Customer-trained Large Event Models
Help your business make more money by predicting the probabilities of future events more accurately. Hayek brings Large Event Model (LEM) technology to your own decisions, data and outcomes—not just the information available to everyone else.
From forecasts to decisions
Pricing, inventory, underwriting and trading decisions rarely have one certain outcome. A single forecast hides the range of possibilities—and the expensive tails. Hayek is designed to model the event itself, using the proposed action and the information available when the decision is made.
A language model can write a percentage. Ordinary next-token training does not, by itself, ensure that the number matches how often the event occurs. Hayek focuses on outcome probabilities that can be evaluated against resolved events.
A forecast of an aggregate trend does not necessarily tell you what changes when you choose a different price, quote or order size. Hayek conditions on the decision context and candidate action, where the available data supports learning that relationship.
Your recorded events + decision context → Hayek outcome probabilities → Your objectives and decision policy
Combine the probability of each outcome with its value or cost to compare expected economics. Your business retains the objectives, constraints and final decision; Hayek supplies the predictive layer.
Your history is the advantage
A shared vendor feed is available to other subscribers. A customer-trained model can learn from your own record of decisions and outcomes. Rather than starting an internal research programme from zero, you start with DeepMM’s LEM architecture and adapt it to your business.
RFQs sent and received, quotes shown, covers, passes, fills and traded-away inquiries. Add counterparty history, axes, inventory, marks and post-trade drift where available. The opportunities you did not trade carry information too.
Timestamped records of prices offered, orders, demand, replenishment, underwriting decisions or other recurring actions, together with what subsequently happened. The model’s scope follows the data and the decision you choose.
Only information knowable at the decision time belongs in the prediction inputs. Outcomes provide the learning and evaluation targets—not a shortcut into the forecast.
One technology, different decisions
Estimate fill likelihood at candidate quotes where RFQ and outcome data supports it. Use price uncertainty and customer context as inputs to your existing optimizer. The desk keeps its objective function, risk limits and final quote.
Use conditional distributions to inform position sizing, hold or hedge decisions and inventory allocation. Route uncertain cases to human review rather than treating every forecast as equally reliable.
Evaluate pricing and discount choices, inventory buffers, procurement, capacity or underwriting decisions against their possible outcomes and economics. Repeated small improvements can matter at business scale.
The credit offering describes nineteen percentiles, from the 5th through the 95th, conditioned on relevant transaction context such as side, size and counterparty. Output definitions and calibration are agreed and validated for the use case.
A trade-price percentile is not automatically an RFQ fill probability. Fill models require suitable quote and outcome observations and their own validation. Economic benefits depend on implementation, costs and constraints; they are not guaranteed.
Private by design
The proposed deployment runs training and inference in your own cloud account or infrastructure. Your proprietary records remain within your environment rather than being handed over as a dataset for DeepMM to retain.
The customer-trained weights are your asset—not a shared feed offered to the desk across the street. Data access, model ownership, security and operating responsibilities are specified in the engagement.
You begin with an established modelling foundation, then adapt the data pipeline, training and evaluation to your event history. The integration is scoped around one decision first, rather than an open-ended promise to transform every workflow at once.
The commercial structure starts with a deployment and initial-training engagement, followed by ongoing model use under agreed terms. Scope and pricing are discussed after reviewing the use case and data.
Prove it on your business
Asymmetric costs are not required, but they can make better probability estimates especially valuable. If the record cannot support a credible evaluation, establish that before committing to a build.
Start with a short introduction, then a working session with the decision owner and data team. Scope a paid prototype on a defined historical dataset, reserve held-out outcomes and compare with your current approach. Measure probability calibration and decision economics—not forecast accuracy alone. You keep the scorecard whichever way the evaluation lands.
AXOR provides evidence of DeepMM’s underlying technology in corporate-bond pricing. That is not a measured performance result for a newly trained Hayek model or a guarantee of realized P&L. Explore AXOR’s published evaluation and caveats →
Choose the right product
A scoped deployment trained on proprietary event history, within your infrastructure, to support the decisions your business owns.
A separate platform for defining outcomes, exploring forecasts and comparing the expected value of actions. It is not the same engagement as deploying a customer-trained Hayek model in your environment.
Explore LEMs.ai →Transaction-conditioned corporate-bond price, spread and yield distributions delivered through an API. Use the market intelligence without commissioning a customer-trained model.
Explore AXOR →Algorithmic trading that brings predictions, quote policy, risk controls and approved execution together under the dealer’s mandate.
Explore Mises →Deployment scope, integrations, ownership and commercial terms are agreed per engagement. Availability of individual predictions depends on the data and validation. Better model accuracy does not by itself establish realized profits.