Our final model was a blend of 6 GBM models (5 Lightgbm, 1 Xgboost) all trained on the same features. We validated on the last month of data and then retrained the model on the entire dataset.
\nOur final feature set had over 1000 features that pretty fell under 3 categories:
\nstrikeouts_1_day_ago, strikeouts_2_days_ago, …, strikeouts_20_days_ago for each player. Some of our most important features were:
\n604.Submission Notebook: https://www.kaggle.com/brandenkmurray/mlb-predict-final
", "messages": [ { "id": "1525183", "postDate": "09/27/2021 07:56:24", "content": "Our final model was a blend of 6 GBM models (5 Lightgbm, 1 Xgboost) all trained on the same features. We validated on the last month of data and then retrained the model on the entire dataset.
\nOur final feature set had over 1000 features that pretty fell under 3 categories:
\nstrikeouts_1_day_ago, strikeouts_2_days_ago, …, strikeouts_20_days_ago for each player. Some of our most important features were:
\n604.Submission Notebook: https://www.kaggle.com/brandenkmurray/mlb-predict-final
", "rawMarkdown": "Our final model was a blend of 6 GBM models (5 Lightgbm, 1 Xgboost) all trained on the same features. We validated on the last month of data and then retrained the model on the entire dataset.\n\n- LightGBM - objective=”regression_l1”\n- LightGBM - objective=”regression_l2” - targets scaled by double square root\n- LightGBM - objective=”regression_l1” - boost_from_average=True\n- LightGBM - objective=”regression_l1”, 'max_depth': -1\n- LightGBM DART - object=’regression_l1’, boosting=’dart’\n- Xgboost - targets scaled by double square root\n\nOur final feature set had over 1000 features that pretty fell under 3 categories:\n\n- Target aggregates - All-time and rolling 12-month historical mean/variance of target for each player.\n- Last 20 days/games of various stats for each player. For example, `strikeouts_1_day_ago`, `strikeouts_2_days_ago`, ..., `strikeouts_20_days_ago` for each player. \n- Features based on our baseball knowledge/what gets people to tweet. For example, a walk-off hit/home run will often trigger some engagement. Other features were no-hitters, win probability added by a player (a proxy for if they played a good game or were involved in \"high leverage\" plays), ranking in the home run race, was a player ejected, ERA and ERA ranking, and more.\n\n\nSome of our most important features were:\n- numberOfFollowers - most recent value for a players number of Twitter followers\n- numberOfFollower_delta - the change in Twitter followers between the most recent 2 months\n- monthday - integer value representing current month and day (mmdd) e.g. June 4th would have a value of `604`.\n- {target}_p_var - Historical variance of {target} for player\n- {target}_p_gameday_mean - Historical mean of {target} for player on gamedays\n- roll12_{target}_p_mean - {target} mean for player for the previous 12 months\n- roll12_{target}_p_var - {target} variance for player for the previous 12 months\n- wpa_daily_max - League-wide daily maximum value of Win Probability Added. An approximate (exact base-out states are not available, so calculation is approximate) WPA value is calculated for each player. \n- homeRuns_rank - Players’ home run ranking \n- walk_off_league - Was there a walk off hit in the league that day?\n- days_since_last_start - number of days since a player last pitched. Useful for the model to learn if a player is likely to pitch on the day engagement is measured.\n\nSubmission Notebook: https://www.kaggle.com/brandenkmurray/mlb-predict-final", "votes": null }, { "id": "1525804", "postDate": "09/27/2021 16:32:28", "content": "Congrats and thanks for sharing! I had tried numberOfFollowers, numberOfFollowers_delta, monthday and they were not useful for me when I add these features to my existing pipeline. What kind of feature importance do you mean when you say these were the most important features?
\nI think league-wise features might have made the difference.
", "rawMarkdown": "Congrats and thanks for sharing! I had tried numberOfFollowers, numberOfFollowers_delta, monthday and they were not useful for me when I add these features to my existing pipeline. What kind of feature importance do you mean when you say these were the most important features?\n\nI think league-wise features might have made the difference.", "votes": null } ], "comments": [ { "id": 1525804, "author_name": "aerdem4", "author_url": "", "post_date": "09/27/2021 16:32:28", "content": "Congrats and thanks for sharing! I had tried numberOfFollowers, numberOfFollowers_delta, monthday and they were not useful for me when I add these features to my existing pipeline. What kind of feature importance do you mean when you say these were the most important features?
\nI think league-wise features might have made the difference.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1525183": "Our final model was a blend of 6 GBM models (5 Lightgbm, 1 Xgboost) all trained on the same features. We validated on the last month of data and then retrained the model on the entire dataset.\n\n- LightGBM - objective=”regression_l1”\n- LightGBM - objective=”regression_l2” - targets scaled by double square root\n- LightGBM - objective=”regression_l1” - boost_from_average=True\n- LightGBM - objective=”regression_l1”, 'max_depth': -1\n- LightGBM DART - object=’regression_l1’, boosting=’dart’\n- Xgboost - targets scaled by double square root\n\nOur final feature set had over 1000 features that pretty fell under 3 categories:\n\n- Target aggregates - All-time and rolling 12-month historical mean/variance of target for each player.\n- Last 20 days/games of various stats for each player. For example, `strikeouts_1_day_ago`, `strikeouts_2_days_ago`, ..., `strikeouts_20_days_ago` for each player. \n- Features based on our baseball knowledge/what gets people to tweet. For example, a walk-off hit/home run will often trigger some engagement. Other features were no-hitters, win probability added by a player (a proxy for if they played a good game or were involved in \"high leverage\" plays), ranking in the home run race, was a player ejected, ERA and ERA ranking, and more.\n\n\nSome of our most important features were:\n- numberOfFollowers - most recent value for a players number of Twitter followers\n- numberOfFollower_delta - the change in Twitter followers between the most recent 2 months\n- monthday - integer value representing current month and day (mmdd) e.g. June 4th would have a value of `604`.\n- {target}_p_var - Historical variance of {target} for player\n- {target}_p_gameday_mean - Historical mean of {target} for player on gamedays\n- roll12_{target}_p_mean - {target} mean for player for the previous 12 months\n- roll12_{target}_p_var - {target} variance for player for the previous 12 months\n- wpa_daily_max - League-wide daily maximum value of Win Probability Added. An approximate (exact base-out states are not available, so calculation is approximate) WPA value is calculated for each player. \n- homeRuns_rank - Players’ home run ranking \n- walk_off_league - Was there a walk off hit in the league that day?\n- days_since_last_start - number of days since a player last pitched. Useful for the model to learn if a player is likely to pitch on the day engagement is measured.\n\nSubmission Notebook: https://www.kaggle.com/brandenkmurray/mlb-predict-final", "1525804": "Congrats and thanks for sharing! I had tried numberOfFollowers, numberOfFollowers_delta, monthday and they were not useful for me when I add these features to my existing pipeline. What kind of feature importance do you mean when you say these were the most important features?\n\nI think league-wise features might have made the difference." }, "source": "meta" }