Table 1.
Prediction performance using all 132 features. Random under-sampling of the majority class (low volatility) was applied and repeated 5 times to make class sizes equal in the training dataset.
| Performance measure, subsamples | Logistic regression (ridge), n (%) | Logistic regression (LASSOa), n (%) | Random forests, n (%) | SVMb, n (%) | |
| Accuracy (low volatility class; n=694) | |||||
|
|
Subsample 1 | 476 (68.6) | 513 (73.9) | 473 (68.2) | 437 (63.0) |
|
|
Subsample 2 | 476 (68.6) | 520 (74.9) | 482 (69.5) | 453 (65.3) |
|
|
Subsample 3 | 492 (70.9) | 514 (74.1) | 499 (71.9) | 465 (67.0) |
|
|
Subsample 4 | 499 (71.9) | 509 (73.3) | 491 (70.7) | 458 (66.0) |
|
|
Subsample 5 | 495 (71.3) | 512 (73.8) | 471 (67.9) | 469 (67.6) |
| Accuracy (high volatility class; n=185) | |||||
|
|
Subsample 1 | 122 (65.9) | 119 (64.3) | 129 (69.7) | 120 (64.9) |
|
|
Subsample 2 | 112 (60.5) | 115 (62.2) | 126 (68.1) | 118 (65.3) |
|
|
Subsample 3 | 117 (63.2) | 115 (62.2) | 127 (68.6) | 117 (63.2) |
|
|
Subsample 4 | 121 (65.4) | 117 (63.2) | 128 (69.2) | 116 (62.7) |
|
|
Subsample 5 | 124 (67.0) | 121 (65.4) | 129 (69.7) | 115 (62.2) |
| Overall accuracy (n=879) | |||||
|
|
Subsample 1 | 598 (68.0) | 632 (71.9) | 602 (68.5) | 557 (63.4) |
|
|
Subsample 2 | 588 (66.9) | 635 (72.2) | 608 (69.2) | 571 (65.0) |
|
|
Subsample 3 | 609 (69.3) | 629 (71.6) | 626 (71.2) | 582 (66.21) |
|
|
Subsample 4 | 620 (70.5) | 626 (71.2) | 619 (70.4) | 574 (65.3) |
|
|
Subsample 5 | 619 (70.4) | 633 (72.0) | 600 (68.3) | 584 (66.4) |
aLASSO: least absolute shrinkage and selection operator.
bSVM: support vector machines.