing values with observations randomly selected from the remaining (either of it or just some section of it)
\n\nTime-Series specific method\n- Last observation carried forward (LOCF)\n- Next observation carried backward (NOCB)
\n\nThis article can be helpfull: https://www.linkedin.com/pulse/data-cleaning-guide-dealing-na-values-karan-rajwanshi/
", "rawMarkdown": "To streamline this though process it is useful to know the 3 categories in which missing data can be classified into:\nMissing Completely at Random (MCAR)\nMissing at Random (MAR)\nMissing Not at Random (MNAR)\n\ndeeper into each one of them and look for potential treatments that can help us circumvent the problems that missing data creates. \n\nSome ideas:\nTime series imputation\n - Non-time-series specific method\n- mean imputation\nmedian - imputation\nmode - imputation\n- calcucate the appropriate measure and replace NAs with the values. Appropriate for stationary time series, for example, white noise data\n- Random sample imputation\nreplace missing values with observations randomly selected from the remaining (either of it or just some section of it)\n\nTime-Series specific method\n- Last observation carried forward (LOCF)\n- Next observation carried backward (NOCB)\n\nThis article can be helpfull: https://www.linkedin.com/pulse/data-cleaning-guide-dealing-na-values-karan-rajwanshi/", "votes": null } ], "comments": [ { "id": 653646, "author_name": "maxhalford", "author_url": "", "post_date": "10/20/2019 18:44:57", "content": "One thing you have to keep in mind is that the classical time series models exploit patterns in the data. We do not have enough data to understand the intrinsic patterns of each building. Moreover forecasting methods are usually best for short term forecasts because they use a bootstrap process of using previous predictions to make future predictions. The errors will accumulate and the method will necessarily be unreliable after a certain amount of time. General regression models will thus usually outperform standard time series models in this competition. You could however mix both up, by using a classical time series models for the short term (e.g January and February 2017) and a regression model for the long term.
", "votes": null, "replies": [ { "id": 653834, "author_name": "hamditarek", "author_url": "", "post_date": "10/21/2019 03:02:48", "content": "Thanks @maxhalford, please can you tell us based on your experience and the available data, which is maybe the best single model that we can use in this contest. I am thinking about XGBoost and LightGBM, is there any other model that we can took it in consideration.\nThanks
", "votes": null, "replies": [] }, { "id": 653874, "author_name": "maxhalford", "author_url": "", "post_date": "10/21/2019 05:11:49", "content": "There is no such thing as a \"best model\" :). You should be spending time thinking about the features, not the model.
", "votes": null, "replies": [] }, { "id": 653912, "author_name": "yourssharmaji", "author_url": "", "post_date": "10/21/2019 06:57:04", "content": "Thanks a lot sir for your inputs Can you please give your 2 cents on CV? I know Stratified KFold won't work. So should I keep 1 month data (Last month) as a validation set for cv? Or any other approach?
", "votes": null, "replies": [] }, { "id": 653923, "author_name": "maxhalford", "author_url": "", "post_date": "10/21/2019 07:11:37", "content": "For the moment I've only tried standard k-fold CV with k set to 3. I'll be putting some effort into designing a better CV strategy this week.
", "votes": null, "replies": [] }, { "id": 653988, "author_name": "yourssharmaji", "author_url": "", "post_date": "10/21/2019 09:14:40", "content": "Yes, actually in my view a standard KFold won't work as sit will leak the future data points and thus its score won't be reliable enough.
", "votes": null, "replies": [] } ] }, { "id": 655253, "author_name": "caesarlupum", "author_url": "", "post_date": "10/22/2019 21:14:10", "content": "To streamline this though process it is useful to know the 3 categories in which missing data can be classified into:\nMissing Completely at Random (MCAR)\nMissing at Random (MAR)\nMissing Not at Random (MNAR)
\n\ndeeper into each one of them and look for potential treatments that can help us circumvent the problems that missing data creates.
\n\nSome ideas:\nTime series imputation\n - Non-time-series specific method\n- mean imputation\nmedian - imputation\nmode - imputation\n- calcucate the appropriate measure and replace NAs with the values. Appropriate for stationary time series, for example, white noise data\n- Random sample imputation\nreplace missing values with observations randomly selected from the remaining (either of it or just some section of it)
\n\nTime-Series specific method\n- Last observation carried forward (LOCF)\n- Next observation carried backward (NOCB)
\n\nThis article can be helpfull: https://www.linkedin.com/pulse/data-cleaning-guide-dealing-na-values-karan-rajwanshi/
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "653623": "I am a noob at kaggle so please pardon my silly question. This problem which seems as time series has data of one year. And if we see the plots then there is a peak time in consumption 2-3 months for most of the building or meter types. Since there is no repeating data, how can someone apply walk forward methodology here? To me it seems that the problem can be solved better by taking it as regression problem. Please enlighten me. Thanks", "653646": "One thing you have to keep in mind is that the classical time series models exploit patterns in the data. We do not have enough data to understand the intrinsic patterns of each building. Moreover forecasting methods are usually best for short term forecasts because they use a bootstrap process of using previous predictions to make future predictions. The errors will accumulate and the method will necessarily be unreliable after a certain amount of time. General regression models will thus usually outperform standard time series models in this competition. You could however mix both up, by using a classical time series models for the short term (e.g January and February 2017) and a regression model for the long term.", "653834": "Thanks @maxhalford, please can you tell us based on your experience and the available data, which is maybe the best single model that we can use in this contest. I am thinking about XGBoost and LightGBM, is there any other model that we can took it in consideration.\nThanks", "653874": "There is no such thing as a \"best model\" :). You should be spending time thinking about the features, not the model.", "653912": "Thanks a lot sir for your inputs Can you please give your 2 cents on CV? I know Stratified KFold won't work. So should I keep 1 month data (Last month) as a validation set for cv? Or any other approach?", "653923": "For the moment I've only tried standard k-fold CV with k set to 3. I'll be putting some effort into designing a better CV strategy this week.", "653988": "Yes, actually in my view a standard KFold won't work as sit will leak the future data points and thus its score won't be reliable enough.", "655253": "To streamline this though process it is useful to know the 3 categories in which missing data can be classified into:\nMissing Completely at Random (MCAR)\nMissing at Random (MAR)\nMissing Not at Random (MNAR)\n\ndeeper into each one of them and look for potential treatments that can help us circumvent the problems that missing data creates. \n\nSome ideas:\nTime series imputation\n - Non-time-series specific method\n- mean imputation\nmedian - imputation\nmode - imputation\n- calcucate the appropriate measure and replace NAs with the values. Appropriate for stationary time series, for example, white noise data\n- Random sample imputation\nreplace missing values with observations randomly selected from the remaining (either of it or just some section of it)\n\nTime-Series specific method\n- Last observation carried forward (LOCF)\n- Next observation carried backward (NOCB)\n\nThis article can be helpfull: https://www.linkedin.com/pulse/data-cleaning-guide-dealing-na-values-karan-rajwanshi/" }, "source": "meta" }