tradefloor: a simulated stock market for testing trading strategies and AI agents

Your first market in Python

The script runs thirty trading days of five generated companies and scores buy-and-hold and momentum on two copies of one market. Then it rebuilds the market from its inputs and checks that it matches the first run. The whole script takes about a second.

import tradefloor as tf

# five generated companies in one simulated market
universe = tf.Universe.random(5, seed=7)
engine = tf.Engine(seed=776644, universe=universe)
engine.run_days(30)  # thirty trading days

# two strategies, each on its own copy of that market.
# Their orders fill against the book and move the price.
cards = tf.evaluate({"hold": tf.baselines.BuyAndHold(),
                     "momentum": tf.baselines.Momentum()},
                    seed=776644, universe=universe, days=30)
for name, card in cards.items():
    print(f"{name:9} pnl {card.pnl:>9,.0f}",
          f"impact {card.impact_bps:+.2f} bp")

# the same inputs rebuild the same market, bit for bit
again = tf.Engine(seed=776644, universe=universe)
again.run_days(30)
print(again.state_hash() == engine.state_hash())

hold      pnl       574 impact +0.09 bp
momentum  pnl  -103,664 impact +1.12 bp
True

Continue in the quickstart

The trouble with backtests

A backtest tells you how a strategy would have done on prices that already happened. It can't tell you what happens when your orders are big enough to move the price, or how the strategy does in a thousand other years. If an AI agent is making the trades, its model may already have seen that history in training.

Common workflows

• Test a strategy

  Write a strategy as a Python class with an act method. tf.rank runs it beside buy-and-hold on many seeds and counts, seed by seed, which one did better.

  Compare strategies

• Test an AI agent

  Adapters for the OpenAI Agents SDK, PydanticAI, LangGraph and FinRobot pass the model each observation and send its orders. Record its answers and the run replays without calling the model. CrewAI agents connect through the local MCP server.

  LLM adapters

• Fork a market

  Copy a running market, change one input in one copy, such as a rate shock, and run both forward. The copies share their past bit for bit. Any later difference comes from the change.

  Fork a market

• Rerun a result

  A run depends on its companies, seed, preset, package version, any scenario and the orders sent into it. Save them in a manifest and someone else can rebuild the same market on the same release.

  Reproducibility

See it working

• Use Cases

  Six worked examples, one for each role. An AI agent team decides which prompt to ship, a bot builder finds out the bot has no edge before going live, and a risk team learns which crises break its drawdown limit.

  See the use cases

• The Model

  pt-v21, the default model, passes all 40 pre-registered checks against real markets. You can watch its variables react as a liquidity crisis hits and read what it doesn't capture yet.

  See the model

Open source and hosted

• Open source

  Available now

  The Python library and the Rust crate, under MIT or Apache-2.0. It runs on your machine with no account and no usage limits.

  • Markets, forks and packaged scenarios
  • Scoring and ranking across seeds
  • LLM adapters and a local MCP server
  • Reproducible runs and manifests
  • A Gymnasium environment for reinforcement learning
  QuickstartGitHub

• Hosted app

  in beta

  The same simulator on a server, in beta with a limited free plan. Paid plans with larger limits are coming.

  • Saved markets an agent can come back to
  • An MCP endpoint, an HTTP API and an API shaped like Alpaca's
  • Python strategies run in a sandbox
  • Evaluation suites for teams and teaching
  Try the hosted betaPlans

Compare plans