Turning broker chats into real-time energy market intelligence
By Varsha Balraj and Ryan Lloyd, Commodities and Global Markets Data Analytics
— 6 minute read
How AI-powered quote extraction revealed the structure of March 2’s oil repricing — and how these insights gave Macquarie’s energy trading experts the capacity to better service clients in volatile markets.
Introduction
The first day of the outbreak of conflict in the Middle East on March 2nd this year brought a sharp change in energy market behaviour — broker-chat quote flow surged, accelerated, and narrowed. In the public market, Brent crude closed at $US77, up from $US72 on the prior Friday, with futures up 6.7% on the day, underscoring the scale of the repricing already underway. More importantly, the move was taking place against the prospect of much larger volatility still to come, with oil price scenarios across the market ranging from the $US60s to as high as $US120 per barrel.
At Macquarie, we had been working with Sense Street, an AI scale-up focused on financial markets dialogue, to extract structured quotes from broker chats. The data from March 2 included some important insights for Macquarie’s experienced energy team, offering a unique window into how quote flow behaves when a market is suddenly pricing a genuinely wide range of outcomes.
Commodity broker quotes
A large proportion of commodities broker prices are communicated to traders via business ‘chat’ channels. In busy periods, traders may have many chat windows open, with multiple brokers providing quotes across different products. Presenting this unstructured chat data as a visual ‘blotter’ dashboard increases a trader’s efficiency in finding the best quote for a product, allowing quicker and sharper prices back to sales desks and hence to clients.
Broker chat quotes pose a complex problem due to the use of natural language with data being provided across multiple messages in a conversation, different conventions such as units or currencies used across different products, and implicit knowledge such as products being quoted by a certain broker. Macquarie’s collaboration with Sense Street resulted in a solution which delivered on our accuracy, scale and latency requirements together with the rigorous accuracy and oversight required within a regulated environment (including anomaly detection, a highly skilled annotation team providing “human in the loop” sampling and labelling, reporting, and live feedback.)
How quote flow changed on March 2
Total extracted quotes — individual price indications posted by brokers in business chat channels, each representing a live bid or offer on a specific contract — rose from an average of 29,248 per day over the prior two weeks to 47,037 on March 2, a 61% increase. But the shift was not just in volume — it was also in continuity. The share of seconds with at least one quote rose from 16.8% to 28.5% — a reflection of just how intensely the market was operating. With more quotes flowing through broker channels, gaps between updates compressed naturally: the median wait between quote-bearing seconds halved from 2 seconds to 1 second. March 2 was not just busier; every available second of broker bandwidth was being used.
61 %
+ 70 %
- 24 %
↑ 29,248 → 47,037 / day
↑ 16.8% → 28.5% of day
↓ 311 → 236 contracts
Figure 1. Continuity through the working day. Share of seconds with at least one quote in each 5-minute block (GMT). March 2 ran consistently above the typical day range from open to close, peaking above 70% at the open versus a typical 47%.
Figure 2. How long traders typically waited for the next information update. Distribution of time gaps between quote-bearing seconds (GMT). On March 2, 73.8% of gaps were 1–2 seconds versus 60.6% on a typical day, with longer gaps shrinking sharply.
Figure 3. Relative quote share before vs after by contract. Top 12 contracts with the largest share shifts. AFE saw the largest absolute gain; TDL the largest loss.
Just as importantly, post-event quoting did not simply rise; it narrowed and shifted toward a smaller set of contracts. Even as average daily quote volumes increased, the number of unique contracts quoted per day fell from 311 to 236, indicating that activity concentrated in fewer contracts rather than broadening across the complex. The strongest gains were in contracts such as AFE (FEI propane) and CEY (Mt. Belvieu vs FEI propane spread), while contracts such as TDL (freight, Middle East Gulf to China) and ULD (gasoil crack) lost share even though their absolute quote volumes increased. Together, these shifts show that streaming quote data captures not just rising volatility, but the market’s preferred channels for repricing it.
Taken together, these patterns point to a market which on that day became both more continuous and more selective. The signal was not simply more activity — it was a faster stream of information, concentrated in a smaller set of contracts which traders did not have to interpret manually.
In March alone, Macquarie extracted over 2.5 million quotes across all global energy broker chats, running a Large Language Model (LLM) in production — a scale of live LLM inference that remains rare even by the standards of today’s use of AI within Financial Services.
Technical implementation
A fine-tuned LLM converts broker chat messages into structured quote data in real time. On average, the system captures approximately 90% of oil quotes — the remaining 10% are messages too ambiguous or fragmented even for a human reader to parse reliably — with 98% of extracted quotes being fully correct.
For context, human traders monitoring the same chats manually would typically achieve lower accuracy under time pressure. This structured data gives traders more time to focus on responding to client requests.
Oversight and continuous improvement
Brokered OTC language is never fully stable — shorthand evolves, new instruments appear, conventions shift across desks and time zones. This is a structural property of any machine learning system: unlike deterministic software, which behaves identically given the same input, ML models are calibrated to a distribution of language that shifts over time. Managing that drift is a core part of operating AI in production.
Macquarie’s Commodity Technology and Data Science teams worked with Sense Street to address this through a closed feedback loop. When the system detects drift or low-confidence outputs in production, those examples are pulled into a human annotation queue, reviewed by specialists providing “human in the loop” evaluation, labelled with the correct extraction, and fed back into the next training cycle — turning production inaccuracies into model improvements.
Further oversight is provided by weekly and monthly accuracy and coverage reporting, highlighting trends, issues and remediations, along with real-time trader feedback to the annotation team.
Figure B. Continuous model-improvement loop. Drift detected in production triggers selective sampling, human annotation, dataset refresh, fine-tuning, and evaluation before the updated model redeploys.
This is how Macquarie clients benefit when AI works with expert teams in ways that are measurable, reliable, and built for the pace of real markets.