ここまでの学習モデル検証

学習データは2004.1.1～2024.12.31の21年分で検証期間は2025.1.1～2025.2.16です。

• ShinbaTimeCK.mlnet RMSE 1.9768
  

  41R	1点	芝(19R)	ダート(22R)	8頭以下(0R)	9～12頭(2R)	13頭以上(39R)	多点
  単勝	12.20% (69.76%)	5.26% (31.05%)	18.18% (103.18%)	-- (--)	0.00% (0.00%)	12.82% (73.33%)	36.59% (75.28%)
  複勝	34.15% (82.93%)	21.05% (33.68%)	45.45% (125.45%)	-- (--)	0.00% (0.00%)	35.90% (87.18%)	80.49% (76.42%)
  枠連	7.32% (67.56%)	5.26% (44.74%)	9.09% (87.27%)	-- (--)	0.00% (0.00%)	7.69% (71.03%)	14.63% (136.36%)
  馬連	7.32% (94.15%)	5.26% (44.74%)	9.09% (136.82%)	-- (--)	0.00% (0.00%)	7.69% (98.97%)	9.76% (36.50%)
  ワイド	12.20% (110.73%)	5.26% (21.05%)	18.18% (188.18%)	-- (--)	0.00% (0.00%)	12.82% (116.41%)	17.07% (52.20%)
  馬単	2.44% (49.51%)	5.26% (106.84%)	0.00% (0.00%)	-- (--)	0.00% (0.00%)	2.56% (52.05%)	9.76% (40.20%)
  三連複	2.44% (34.15%)	0.00% (0.00%)	4.55% (63.64%)	-- (--)	0.00% (0.00%)	2.56% (35.90%)	7.32% (289.45%)
  三連単	0.00% (0.00%)	0.00% (0.00%)	0.00% (0.00%)	-- (--)	0.00% (0.00%)	0.00% (0.00%)	7.32% (335.15%)
  総合	34.15% (63.60%)	21.05% (35.26%)	45.45% (88.07%)	-- (--)	0.00% (0.00%)	35.90% (66.86%)	80.49% (215.97%)

• SouhaTimeCK.mlnet RMSE 1.3958
  

  415R	1点	芝(171R)	ダート(226R)	障害(18R)	8頭以下(13R)	9～12頭(81R)	13頭以上(321R)	多点
  単勝	17.35% (61.06%)	10.53% (49.36%)	22.12% (70.75%)	22.22% (50.56%)	23.08% (62.31%)	19.75% (62.96%)	16.51% (60.53%)	45.78% (73.25%)
  複勝	47.95% (87.66%)	40.94% (89.36%)	52.21% (87.30%)	61.11% (76.11%)	53.85% (73.85%)	54.32% (81.48%)	46.11% (89.78%)	85.06% (84.35%)
  枠連	12.02% (81.79%)	12.03% (101.52%)	11.87% (55.48%)	14.29% (270.71%)	-- (--)	18.52% (130.86%)	9.97% (66.60%)	26.85% (79.90%)
  馬連	6.99% (49.59%)	4.68% (37.43%)	8.41% (48.23%)	11.11% (182.22%)	7.69% (23.08%)	13.58% (102.35%)	5.30% (37.35%)	18.07% (59.65%)
  ワイド	17.83% (86.53%)	14.04% (91.52%)	20.35% (82.92%)	22.22% (84.44%)	23.08% (47.69%)	24.69% (74.32%)	15.89% (91.18%)	36.87% (75.12%)
  馬単	3.37% (32.05%)	0.00% (0.00%)	5.75% (54.65%)	5.56% (52.78%)	7.69% (27.69%)	4.94% (64.57%)	2.80% (24.02%)	18.07% (57.34%)
  三連複	4.34% (57.40%)	4.09% (60.18%)	4.87% (59.87%)	0.00% (0.00%)	7.69% (98.46%)	6.17% (58.52%)	3.74% (55.45%)	10.36% (70.19%)
  三連単	0.48% (9.35%)	0.00% (0.00%)	0.88% (17.17%)	0.00% (0.00%)	0.00% (0.00%)	0.00% (0.00%)	0.62% (12.09%)	10.36% (55.31%)
  総合	49.64% (58.01%)	42.11% (53.21%)	54.42% (59.56%)	61.11% (84.43%)	53.85% (47.58%)	58.02% (71.88%)	47.35% (54.63%)	85.30% (62.51%)

• RaceJuniFullCK.mlnet
  N.A.

• ShinbaTimeCK.lgbm RMSE 1.6378
  

  41R	1点	芝(19R)	ダート(22R)	8頭以下(0R)	9～12頭(2R)	13頭以上(39R)	多点
  単勝	9.76% (27.32%)	5.26% (12.63%)	13.64% (40.00%)	-- (--)	50.00% (175.00%)	7.69% (19.74%)	29.27% (50.49%)
  複勝	34.15% (55.12%)	36.84% (56.32%)	31.82% (54.09%)	-- (--)	50.00% (70.00%)	33.33% (54.36%)	75.61% (63.58%)
  枠連	7.32% (37.56%)	15.79% (81.05%)	0.00% (0.00%)	-- (--)	0.00% (0.00%)	7.69% (39.49%)	17.07% (51.62%)
  馬連	2.44% (20.73%)	5.26% (44.74%)	0.00% (0.00%)	-- (--)	0.00% (0.00%)	2.56% (21.79%)	7.32% (50.00%)
  ワイド	4.88% (19.76%)	5.26% (21.05%)	4.55% (18.64%)	-- (--)	0.00% (0.00%)	5.13% (20.77%)	21.95% (43.66%)
  馬単	0.00% (0.00%)	0.00% (0.00%)	0.00% (0.00%)	-- (--)	0.00% (0.00%)	0.00% (0.00%)	7.32% (43.82%)
  三連複	0.00% (0.00%)	0.00% (0.00%)	0.00% (0.00%)	-- (--)	0.00% (0.00%)	0.00% (0.00%)	2.44% (93.96%)
  三連単	0.00% (0.00%)	0.00% (0.00%)	0.00% (0.00%)	-- (--)	0.00% (0.00%)	0.00% (0.00%)	2.44% (54.43%)
  総合	39.02% (20.06%)	47.37% (26.97%)	31.82% (14.09%)	-- (--)	50.00% (30.63%)	38.46% (19.52%)	78.05% (55.59%)

• SouhaTimeCK.lgbm RMSE 1.3662
  

  415R	1点	芝(171R)	ダート(226R)	障害(18R)	8頭以下(13R)	9～12頭(81R)	13頭以上(321R)	多点
  単勝	23.86% (106.55%)	19.88% (150.76%)	26.99% (77.04%)	22.22% (57.22%)	30.77% (65.38%)	32.10% (158.27%)	21.50% (95.17%)	56.14% (95.72%)
  複勝	52.29% (88.22%)	50.29% (104.62%)	53.98% (78.01%)	50.00% (60.56%)	61.54% (73.08%)	61.73% (93.33%)	49.53% (87.54%)	86.51% (82.86%)
  枠連	12.53% (103.73%)	8.86% (109.87%)	15.53% (102.05%)	7.14% (60.71%)	-- (--)	12.35% (147.28%)	12.15% (89.19%)	29.16% (87.10%)
  馬連	10.36% (108.48%)	7.02% (116.14%)	13.72% (111.33%)	0.00% (0.00%)	15.38% (50.00%)	11.11% (120.00%)	9.97% (107.94%)	22.89% (81.43%)
  ワイド	20.00% (75.01%)	15.79% (75.44%)	23.89% (76.95%)	11.11% (46.67%)	30.77% (66.15%)	27.16% (95.93%)	17.76% (70.09%)	38.31% (71.82%)
  馬単	5.30% (111.69%)	4.09% (171.81%)	6.64% (75.09%)	0.00% (0.00%)	7.69% (41.54%)	6.17% (199.63%)	4.98% (92.34%)	22.89% (78.91%)
  三連複	5.30% (64.43%)	3.51% (62.57%)	6.64% (68.36%)	5.56% (32.78%)	15.38% (69.23%)	8.64% (56.67%)	4.05% (66.20%)	12.53% (67.32%)
  三連単	0.96% (105.23%)	1.17% (232.28%)	0.88% (17.48%)	0.00% (0.00%)	7.69% (136.92%)	1.23% (26.79%)	0.62% (123.74%)	12.53% (58.56%)
  総合	53.01% (95.36%)	51.46% (128.11%)	54.42% (75.69%)	50.00% (31.43%)	69.23% (71.76%)	61.73% (112.24%)	50.16% (91.53%)	86.75% (69.36%)

• RaceJuniFullCK.lgbm
  

  456R	1点	芝(190R)	ダート(248R)	障害(18R)	8頭以下(13R)	9～12頭(83R)	13頭以上(360R)	多点
  単勝	21.71% (98.25%)	18.95% (133.53%)	24.19% (75.24%)	16.67% (42.78%)	38.46% (95.38%)	22.89% (84.10%)	20.83% (101.61%)	49.78% (86.12%)
  複勝	53.07% (93.00%)	49.47% (101.47%)	55.24% (87.30%)	61.11% (82.22%)	69.23% (93.08%)	56.63% (81.81%)	51.67% (95.58%)	86.84% (85.42%)
  枠連	13.19% (91.60%)	9.04% (60.79%)	15.77% (104.48%)	21.43% (259.29%)	-- (--)	18.07% (101.57%)	11.67% (86.50%)	28.01% (87.52%)
  馬連	8.99% (78.82%)	4.74% (33.47%)	12.10% (107.02%)	11.11% (168.89%)	38.46% (162.31%)	10.84% (71.57%)	7.50% (77.47%)	19.74% (75.86%)
  ワイド	19.96% (85.68%)	14.21% (79.68%)	23.79% (88.63%)	27.78% (108.33%)	53.85% (123.08%)	27.71% (81.33%)	16.94% (85.33%)	38.16% (77.24%)
  馬単	4.82% (71.71%)	2.11% (23.32%)	6.85% (97.58%)	5.56% (226.11%)	15.38% (168.46%)	7.23% (83.25%)	3.89% (65.56%)	19.74% (73.52%)
  三連複	4.17% (60.90%)	2.63% (24.42%)	5.65% (93.27%)	0.00% (0.00%)	15.38% (120.77%)	7.23% (47.11%)	3.06% (61.92%)	10.96% (77.10%)
  三連単	0.88% (129.82%)	0.00% (0.00%)	1.61% (238.71%)	0.00% (0.00%)	0.00% (0.00%)	1.20% (26.14%)	0.83% (158.42%)	10.96% (70.20%)
  総合	54.17% (88.70%)	50.53% (57.05%)	56.05% (111.55%)	66.67% (106.71%)	69.23% (109.01%)	59.04% (72.11%)	52.50% (91.55%)	87.94% (74.82%)

モデルビルダーはまだまだな感じは否めない。LightGBMは若干悪くない感じですが、新馬戦用がもう少し改善すればって思います。ただ、あまりにもユーザー様を待たせてしまってるので、妥協してこの辺りのモデルで順位予測も学習させてリリースさせるべきかと思います。

追記 2025.2.22 7:58
最初アップした時点ではまだ順位予測全く学習させてませんでしたが、LightGBMでは試しにv4.6.0でしてみた結果を入れてみました。