For my first solution I used a global linear multiplier were I pick the coefficient optimizing SMAPE for each split month and forecast span separately.
\nThe second solution is based in SVR. You can check my notebook here: Super Fast RAPIDS SVR
\nOnce microbusiness_density can be calculated using active and county population, I used active as the target labels in my models.
\nFor validation I used last 12 months to calculate the SMAPE for each forecast range. To ease development I tracked only the gain each algorithm gives compared with last value benchmark. For example, predicting 3 months ahead with last value gives SMAPE 2.717, if my algorithm scores 2.617, then the gain will be 0.10 over the baseline.
The SVR gain for each forecast period is:
\n| Forecast | \nSMAPE Gain(12 months avg) | \n
|---|---|
| 1 | \n0.0171 | \n
| 2 | \n0.0611 | \n
| 3 | \n0.1555 | \n
| 4 | \n0.2621 | \n
| 5 | \n0.3524 | \n
| 6 | \n0.4750 | \n
As you can see the gain is not much. But all other algorithms I tried (including GBDTs) scored worse than SVR in my approach.
\nobs.
\nIt's great to see your innovative approach using RAPIDS SVR, and the gains you achieved compared to the last value benchmark. The significant improvement in computational time from using RAPIDS SVR over sklearn SVR is impressive as well. Thank you for sharing your solution and notebook. Good luck on the private leaderboard, and I hope your solution continues to perform well!
", "votes": 3, "replies": [] }, { "id": 2185134, "author_name": "Chris Deotte", "author_url": "", "post_date": "2023-03-16T20:47:51.897000", "content": "Great job Giba. This is a strong SVR model with good CV score. I'm eager to see whether global linear multiplier or SVR will do better on final private LB!
", "votes": 1, "replies": [] }, { "id": 2183237, "author_name": "YaGana Sheriff-Hussaini", "author_url": "", "post_date": "2023-03-15T14:42:07.590000", "content": "Thanks for sharing @titericz. Your solution is interesting as usual and will find time to look carefully later. Might have questions for you then. Good luck.
", "votes": 1, "replies": [] }, { "id": 2185666, "author_name": "CPMP", "author_url": "", "post_date": "2023-03-17T08:01:26.770000", "content": "Interesting. How much better is SVR compared to your other models?
", "votes": 2, "replies": [ { "id": 2186618, "author_name": "Giba", "author_url": "", "post_date": "2023-03-18T00:32:03.267000", "content": "SVR around ~20-30% better than XGB/LGB.
", "votes": 2, "replies": [] } ] }, { "id": 2186641, "author_name": "jr3", "author_url": "", "post_date": "2023-03-18T01:08:21.217000", "content": "Thanks for sharing. Good luck.
", "votes": 1, "replies": [] } ], "comment_count": 6, "messages": [ { "id": 2183213, "postDate": "2023-03-15T14:24:19.943Z", "content": "For my first solution I used a global linear multiplier were I pick the coefficient optimizing SMAPE for each split month and forecast span separately.
\nThe second solution is based in SVR. You can check my notebook here: Super Fast RAPIDS SVR
\nOnce microbusiness_density can be calculated using active and county population, I used active as the target labels in my models.
\nFor validation I used last 12 months to calculate the SMAPE for each forecast range. To ease development I tracked only the gain each algorithm gives compared with last value benchmark. For example, predicting 3 months ahead with last value gives SMAPE 2.717, if my algorithm scores 2.617, then the gain will be 0.10 over the baseline.
The SVR gain for each forecast period is:
\n| Forecast | \nSMAPE Gain(12 months avg) | \n
|---|---|
| 1 | \n0.0171 | \n
| 2 | \n0.0611 | \n
| 3 | \n0.1555 | \n
| 4 | \n0.2621 | \n
| 5 | \n0.3524 | \n
| 6 | \n0.4750 | \n
As you can see the gain is not much. But all other algorithms I tried (including GBDTs) scored worse than SVR in my approach.
\nobs.
\nIt's great to see your innovative approach using RAPIDS SVR, and the gains you achieved compared to the last value benchmark. The significant improvement in computational time from using RAPIDS SVR over sklearn SVR is impressive as well. Thank you for sharing your solution and notebook. Good luck on the private leaderboard, and I hope your solution continues to perform well!
", "rawMarkdown": "It's great to see your innovative approach using RAPIDS SVR, and the gains you achieved compared to the last value benchmark. The significant improvement in computational time from using RAPIDS SVR over sklearn SVR is impressive as well. Thank you for sharing your solution and notebook. Good luck on the private leaderboard, and I hope your solution continues to perform well!", "votes": 3 }, { "id": 2185134, "postDate": "2023-03-16T20:47:51.897Z", "content": "Great job Giba. This is a strong SVR model with good CV score. I'm eager to see whether global linear multiplier or SVR will do better on final private LB!
", "rawMarkdown": "Great job Giba. This is a strong SVR model with good CV score. I'm eager to see whether global linear multiplier or SVR will do better on final private LB!", "votes": 1 }, { "id": 2183237, "postDate": "2023-03-15T14:42:07.590Z", "content": "Thanks for sharing @titericz. Your solution is interesting as usual and will find time to look carefully later. Might have questions for you then. Good luck.
", "rawMarkdown": "Thanks for sharing @titericz. Your solution is interesting as usual and will find time to look carefully later. Might have questions for you then. Good luck.", "votes": 1 }, { "id": 2185666, "postDate": "2023-03-17T08:01:26.770Z", "content": "Interesting. How much better is SVR compared to your other models?
", "rawMarkdown": "Interesting. How much better is SVR compared to your other models?", "votes": 2, "replies": [ { "id": 2186618, "postDate": "2023-03-18T00:32:03.267Z", "content": "SVR around ~20-30% better than XGB/LGB.
", "rawMarkdown": "SVR around ~20-30% better than XGB/LGB. ", "votes": 2 } ] }, { "id": 2186641, "postDate": "2023-03-18T01:08:21.217Z", "content": "Thanks for sharing. Good luck.
", "rawMarkdown": "Thanks for sharing. Good luck.", "votes": 1 } ], "complete": true, "source": "live", "fetched_at": "2026-10-19T19:27:51Z", "competition_id": 41881, "competition_slug": "godaddy-microbusiness-density-forecasting", "team_id": 9628638, "solution_url": "/competitions/godaddy-microbusiness-density-forecasting/writeups/xgiba-giba-rapids-svr-solution" }