Experimentos realizados y sus respectivas ganancias proyectadas

Línea Principal de Experimentos

Modelo seleccionado 8122j con las siguientes características:

  • Nuevas variables: sobre la base de poteciar la señal vinculada al ‘volumen transaccional’ del cliente se crearon variables sobre actividad financiera (prestamos, movimientos y consumo tajetas), af_edad_antiguedad cliente, y evolución temporal de dichas variables (ej.tendencias).
  • Entrenamiento final del modelo sobre trimestre septiembre-octubre 2020, con buenos resultados en los experimentos.
  • Ensamble final de 100 modelos con distintas semillas.

Experimentos Secundarios

Tabla de Estrategia de Entrenamiento y sus respectivas ganancias

filename experimento train.rango.desde train.rango.hasta validate.periodos test.periodos train_final.rango.desde train_final.rango.hasta excluir ganancia gananciaPublic
8122HTg 8122TSg/8122TSg.yml 202007 202009 202010 202011 202009 202011 5 17116000 20.57961
8122HTc 8122TSc/8122TSc.yml 202007 202009 202010 202011 202009 202011 5 16932000 21.41459
8122HTh 8122TSh/8122TSh.yml 202007 202009 202010 202011 202009 202011 5 16778000 21.65959
8122HTj 8122TSj/8122TSj.yml 202007 202009 202010 202011 202009 202011 5 16652000 21.59959
8122HTk 8122TSk/8122TSk.yml 202007 202009 202010 202011 202009 202011 5 16618000 21.29960
8122HTa 8122TSa/8122TSa.yml 202007 202009 202010 202011 202009 202011 6 16582000 20.56961
8121HTb 8121TSb/8121TSb.yml 202007 202009 202010 202011 202001 202011 6 16484000 19.40963
8220HTi 8220TSi/8220TSi.yml 202009 202010 202010 202011 202001 202011 6 16320000 21.38960
8122HTf 8122TSf/8122TSf.yml 202006 202009 202010 202011 202006 202011 5 16118000 17.42967
8220HTh 8220TSh/8220TSh.yml 202008 202009 202010 202011 202001 202011 6 16070000 20.28962
8122HTb 8122TSb/8122TSb.yml 202009 202009 202010 202011 202011 202011 6 16068000 20.95960
8120HTa 8120TSa/8120TSa.yml 202001 202009 202011 202011 202001 202011 6 15522000 21.74459
8220HTg 8220TSg/8220TSg.yml 202007 202008 202010 202011 202001 202011 6 15416000 21.26960
8121HTc 8121TSc/8121TSc.yml 201912 202009 202010 202011 201912 202011 6 15310000 19.35463
8220HTf 8220TSf/8220TSf.yml 202006 202007 202010 202011 202001 202011 NA 15250000 21.46459
8220HTd 8220TSd/8220TSd.yml 202004 202005 202010 202011 202001 202011 6 14818000 19.97462
8122HTd 8122TSd/8122TSd.yml 202003 202009 202010 202011 202003 202011 5 14804000 20.09462
8122HTi 8122TSi/8122TSi.yml 201912 202009 202010 202011 201912 202011 1, 2, 5, 6 14802000 18.95964
8000HTa 8000TSa/8000TSa.yml 202001 202009 202010 202011 202001 202011 6 14412000 19.43963
8121HTa 8121TSa/8121TSa.yml 202001 202009 202010 202011 202001 202011 5, 6 14270000 17.86466
8220HTe 8220TSe/8220TSe.yml 202005 202006 202010 202011 202001 202011 6 14074000 21.49459
8320HTa 8320TSa/8320TSa.yml 202001 202002 202010 202011 202001 202011 6 11254000 18.79964
8101HTa 8101TSa/8101TSa.yml 202001 202009 202011 202011 202001 202011 6 NA NA
8102HTa 8102TSa/8102TSa.yml 202001 202011 NA NA 202001 202011 6 NA NA
8102HTb 8102TSb/8102TSb.yml 202001 202009 202011 202011 202001 202011 6 NA NA
8106HTa 8106TSa/8106TSa.yml 202001 202009 202011 202011 202001 202011 6 NA NA
8115HTa 8115TSa/8115TSa.yml 202001 202009 202011 202011 202001 202011 6 NA NA
8120HTb 8120TSb/8120TSb.yml 202001 202009 202011 202011 202001 202011 6 NA NA
8122HTl 8122TSl/8122TSl.yml 201911 202009 202010 202011 202001 202011 6 NA NA
8200HTa 8200TSa/8200TSa.yml 201912 202011 202010 202011 201912 202011 1, 2, 5, 6 NA NA
8220HTa 8220TSa/8220TSa.yml 202001 202002 202010 202011 202001 202011 6 NA NA
8220HTb 8220TSb/8220TSb.yml 202002 202003 202010 202011 202001 202011 6 NA NA
8220HTc 8220TSc/8220TSc.yml 202003 202004 202010 202011 202001 202011 6 NA NA
8220HTj 8220TSj/8220TSj.yml 202010 202011 202010 202011 202001 202011 6 NA NA
8220HTk 8220TSk/8220TSk.yml 202011 202012 202010 202011 202001 202011 6 NA NA
termonu termonuclear-TSa/termonuclear-TSa.yml 201911 202009 202010 202011 202001 202011 6 NA NA

Tabla de Hiperparámetros de Optimizaciones Bayesianas y sus Ganancias

filename learning_rate1 learning_rate2 feature_fraction1 feature_fraction2 num_leaves1 num_leaves2 min_data_in_leaf1 min_data_in_leaf2 max_depth BO.iterations ganancia gananciaPublic
8122HTg 0.01 0.06 0.10 0.60 16 1824 200 2800 -1 100 17116000 20.57961
8122HTc 0.01 0.30 0.10 1.00 16 2024 0 25000 -1 100 16932000 21.41459
8122HTh 0.01 0.06 0.10 0.60 16 1824 200 2800 6 150 16778000 21.65959
8122HTj NA NA NA NA 10 500 100 2000 -1 10 16652000 21.59959
8122HTk NA NA NA NA 10 500 100 2000 6 10 16618000 21.29960
8122HTa 0.01 0.30 0.10 1.00 16 2024 0 25000 -1 100 16582000 20.56961
8121HTb 0.01 0.30 0.10 1.00 16 2024 0 25000 -1 100 16484000 19.40963
8122HTe 0.01 0.30 0.10 1.00 16 2024 0 50000 -1 200 16406000 19.66463
8220HTi 0.02 0.30 0.10 1.00 16 1024 0 8000 -1 50 16320000 21.38960
8122HTf 0.01 0.30 0.10 1.00 16 2024 0 25000 -1 100 16118000 17.42967
8220HTh 0.02 0.30 0.10 1.00 16 1024 0 8000 -1 50 16070000 20.28962
8122HTb 0.01 0.30 0.10 1.00 16 2024 0 25000 -1 100 16068000 20.95960
8120HTa 0.02 0.30 0.10 1.00 16 1024 0 8000 -1 50 15522000 21.74459
8120HTa 0.02 0.30 0.10 1.00 16 1024 0 8000 -1 50 15522000 21.74459
8220HTg 0.02 0.30 0.10 1.00 16 1024 0 8000 -1 50 15416000 21.26960
8121HTc 0.01 0.30 0.10 1.00 16 2024 0 25000 -1 100 15310000 19.35463
8220HTf 0.02 0.30 0.10 1.00 16 1024 0 8000 -1 50 15250000 21.46459
8220HTd 0.02 0.30 0.10 1.00 16 1024 0 8000 -1 50 14818000 19.97462
8122HTd 0.01 0.30 0.10 1.00 16 2024 0 25000 -1 100 14804000 20.09462
8122HTi 0.01 0.30 0.10 1.00 16 2024 0 4000 -1 100 14802000 18.95964
8000HTa 0.02 0.30 0.10 1.00 16 2524 0 20000 -1 50 14412000 19.43963
8121HTa 0.02 0.30 0.10 1.00 16 1024 0 12000 -1 50 14270000 17.86466
8220HTe 0.02 0.30 0.10 1.00 16 1024 0 8000 -1 50 14074000 21.49459
8320HTa 0.02 0.30 0.10 1.00 16 2524 0 50000 -1 50 11254000 18.79964
8101HTa 0.02 0.30 0.10 1.00 16 3024 1 50000 -1 100 NA NA
8102HTa 0.02 0.30 0.10 1.00 16 1024 1 20000 -1 50 NA NA
8102HTb 0.02 0.30 0.10 1.00 16 1024 1 20000 -1 50 NA NA
8106HTa 0.02 0.30 0.10 1.00 16 1024 1 20000 -1 50 NA NA
8115HTa NA NA 0.91 0.91 NA NA NA NA -1 50 NA NA
8122HTl NA NA NA NA 10 500 100 2000 -1 20 NA NA
8200HTa 0.02 0.30 0.10 1.00 16 2024 0 20000 -1 50 NA NA
8220HTa 0.02 0.30 0.10 1.00 16 1024 0 8000 -1 50 NA NA
8220HTb 0.02 0.30 0.10 1.00 16 1024 0 8000 -1 50 NA NA
8220HTc 0.02 0.30 0.10 1.00 16 1024 0 8000 -1 50 NA NA
8220HTj 0.02 0.30 0.10 1.00 16 1024 0 8000 -1 50 NA NA
8220HTk 0.02 0.30 0.10 1.00 16 1024 0 8000 -1 50 NA NA
termonu NA NA NA NA 10 500 100 2000 -1 10 NA NA
termonu 0.02 0.10 0.20 1.00 NA NA NA NA -1 10 NA NA

Top 10 Modelos según mayor ganancia y sus respectivos hiperparámetros

experimento cols rows learning_rate feature_fraction num_leaves min_data_in_leaf num_iterations estimulos ganancia gananciaPublic
8122HTg 785 49403 0.0213877 0.5904652 1483 1438 530 8924 17116000 20.57961
8122HTc 785 49403 0.0100494 0.3825268 560 1344 965 8988 16932000 21.41459
8122HTh 785 49403 0.0213877 0.5904652 1483 1438 539 8782 16778000 21.65959
8122HTh 785 49403 0.0212066 0.2487032 943 1735 950 9794 16726000 21.65959
8122HTh 785 49403 0.0277666 0.3911781 499 1941 729 10046 16714000 21.65959
8122HTg 785 49403 0.0138663 0.3296068 1769 1765 800 9120 16680000 20.57961
8122HTg 785 49403 0.0242742 0.3422475 535 2240 816 9968 16672000 20.57961
8122HTg 785 49403 0.0217941 0.5326392 1729 1707 595 9022 16658000 20.57961
8122HTj 785 49403 0.0500000 0.5000000 11 927 335 9028 16652000 21.59959
8122HTc 785 49403 0.0123253 0.2935178 95 2204 1262 9782 16618000 21.41459
8122HTk 785 49403 0.0500000 0.5000000 336 116 225 9902 16618000 21.29960
8122HTg 785 49403 0.0162318 0.5713234 457 1003 701 8710 16610000 20.57961
8122HTa 785 49403 0.0230265 0.3589839 16 1811 941 8978 16582000 20.56961
8122HTh 785 49403 0.0123953 0.1284983 1163 2603 1864 9460 16580000 21.65959
8122HTg 785 49403 0.0215486 0.3418625 250 2147 741 9840 16560000 20.57961
8122HTh 785 49403 0.0302698 0.2282204 244 1510 430 8294 16546000 21.65959
8122HTk 785 49403 0.0500000 0.5000000 11 927 283 9738 16542000 21.29960
8122HTg 785 49403 0.0351382 0.1680806 1193 972 272 8668 16532000 20.57961
8122HTc 785 49403 0.0100197 0.3627411 204 1549 1737 9524 16516000 21.41459
8122HTh 785 49403 0.0441813 0.5784702 966 2233 528 10612 16508000 21.65959

Gráficos de Features según su ganancia en las Optimizaciones Bayesianas

Gráficos de Features según su ganancia en las Optimizaciones Bayesianas

Experimentos en datos mensuales para determinar ventana temporal de entrenamiento

ANEXO: Tabla para consulta de parámetros por experiemento