Update BOOM results with refreshed Toto 2.0 evaluations

results/leaderboards/BOOMLET_leaderboard.csv

@@ -1,24 +1,24 @@
 model,MASE,CRPS,Rank
-toto_2.0_2.5b,0.583,0.485,2.589
-toto_2.0_1b,0.585,0.485,2.6
-toto_2.0_313m,0.586,0.485,2.811
-toto_2.0_22m,0.601,0.5,4.389
-toto_2.0_4m,0.624,0.517,5.967
-Toto-Open-Base-1.0,0.617,0.519,6.411
-chronos_2,0.623,0.522,6.711
-timesfm_2_5_200m,0.638,0.532,7.722
+toto_2.0_2.5b,0.588,0.486,2.633
+toto_2.0_1b,0.584,0.486,2.711
+toto_2.0_313m,0.585,0.486,2.9
+toto_2.0_22m,0.601,0.5,4.322
+toto_2.0_4m,0.624,0.517,5.878
+Toto-Open-Base-1.0,0.617,0.519,6.4
+chronos_2,0.623,0.522,6.667
+timesfm_2_5_200m,0.638,0.532,7.7
 moirai_1.1_large,0.767,0.621,11.478
 timesfm_2_0_500m,0.685,0.603,11.511
-moirai_1.1_base,0.779,0.63,11.767
-moirai_1.1_small,0.786,0.631,12.133
-moirai_2_small,0.902,0.467,12.278
+moirai_1.1_base,0.779,0.63,11.756
+moirai_1.1_small,0.786,0.631,12.144
+moirai_2_small,0.902,0.467,12.267
 chronos_bolt_base,0.711,0.637,12.844
 chronos_bolt_small,0.717,0.642,13.078
 time-moe-200M,0.793,0.78,17.178
-time-moe-50M,0.810,0.788,17.222
+time-moe-50M,0.81,0.788,17.222
 timer,0.807,0.793,17.622
 autoarima,0.922,0.88,18.011
 visionts,0.912,0.885,19.6
-seasonalnaive,1.000,1.0,20.122
+seasonalnaive,1.0,1.0,20.122
 autoets,0.969,15.664,20.756
-autotheta,1.030,1.182,21.2
+autotheta,1.03,1.182,21.2

results/leaderboards/BOOM_domain_leaderboard.csv

@@ -16,9 +16,9 @@ timer,0.871,0.716,0.728,0.871,0.828,0.636,0.619,0.655,0.725,0.664
 timesfm_2_0_500m,0.736,0.765,0.679,0.765,0.717,0.441,0.44,0.471,0.493,0.525
 timesfm_2_5_200m,0.654,0.655,0.577,0.65,0.686,0.384,0.372,0.397,0.408,0.476
 Toto-Open-Base-1.0,0.639,0.635,0.567,0.635,0.682,0.378,0.362,0.391,0.4,0.476
-toto_2.0_1b,0.603,0.596,0.536,0.605,0.654,0.352,0.335,0.369,0.382,0.454
-toto_2.0_2.5b,0.601,0.595,0.534,0.61,0.654,0.353,0.339,0.371,0.382,0.456
-toto_2.0_22m,0.623,0.613,0.552,0.627,0.666,0.367,0.349,0.38,0.391,0.462
-toto_2.0_313m,0.607,0.6,0.539,0.611,0.651,0.357,0.34,0.366,0.381,0.453
-toto_2.0_4m,0.648,0.642,0.569,0.652,0.68,0.382,0.361,0.39,0.406,0.473
+toto_2.0_1b,0.603,0.597,0.536,0.605,0.651,0.352,0.336,0.363,0.379,0.452
+toto_2.0_2.5b,0.602,0.597,0.534,0.609,0.653,0.352,0.336,0.362,0.379,0.455
+toto_2.0_22m,0.623,0.613,0.552,0.627,0.666,0.366,0.348,0.375,0.391,0.462
+toto_2.0_313m,0.605,0.6,0.538,0.613,0.654,0.354,0.339,0.365,0.383,0.454
+toto_2.0_4m,0.648,0.642,0.569,0.652,0.68,0.382,0.361,0.389,0.406,0.473
 visionts,1.044,1.003,0.858,1.045,0.92,0.692,0.637,0.663,0.736,0.735

results/leaderboards/BOOM_leaderboard.csv

@@ -1,24 +1,24 @@
 model,MASE,CRPS,Rank
-toto_2.0_2.5b,0.581,0.353,3.902
-toto_2.0_1b,0.582,0.351,4.006
-toto_2.0_313m,0.586,0.353,4.371
-toto_2.0_22m,0.601,0.365,5.558
-Toto-Open-Base-1.0,0.617,0.375,6.926
-toto_2.0_4m,0.624,0.378,7.175
-chronos_2,0.641,0.392,7.367
-timesfm_2_5_200m,0.632,0.382,7.68
-moirai_1.1_base,0.710,0.428,10.375
-moirai_1.1_large,0.720,0.436,10.727
-moirai_1.1_small,0.729,0.442,11.228
-moirai_2_small,0.672,0.427,11.273
-timesfm_2_0_500m,0.725,0.447,11.9
-chronos_bolt_base,0.726,0.451,12.321
-chronos_bolt_small,0.733,0.455,12.725
-autoarima,0.824,0.736,17.128
-time-moe-50M,0.806,0.649,17.514
-time-moe-200M,0.881,0.643,17.531
-timer,0.796,0.639,17.581
-visionts,0.988,0.673,18.721
-autoets,0.842,1.975,19.349
-autotheta,1.123,1.018,20.233
-seasonalnaive,1.000,1.000,20.41
+toto_2.0_2.5b,0.581,0.349,3.876
+toto_2.0_1b,0.582,0.349,3.964
+toto_2.0_313m,0.585,0.351,4.258
+toto_2.0_22m,0.601,0.363,5.525
+Toto-Open-Base-1.0,0.617,0.375,6.939
+toto_2.0_4m,0.624,0.377,7.173
+chronos_2,0.641,0.392,7.39
+timesfm_2_5_200m,0.632,0.382,7.702
+moirai_1.1_base,0.71,0.428,10.395
+moirai_1.1_large,0.72,0.436,10.743
+moirai_1.1_small,0.729,0.442,11.244
+moirai_2_small,0.672,0.427,11.286
+timesfm_2_0_500m,0.725,0.447,11.907
+chronos_bolt_base,0.726,0.451,12.339
+chronos_bolt_small,0.733,0.455,12.74
+autoarima,0.824,0.736,17.14
+time-moe-50M,0.806,0.649,17.522
+time-moe-200M,0.881,0.643,17.538
+timer,0.796,0.639,17.59
+visionts,0.988,0.673,18.725
+autoets,0.842,1.975,19.352
+autotheta,1.123,1.018,20.235
+seasonalnaive,1.0,1.0,20.418

results/leaderboards/BOOM_metric_type_leaderboard.csv

@@ -16,9 +16,9 @@ timer,0.663,0.89,0.721,0.864,0.662,0.608,0.658,0.63
 timesfm_2_0_500m,0.919,0.745,0.706,0.726,0.403,0.44,0.466,0.431
 timesfm_2_5_200m,0.71,0.672,0.591,0.659,0.325,0.384,0.386,0.382
 Toto-Open-Base-1.0,0.687,0.658,0.583,0.634,0.317,0.382,0.382,0.369
-toto_2.0_1b,0.633,0.627,0.541,0.608,0.289,0.359,0.353,0.351
-toto_2.0_2.5b,0.639,0.622,0.541,0.607,0.303,0.36,0.355,0.351
-toto_2.0_22m,0.67,0.647,0.56,0.626,0.313,0.373,0.367,0.362
-toto_2.0_313m,0.631,0.632,0.546,0.61,0.281,0.364,0.354,0.353
+toto_2.0_1b,0.634,0.626,0.541,0.609,0.279,0.358,0.35,0.35
+toto_2.0_2.5b,0.626,0.625,0.54,0.609,0.278,0.36,0.349,0.351
+toto_2.0_22m,0.67,0.647,0.56,0.626,0.303,0.373,0.364,0.362
+toto_2.0_313m,0.626,0.629,0.545,0.61,0.281,0.361,0.353,0.351
 toto_2.0_4m,0.704,0.664,0.584,0.65,0.313,0.38,0.379,0.38
 visionts,1.199,1.03,0.925,1.032,0.591,0.672,0.674,0.68

results/leaderboards/BOOM_term_leaderboard.csv

@@ -16,9 +16,9 @@ timer,0.809,0.804,0.779,0.661,0.671,0.597
 timesfm_2_0_500m,0.817,0.78,0.619,0.522,0.499,0.359
 timesfm_2_5_200m,0.7,0.67,0.554,0.424,0.411,0.331
 Toto-Open-Base-1.0,0.688,0.657,0.535,0.424,0.406,0.318
-toto_2.0_1b,0.654,0.618,0.502,0.399,0.379,0.296
-toto_2.0_2.5b,0.653,0.618,0.501,0.402,0.382,0.297
-toto_2.0_22m,0.671,0.642,0.52,0.413,0.395,0.308
-toto_2.0_313m,0.657,0.622,0.507,0.399,0.383,0.3
-toto_2.0_4m,0.695,0.663,0.543,0.426,0.409,0.32
+toto_2.0_1b,0.653,0.617,0.504,0.393,0.376,0.298
+toto_2.0_2.5b,0.65,0.616,0.505,0.393,0.377,0.298
+toto_2.0_22m,0.671,0.642,0.52,0.408,0.393,0.308
+toto_2.0_313m,0.656,0.621,0.506,0.397,0.379,0.299
+toto_2.0_4m,0.695,0.663,0.543,0.425,0.409,0.32
 visionts,1.028,1.006,0.943,0.698,0.694,0.637