Zentos

Zentos: Parallel Strategy Suite

Multi-agent experiment design for ZenTraders.

Overview

ZenTraders is an experimental quantitative trading platform. The core question: can an autonomous agent beat a human-written trading strategy?

Three independent systems run in parallel on shared market data, each deploying to Alpaca paper trading for comparable evaluation.

The Experiment

┌─────────────────────────────────────────────────────┐
│                  Shared Market Data                   │
│          (OHLCV, fundamentals, alt data)              │
└──────┬──────────────────┬──────────────────┬─────────┘
       │                  │                  │
       ▼                  ▼                  ▼
┌──────────────┐  ┌──────────────┐  ┌──────────────────┐
│  Human-Agent │  │    Pure      │  │   Prediction     │
│  Collab      │  │  Automatic   │  │   Market Hedging │
│  Server      │  │  Agent       │  │   Agent          │
│              │  │              │  │                  │
│  Human sets  │  │  Agent tries │  │  Monitors volume │
│  baseline    │  │  to beat it  │  │  acceleration    │
│  strategies  │  │  autonomously│  │  for hedging     │
│              │  │              │  │                  │
│ 64.23.228.132│  │ 64.23.179.43│  │  64.23.179.43    │
└──────┬───────┘  └──────┬───────┘  └────────┬─────────┘
       │                  │                   │
       ▼                  ▼                   ▼
┌─────────────────────────────────────────────────────┐
│              Alpaca Paper Trading                     │
│         Comparable live-market evaluation              │
│              trading-tracker logging                   │
└─────────────────────────────────────────────────────┘

Baseline: Human-Agent Collaboration Server

Human writes core strategy logic in JupyterHub. Agent assists with parameter optimization. Result serves as the performance baseline.

  • Server: 64.23.228.132
  • Tools: JupyterHub, vectorbt, Alpaca SDK

Learn more →

Challenger: Autonomous Execution Agent

Autonomous agent adapted from Microsoft's RD-Agent(Q). Runs 24/7, generating and testing strategies to outperform the human baseline.

  • Server: 64.23.179.43
  • Tools: RD-Agent(Q), Qlib, Papermill

Learn more →

Hedging: Prediction Market Agent

Tracks volume acceleration on Polymarket and Kalshi. Detects informed capital flow ahead of public news and generates hedging signals.

  • Server: 64.23.179.43
  • Tools: Polymarket API, Kalshi API, volume acceleration detector

Learn more →

Design Principles

Alpha comes from humans, scale comes from agents. Core strategy logic must be defined by experienced researchers. Agents test thousands of variations within that framework.

Independent evaluation on the same benchmark. Both human and agent deploy to the same Alpaca paper trading and trading-tracker for apples-to-apples comparison.

Single-machine deployment. Each system runs co-located on one server. No cross-network calls, no serialization overhead.

Anti-overfitting by design. All systems enforce backtest → paper trading → live validation. Walk-forward validation is mandatory.

Every iteration is logged. All hypotheses, parameters, and results are persisted for review and agent learning.

On this page