Multiple-input-multiple-output orthogonal frequency-division multiplexing (MIMO-OFDM) wireless systems use link adaptation to exploit the dynamic nature of wireless environments. Link adaptation maximizes throughput while maintaining target reliability by adaptively selecting the modulation order and coding rate. Link adaptation is extremely challenging, however, due to the difficulty in predicting error rates in OFDM with binary convolutional codes, bit interleaving, MIMO processing, and real channel impairments. This paper proposes a new machine-learning framework that exploits past observations of the error rate and the associated channel-state information to predict the best modulation order and coding rate for new realizations of the channel state without modeling the input-output relationship of the wireless transceiver. Our approach is enabled through our new error-rate expression that is only parameterized by postprocessing signal-to-noise ratios (SNRs), ordered over subcarriers and spatial streams. Using ordered SNRs, we propose a low-dimensional feature set that enables machine learning to increase the accuracy of link adaptation. An IEEE 802.11n simulation study validates the application of this machine-learning framework in real channels and demonstrates the improved performance of SNR ordering as it compares with competing link-quality metrics.