Global STAC Live, Fall 2021

• Ahmad Byagowi, Research Scientist, Facebook

To deal with huge data growth, Facebook has been moving to a tightly coupled distributed data infrastructure. But preserving event order across highly dispersed nodes requires a time synchronization solution that is both highly accurate and highly scalable. So Facebook created Time Card, a PCIe card with a multiband GNSS receiver, an onboard miniature atomic clock, and a time engine implemented in FPGA. Then they open-sourced it within the Open Compute Project in an effort to drive costs low enough to allow deployment in massive numbers. Join us as Ahmad explains Time Card and how Facebook uses it to enable faster performance from distributed databases as well as more advanced management of global infrastructure.

• David Cohen, Senior Principal Engineer, Intel Corporation
• Dan Biederman, Senior ASIC & FPGA Design Architect, Intel Corporation

As just described by Ahmad Byagowi of Facebook, Time Card brings highly accurate and scalable time sync to Facebook's global infrastructure. But Facebook open-sourced Time Card on the premise that other industries could benefit. What about financial services? One use case is trade execution, where inexpensive atomic clocks in every node could be very advantageous. But in Intel’s opinion, that’s not where it ends. David will argue that changing the economics of time sync will benefit databases, AI/ML, and other architectures that rely on understanding or controlling a sequence of events—from market data analytics and pricing algorithms to risk management and fraud detection. Dan will go on to explain how application developers can integrate the precise time functionality of Intel’s time cards into their mission-critical operations. If you care about big workloads or fast workloads, you'll want to attend this talk.

• Peter Durkan, CEO, Lucera
• Mike Galime, Software Engineering Manager, Keysight Technologies, Inc.
• Ron Nevo, CTO, cPacket Networks
• Gareth Mason, Senior Innovation and Engineering Lead, Options Technology

As any firm that’s been burned by a production glitch knows, managing operational risk is just as critical as managing market, credit, or liquidity risk. They all threaten the firm’s capital. As a first line of defense, operations staff need system visibility that is immediate, accurate, precise, and comprehensive. However, when markets move in milli-, micro-, or nanoseconds, providing visibility to humans is not enough. Today’s trading demands automated actions in response to problems. Yet many questions stand in the way. How do we best correlate events across systems? How do we automate root-cause analysis? How far can trading algorithms take their use of realtime IT telemetry input? Our panel of operations experts will tackle these questions and others on how to break down the walls between the conservative world of IT and the risk-taking world of trading. Join us to add your questions to the mix. To kick off, some of the panelists will provide a short presentation:

• Peter Nabicht, President, STAC

Peter will present the latest benchmark results for big compute stacks and announce STAC's Inference Benchmark.

• Stefan Zohren, Faculty Member at the Oxford-Man Institute of Quantitative Finance

Using ML to predict prices requires firms to confront a fundamental tension: finding smarter models tends to require more complex modeling techniques, while more complex techniques tend to require substantially more compute (which translates into higher training cost, longer time-to-market, or both). In this talk, Stefan will explain how his team resolved this tension in a recent case. According to Stefan, the team developed a novel multi-horizon forecasting model for limit order books that significantly outperformed existing methods but was quite computationally expensive. After some research, the team overcame the compute bottleneck by moving from traditional hardware to Graphcore's Intelligence Processing Unit. Come to his talk to hear (and ask questions) about the model, the technical implementation, and the numbers.

• John Ashley, General Manager, Financial Services and Technology, NVIDIA

Last fall at STAC, John argued that the future belongs to the trader who leverages the Python ecosystem, including ML/AI tools, to create smarter strategies faster. Part of this was using NLP to provide better contextual intelligence to trading and investing. That was so last year. This fall, John will discuss some of the latest advanced methods the NLP community has developed to use massive compute to create new kinds of natural language understanding (NLU). In particular, he will focus on how financial firms can use NLU to generate synthetic data that can be put to profitable use (and no, he doesn’t mean Reddit posts).

• Inderdeep Singh, Sr. Director Architecture & Product Management, CME Group
• Gary Kazantsev, Head of Quant Technology Strategy, Bloomberg
• James Coomer, Sr. Vice President Products, DDN Storage
• Bob Gaines, Managing Director Financial Services, SambaNova

Machine learning research continues to grab headlines with increasingly powerful achievements. But practitioners know that making ML work on the ground is a very different story. In particular, they sing a constant refrain: "I don't have an ML problem; I have a data problem." Good data pipelines are absolutely essential for ML projects to succeed. But they aren’t easy. Data scientists must leverage new techniques—sometimes themselves even including ML—to ensure data is appropriately cleaned, enriched, and otherwise ready for feature engineering. At the same time, data engineers must ensure constant availability and minimize costs even while data sizes and data demand continue to expand. Join in as our panel of experts discuss how to solve these challenges. To kick off, some of the panelists will provide a short presentation:

• Markus Löning, Data Scientist and sktime contributor

Python's data science ecosystem has been an indisputable boon to quantitative analysts. However, the most common data type that quants deal with—time series—has been a second-class citizen in the Python world. Compatibility issues between time series libraries slow research, while popular ML libraries, such as scikit-learn, focus on non-temporal data. Stepping up to solve these problems is sktime, an open source unified framework for machine learning with time series. As a key contributor to sktime, Markus will walk us through its critical features, such as time series transformation, classification, and forecasting. He'll cover recent updates to the algorithms and pipeline tools that make this possible, as well as how sktime interfaces with existing libraries, such as scikit-learn, statsmodels, and fbprophet. Come to learn more about sktime and ask Markus your questions.