London5 Research

Models that understand markets. Small enough to run on your desk.

London5 Research is the investigation team behind the London5 agent. We build the models the product runs on: models that (1) learn the structure of a market rather than its noise, (2) hold specialised knowledge of trading and financial advice, (3) run on the trader's own machine, and (4) earn their place in trading simulations before they reach a real desk.

Representation space · market stateLondon5 Research
s₁ s₂ QUIET TREND STRESS sₜ ŝₜ₊ₖ market state trajectory predicted representation
What we work on

Two lines of research. One product they feed.

Our main goal is to build models that understand markets. Market data is continuous, high-dimensional and noisy — prices, order books, flow, news — and much of what happens next is not predictable from what came before. A model trained to predict the next price is trained to fit that noise: it does well on the past and poorly on the desk. Our work follows from that.

JEPA · Joint-Embedding Predictive Architectures for markets

Representations, not prices.

We are developing world models of market state: architectures that learn abstract representations of a market, ignore the unpredictable detail, and make their predictions in representation space. What the agent receives is not a number to trust but a state it can reason and plan against — regime, pressure, the shape of the book.

market window t − w … t market window t … t + k encoder encoder sₓ sᵧ predictor ŝᵧ loss in embedding space never “next price”
Data
The market history the engine already records for its own desks, and replays from the London5 simulator.
Into the product
A context signal the agent's reasoning subscribes to, once a representation has proved itself on the desk.
Status
Exploratory

SLM · Small language models with trading knowledge

Knowledge that runs on the desk.

Trading and financial advice is specialised knowledge — scarce in the text large models are trained on, and the frontier models that hold some of it run in someone else's data centre. We start from open-weight models in the 20B-to-70B class — Qwen 3.8 27B first — and train them further on that knowledge, from two sources that have it in the right shape: the Londron network, where real agents trade on a real feed with outcomes on the record, and synthetic desks run in the London5 simulator, where the answer is known.

Londron network feed · agents · outcomes synthetic desks simulator · answer known SLM on an open 20B–70B base trading simulation same desk, same rules base · trained
Base
Open-weight models in the 20B–70B class, benchmarked in the simulator first; Qwen 3.8 27B is the first base.
Data
The Londron network — the feed, the agents and their journals, outcomes on the record. Synthetic data — simulator desks with the answer known, as many scenarios as the training needs.
Into the product
Weights the app runs locally, taking the general model's place on the desk where they grade better.
Status
Data pipeline design
How we compare

We compare how models behave, and report where they diverge.

A score says how well a model did. It does not say what it did. We put every model through the same three subjects — a trading desk, an investor's question, a research brief — and compare the behaviour, not only the result: the open bases in the 20B-to-70B class beside the very large models at the frontier. Where they converge, a model on a trader's own machine already knows enough. Where they diverge, the research goes next. We compile the comparisons into reports by subject; the trading-simulation grade is one of the scores in them, not the whole of them.

Trading

The same simulated desk — instruments, data windows, fees, risk limits. Compared: every decision and the reasoning given for it. Scored by the simulator — valid, grounded, defensible, aligned, disciplined — out of 100.

Investor advice

The same investor, the same portfolio, the same constraints, the same question. Compared: the advice given, what it weighed, and what it left out.

Research

The same brief — a company, a sector, an event. Compared: what each model found, what it cited, and what it concluded.

Open models · 20B–70B

20B27B32B70B

Open weights. Run on the trader's own machine. Candidates for the desk.

Very large models

trillion-scale · open weightsfrontier APIs

The yardstick, not a candidate.

The same three subjects

  • Trading
  • Investor advice
  • Research

Same inputs to every model. Behaviour compared, not only results.

Converge

The base already knows enough on that point.

Diverge

Where the research goes next.

Report · by subject schematic

Tradingdiverge
open
very large
Investor adviceconverge
open
very large
Researchdiverge
open
very large

Scores side by side per subject — for trading, the simulator's grade — and where they converge or diverge. Shape only: no result is published yet.