Mar 2, 2018 · 23m · mad
Text Analytics for Finance // Amanda Stent, Bloomberg (FirstMark's Data Driven)
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
At a Data Driven NYC event, Bloomberg NLP Researcher Amanda Stent explores how text analytics and natural language processing are applied to unstructured financial data to generate actionable market insights. She outlines core application areas, real-time speed requirements, high-precision entity extraction, and the importance of hybrid human-in-the-loop workflows.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Matt holds 3.9% of the talking time here. How this is scored →
Matt as informed peer 0.9 Guest teaching 1.0 Guest disagreement 0.6 Matt pushing back 0.1
speaking balance: gold is Matt, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 14:09 Clarifying Bloomberg's infrastructure architecture
Stent politely rejects the host's premise of a single data pipeline, explaining that Bloomberg runs 15,000 to 40,000 separate functional pipelines.
Hardest push from Matt ▶ 12:52 Host pushes on practical human-in-the-loop executionMatt Turck anchors his question in his own four-year experience at Bloomberg and presses for practical details on how human-machine handoffs actually function.
Biggest teaching moment ▶ 14:09 Educating on multi-function pipeline complexityStent corrects the host's impression of backend architecture by breaking down how thousands of feeds feed tens of thousands of individual functions.
Matt holds his own ▶ 12:26 Host establishes domain authority and Bloomberg historyMatt Turck highlights Bloomberg's foundational tech heritage and mentions his four-year tenure at the company to frame his line of questioning.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Matt as informed peer | Guest teaching | Guest disagreement | Matt pushing back | Why |
|---|---|---|---|---|---|---|
| Four Key NLP Applications at Bloomberg | 0 | 0 | 0 | 0 | Monologue presentation segment delivered by Amanda Stent introducing Bloomberg's core mission and NLP enrichment applications. The host is not active during this segment, requiring host-side scores to be zero. | |
| Three Core Challenges in Financial Text Analytics | 0 | 0 | 0 | 0 | Monologue presentation detailing idiomatic financial language and document understanding challenges. The host is inactive throughout this portion. | |
| The Need for Speed and High-Volume Social Processing | 0 | 0 | 0 | 0 | Monologue segment focusing on latency, processing speed, and tweet funnels. The host remains silent throughout the talk. | |
| Precision vs. Recall and Ethical Data Science | 0 | 0 | 0 | 0 | Monologue presentation on precision versus recall, ethical data science, and human-in-the-loop workflows. Host activity is zero. | |
| Presentation Wrap-Up and Transition to Q&A | 4 | 3 | 2 | 1 | Matt Turck steps in, citing his four-year background at Bloomberg and asking about human-in-the-loop mechanics and data pipelines. Stent gently reframes his question about a single data pipeline, clarifying that Bloomberg operates between 15,000 and 40,000 distinct functions. | |
| Audience Q&A: Social Media Manipulation, Speed Limits, and Market Truth | 1 | 2 | 1 | 0 | Audience members ask detailed questions on bot manipulation, speed calculus, and market expectations. The host acts only as a moderator, while Stent provides detailed technical and philosophical answers. | |
| Audience Q&A: Competitor Usage of Automated Content | 1 | 2 | 1 | 0 | Audience questions cover competitor usage of automated text and sentiment extraction from earnings calls. Stent shares informative responses while the host maintains a minimal moderating role. |
Statements from this episode (14)
Stent: One-minute news reporting delay corresponded to 3% stock price move
“The distance between 10 38 and 10 39 Was about three percent in the stock price. The distance between the SEC headline and 1055, 20 minutes later, was a 12% absolute difference in the stock price.”
Stent: SEC declared Twitter valid for official corporate disclosures
“SEC has declared that Twitter, in particular, can be used to provide company, official company announcements.”
Stent: Bloomberg NLP aims to inform trading decisions, not track sentiment
“At Bloomberg, we're not trying to identify net promoter score. We're trying to really understand what's going on so that someone else can decide whether they should buy or sell this company.”
Stent: Vertical NLP faces key hurdles in domain language, speed, and precision
“Here are three key challenges that we face that would be present to a greater or lesser extent in any other highly highly vertical industry. The first is financial language, idiomatic language. The second is the need for speed. Remember, I said a minute. And t…”
Stent: No generic text analytics platform can handle financial language
“There is no on the market generic text analytics platform that can handle this type of language. No way.”
Bloomberg can automatically generate news articles without humans in two minutes
“And this kind of article we can now automatically generate. So you can get this with no human intervention within two minutes, and then within five minutes a human journalist can come along in art color.”
Bloomberg NLP models feed over 100 Bloomberg Terminal screens
“And they feed over a hundred different screens within the Bloomberg terminal, which provide company trend information, company tweet information, sentiment information, sentiment alerts, tweets on mergers, tweets by the president of the United States, et ceter…”
Generic NLP mistakenly linked Bill O'Reilly news to O'Reilly Auto Parts
“Okay, so Bill O'Reilly said something, and there's a small and innocent company called O'Reilly Auto Parts somewhere in the Midwest, and a bunch of generic text analytics engines linked Bill O'Reilly to O'Reilly Auto Parts, and the O'Reilly Auto Parts stock dr…”
Stent: Bloomberg shifted data labeling to deep learning and decision trees
“And that's something that historically has been done mostly with rule-based systems, but today we do it with deep learning and a lot of decision trees.”
Stent: Bloomberg Terminal contains between 15,000 and 40,000 functions
“There are, nobody has told me the actual number. I've heard numbers from between 15 and 40,000 individual functions inside the Bloomberg. Each of them is fed by at least one data feed, and in many cases, several.”
Bloomberg filters social media manipulation using account metadata and supply chain data
“We do a lot of historical monitoring, so we can know how old a social media account is. We also have a very, very large collection of pieces of information about entities in the world, so we can know how closely associated a particular handle is. With a decisi…”
Stent: Bloomberg's human-written stories are much more informative than automated ones
“Our manually generated stories, which are frankly much more informative”