MVTRX

Building realistic training datasets for market microstructure models requires either access to expensive exchange data or synthetic data from oversimplified generators. Existing approaches lack the diversity of market conditions and agent behaviors needed to train models that generalize to real trading. Additionally, training AI models across distributed contributors without trusted infrastructure requires coordination, security, and fair scoring that most teams cannot implement themselves.

A decentralized network can solve this by combining agent-based simulation with collaborative model training. Distributed agents compete in a shared simulated market, generating realistic L3 order-book data as a side effect of their trading. Simultaneously, those agents can train a shared model on the data they generate, uploading compressed gradients to personal buckets while validators score contributions fairly and publish the best updates on-chain. This creates both a high-quality dataset and a trained model in one operation, with incentives that reward both trading performance and gradient quality.

MVTRX (SN-79) implements this on Bittensor. Miners run automated trading strategies against a C++ limit-order-book simulator with background agents. Validators score trading performance (Kappa-3 and PnL) across multiple order-book realizations. Miners can optionally enable GenTRX: each round, the agent trains a ~12M-parameter transformer on its assigned slice of order-book data, uploads the compressed gradient to S3, and validators score it against held-out books. The best gradient updates the canonical checkpoint published on-chain. Rewards split ~95% trading / ~5% training by default, with the training pool scaling to active participation.

MVTRX (SN-79) is a Bittensor subnet that combines agent-based market simulation (τaos) with distributed AI model training (GenTRX). Miners act as automated trading agents in a C++ limit-order-book simulation, competing to maximize risk-adjusted performance (Kappa-3 ratio) across multiple simulated order books. Optionally, miners can also participate in GenTRX, which trains a shared ~12M-parameter transformer model on simulation order-book data: each round, miners compute and upload compressed gradients to personal S3 buckets; validators score these against held-out data and publish the best checkpoint on-chain.

Rewards are split ~95% trading / ~5% training by default, with the training allocation scaling with active participation.