Narrow or general?
When does a population of specialized models beat a general one, and what does coordination cost?
AI RESEARCH LAB / QUANTITATIVE TRADING
Collectiveintelligence.Models that learn to coordinate.
MASSTEC is an AI research lab. We train populations of specialized models, evaluate them against each other in multi-agent market simulation, and apply them to quantitative trading.
Enter the fieldTRAIN then EVALUATE then DEPLOY then ORCHESTRATE
THE FIELD
ILLUSTRATIVE · NOT TO SCALE
ILLUSTRATIVE · NOT TO SCALE
01 / RESEARCHMASS, NOT MONOLITHS
Every price is
many decisions,
meeting at once.
Markets are multi-agent systems: non-stationary, adversarial, and shaped by the agents acting in them. We think the intelligence that operates there should be collective too. Populations of specialized models, trained with multi-agent reinforcement learning, rather than a monolithic one.
See the method
Together, not identical.
How do independently trained agents learn to act together without converging on the same view?
A target that moves.
How do you evaluate a policy in a market that reacts to it, and changes because of it?
FOUNDATIONS
The work we build on, from swarm intelligence to multi-agent reinforcement learning.
- 1987Flocks, herds and schools: a distributed behavioral modelGlobal order from local rules.
- 1995Particle swarm optimizationA population searches better than any one member.
- 1996Ant system: optimization by a colony of cooperating agentsCoordination through a shared environment.
- 2017Population based training of neural networksTrain populations, not individuals.
- 2019ABIDES: towards high-fidelity market simulation for AI researchMarkets as agent-based simulations.
- 2019Grandmaster level in StarCraft II using multi-agent reinforcement learningA league of agents, each improving against the others.
- 2020Emergent tool use from multi-agent autocurriculaCompetition as curriculum.
02 / THE METHODFrom training
SCROLL, OR SELECT A STAGEFrom training
to inference.
ENV 01 · TRAINING01 / 04
Specialized agents are trained independently, each on a narrow objective.
- OFFLINE
- SIMULATED
- LIVE
03 / INFRASTRUCTUREHYPOTHETICAL UNTIL LIVE
Four environments.
Three gates.
Every swarm passes through four environments in order. Everything before live inference is hypothetical.
- ENV 01
Training
Where policies learn.
Agents are trained independently with reinforcement learning on narrow objectives, then composed into swarms around a mandate.
- STATUS
- Research
- LEAVES AS
- A composed swarm
- ENV 02
Offline evaluation
Held-out data.
Swarms are scored on market data excluded from training. Offline data cannot react, so this measures generalization, not market impact.
- STATUS
- Hypothetical
- ENTRY GATE
- Offline
- ENV 03
Simulation
A market that reacts.
A forecast assumes the market ignores you; a simulation lets it answer back. In a multi-agent market simulator the swarm's orders meet other agents and move prices. Simulated capital; nothing is at risk.
- STATUS
- Hypothetical
- ENTRY GATE
- Simulated
- ENV 04
Live inference
Production.
Swarms that pass both gates run in live inference. We coordinate them across the book, each kept distinct.
- STATUS
- Live
- ENTRY GATE
- Live
OFFLINE · SIMULATED · LIVE
Results from offline evaluation and simulation are hypothetical and differ from live trading.
04 / BRIEFINGThe method,
The method,
on one page.
How we train agents, what each environment tests, and where hypothetical ends and live begins.
