tradefloorStart free

tradefloor compared

There are already good tools for testing trading strategies. This page covers what each one does well, where tradefloor is different, and when you'd be better off with something else.

The projects at a glance

ProjectWhere prices come fromYour orders move pricesFork a running marketLLM agentsChecked against real data
tradefloorA simulated economy and an order book per stockYes, through the order bookYesAdapters for four frameworks, an MCP server, recorded answers40 criteria registered before grading
Backtrader, Zipline, vectorbt, LEANHistorical dataThrough a slippage model applied to recorded pricesNot documentedNot built inThey replay real prices
FinRLHistorical dataFixed transaction costs onlyNot documentedIn separate projectsIt replays real prices
gym-anytradingHistorical dataFixed fees onlyNot documentedNot built inIt replays real prices
ABIDESSimulated agent populations, or a replay of one historical stockYes, through a message-level exchangeNot documentedNot built inScripts for stylized facts
MarSA generative model of order flowYesSeparate rollouts from one starting pointNot built in9 of 11 stylized facts
StockSimHistorical prices, or an order book modeIts README lists latency, slippage and market impactNot documentedA built-in LLM agentNot documented

Backtesters

Backtrader, Zipline-reloaded, vectorbt and QuantConnect LEAN all replay historical prices. Backtrader, Zipline and LEAN step through events one at a time, while vectorbt tests thousands of parameter sets at once as arrays. When your strategy trades, it gets the recorded price adjusted by a slippage model. LEAN's most detailed model estimates impact using the work of Almgren and colleagues, but in every case the rest of the market never sees your order.

If you want to know how a rule would have done on real history, with real tickers and real dates, use a backtester. The same goes if you want to run the same code live, which LEAN supports. tradefloor is for the things history can't tell you: how much your own orders move the price, how a strategy holds up across lots of different markets, what a single change does when everything else stays the same, and how an LLM trades a market it can't have seen in its training data. Plenty of people will end up using both.

Reinforcement-learning environments

FinRL and gym-anytrading wrap historical prices in a Gymnasium environment. FinRL charges a fixed percentage per trade, and gym-anytrading charges fixed fees and only lets the agent buy or sell. FinRL is popular for teaching and research, with over sixteen thousand GitHub stars, while gym-anytrading hasn't had a commit since 2023.

FinRL is the better pick for deep reinforcement learning on real price histories, and it has a big community behind it. Use tradefloor's TradingEnv when you want the policy to pay what the order book would really charge for its trades. Otherwise a policy can learn an edge that only exists because its orders cost nothing.

Market simulators

ABIDES simulates an exchange message by message, modelled on NASDAQ's ITCH and OUCH protocols, with more than a thousand trading agents. JPMorgan released a version with Gym environments, and the repo has scripts that check the output against stylized facts. JPMorgan archived its repo in 2023, and the original hasn't changed since 2020. It's still the one to use for research on market microstructure at the message level.

Microsoft's MarS generates order-level events from a model trained on real order data, and its paper reports that it reproduces 9 of 11 stylized facts on Chinese stocks. It runs on Linux in Docker, and the larger models need GPUs. It's a good fit for research on generative order flow and market impact, if you have the hardware for it.

tradefloor is aimed at something else: testing strategies and agents on a laptop. It installs with pip on Linux, macOS and Windows, a year of trading runs in seconds, and the order books sit on top of an economy with interest rates, earnings and a VIX. The default model is graded on crash replays, rates and trading costs as well as stylized facts, against criteria written down before grading started.

LLM trading simulators

StockSim comes with its own LLM trading agent, split into analyst roles for markets, news and fundamentals, and runs it against historical prices or an order book using OpenAI, Anthropic or Bedrock models. Alejandro Lopez-Lira's llm_trading_sim, from the paper "Can Large Language Models Trade?", puts LLM agents in a synthetic market and has them trade in rounds to test financial theories. Both are built for studying those agents or reproducing that research.

tradefloor is for testing the agent you're building yourself. There are adapters for the OpenAI Agents SDK, PydanticAI, LangGraph and FinRobot, and an MCP server that works with Claude, Cursor and VS Code. Every answer the model gives is recorded. Anyone can then replay the run later without an API key.

Features no other project here documents

  • Forking a running market: copy it on any day, change one thing in one copy, and run both forward from the same past.
  • The same market, bit for bit, on Linux, macOS and Windows, given the same inputs and the same release. Reproducibility has the details.
  • Adapters for the main agent frameworks, with the model's answers recorded so any run can be replayed.
  • A model graded against criteria fixed in advance, with the scripts and results published.

Sources

Each project's own pages, as read on 8 October 2026. If something here is out of date, let us know through the contact page.