Methods and systems for low latency generation and distribution of hidden liquidity size estimates
Systems and methods are disclosed herein that compute trading signals with low latency and high throughput using highly parallelized compute resources such as integrated circuits, reconfigurable logic devices, graphics processor units (GPUs), multi-core general purpose processors, and/or chip multi-processors (CMPs). For example, a liquidity estimation that estimates an amount of hidden liquidity for a financial instrument can be generated from streaming financial market data.
1. A field programmable gate array (FPGA) or graphics processor unit (GPU) for accelerated processing of streaming financial market data that pertains to a plurality of financial instruments to derive trading signals at low latency, the FPGA or GPU comprising:
a plurality of feature compute stage circuits arranged in parallel to define a plurality of parallel paths within the FPGA or GPU, wherein the parallel feature compute stage circuits comprise parallelized hardware logic and state memory that are configured to compute a plurality of features of the streaming financial market data in parallel; and
a combine stage circuit connected to the feature compute stage circuits, wherein the combine stage circuit comprises parallelized hardware logic configured to compute a hidden liquidity size estimation based on a weighted combination of the computed features, wherein the hidden liquidity size estimation represents an estimated size of a hidden order for a financial instrument, wherein each computed feature has a corresponding weight for the weighted combination;
wherein the computed features comprise features that are correlated to and predictive of the estimated size for the hidden order; and
wherein a selection of which features to use for the correlated and predictive features and what values to use for the corresponding weights are derived from a supervised machine learning model.
2. The FPGA or GPU of claim 1 wherein the FPGA or GPU is part of a system that further comprising a processor configured to train the supervised machine learning model based on labeled training data representing known sizes for hidden orders to identify the correlated and predictive features and their corresponding weights.
3. The FPGA or GPU of claim 1 wherein the computed hidden liquidity size estimation comprises a probability that the hidden order has a size above or below a defined threshold.
4. The FPGA or GPU of claim 3 wherein the defined threshold comprises a partition within a range of different sizes for the hidden order.
5. The FPGA or GPU of claim 3 wherein the defined threshold comprises a value within a range of sizes corresponding to 100 to 500 shares of the hidden order.
6. The FPGA or GPU of claim 1 wherein the estimated size comprises an estimated number of shares for the hidden order multiplied by a price for the hidden order.
7. The FPGA or GPU of claim 1 further comprising parallelized circuitry configured to (1) process the streaming financial market data through parallelized processing logic to compute a plurality of additional features of the streaming financial market data and (2) detect a hidden order for a financial instrument based on the additional features, and wherein the hidden liquidity estimation is computed for the detected hidden order.
8. The FPGA or GPU of claim 1 wherein the features comprise features that represent (i) a cumulative volume of hidden orders for the financial instrument or a market and/or (ii) a volume of orders executed as part of a currently detected hidden order for the financial instrument.
9. The FPGA or GPU of claim 1 wherein the streaming financial market data comprises a plurality of messages that pertain to financial instruments, wherein each of a plurality of the messages comprises a plurality of fields of financial market data, and wherein the FPGA or GPU further comprises a plurality of field select circuits that are arranged in parallel upstream from the feature compute stage circuits, each field select circuit being connected to a corresponding feature compute stage circuit and configured to select one or more fields of the financial market data for processing by the corresponding feature compute stage circuit connected thereto to support the feature computations thereby.
10. The FPGA or GPU of claim 9 wherein the feature compute stage circuits and the combine stage circuit are further configured to evaluate the messages for the hidden liquidity size estimations on a message-specific basis.
11. The FPGA or GPU of claim 10 further comprising logic configured to append a subject message with the hidden liquidity size estimation that has been computed for the subject message.
12. The FPGA or GPU of claim 9 further comprising logic configured to normalize the financial market data, and wherein the feature compute stage circuits are further configured to operate on fields of the normalized financial market data.
13. The FPGA or GPU of claim 1 wherein the FPGA or GPU is configured to present hidden liquidity size estimations to one or more trading applications synchronously with the financial market data to which the hidden liquidity size estimations pertain.
14. The FPGA or GPU of claim 1 wherein the FPGA or GPU is configured to generate or modify an order based on the hidden liquidity size estimation.
15. The FPGA or GPU of claim 1 wherein the hidden order comprises a reserve order.
16. The FPGA or GPU of claim 1 wherein the feature compute stage circuits and the combine stage circuit are defined by gate level logic and interconnections therebetween that are resident on the FPGA.
17. The FPGA or GPU of claim 1 wherein the feature compute stage circuits and the combine stage circuit are deployed on the GPU as a plurality of interconnected compute cores of the GPU.
18. The FPGA or GPU of claim 9 wherein the FPGA or GPU is part of a system that further comprises a host that supports processing of the streaming financial market data, wherein the FPGA or GPU offloads computations of the features and the hidden liquidity size estimation from the host.
19. The FPGA or GPU of claim 1 wherein the streaming financial market data comprises real-time financial market data, and wherein the FPGA or GPU is configured to compute the hidden liquidity size estimation in real-time.
20. The FPGA or GPU of claim 19 wherein the FPGA or GPU is further configured to normalize the financial market data and deliver the real-time hidden liquidity size estimation in synchronicity with the normalized financial market data.
21. The FPGA or GPU of claim 19 wherein the FPGA or GPU computes the real-time hidden liquidity size estimation without impeding delivery of the real-time financial market data.
22. The FPGA or GPU of claim 19 wherein the FPGA or GPU is further configured to (1) normalize the real-time financial market data and (2) deliver the normalized real-time financial market data and the real-time hidden liquidity size estimation with single-digit microsecond latency.
23. The FPGA or GPU of claim 1 wherein the FPGA or GPU is part of a dedicated signal plant that operates on streaming financial market data.
24. The FPGA or GPU of claim 1 wherein the FPGA or GPU is part of a ticker plant that consumes, normalizes, aggregates, and distributes streaming financial market data.
25. The FPGA or GPU of claim 1 wherein the FPGA or GPU is part of a trading application server.
26. A method for accelerated processing of streaming financial market data to derive trading signals at low latency, the method comprising:
streaming financial market data through a field programmable gate array (FPGA) or a graphics processor unit (GPU), wherein the streaming financial market data pertains to a plurality of financial instruments, wherein the FPGA or GPU comprises (1) a plurality of feature compute stage circuits arranged in parallel to define a plurality of parallel paths within the FPGA or GPU and (2) a combine stage circuit connected to the feature compute stage circuits, wherein each feature compute stage circuit comprises parallelized hardware logic and state memory, and wherein the combine stage circuit comprises parallelized hardware logic;
the feature compute stage circuits computing a plurality of features of the streaming financial market data in parallel; and
the combine stage circuit computing a hidden liquidity size estimation based on a weighted combination of the computed features, wherein the hidden liquidity size estimation represents an estimated size of a hidden order for a financial instrument;
wherein the computed features comprise features that are correlated to and predictive of the estimated size for the hidden order; and
wherein a selection of which features to use for the correlated and predictive features and what values to use for the corresponding weights are derived from a supervised machine learning model.