Optiver researcher Karun Rao’s October 6 essay explains why market-data deep learning resists a simple language-model transplant. Best quotes, aggregated depth and individual orders contain different information. Two books can show identical depth while concealing different queue positions and cancellation histories. Event timing is irregular, signals are weak, and a long token window can still represent a very short trading interval. The essay describes a research problem without publishing a live profitability result. Read Rao’s practitioner account.
The engineering complement is an Optiver/NVIDIA discussion of accelerated computing. Model weights can be loaded before an event arrives; memory movement and physical placement still affect response time. For research, the valuable improvement may be how quickly a researcher reaches a sound result across the whole pipeline. Read the hardware discussion. For the earlier research lineage, DeepLOB combined convolutional book structure with recurrent temporal modelling; its historical prediction tests establish neither today’s architecture nor today’s tradable alpha.