Tradefloor

tradefloor

tradefloor is a market simulator you can run a strategy against. It has a Rust core and a Python API.

Give it a seed and a list of companies. It runs a market forward: prices, a limit order book, fills, and an economy that moves each day. Your orders match against the book's depth, so your trades move the price.

Real market data can't tell you what would have happened if you had traded differently, or what caused a move. tradefloor can, because it computed every price. You can fork a running market, change one thing in one branch (a rate rise, a liquidity crisis, a different agent), and measure where the two branches came apart. engine.truth() splits each move in the gap between a price and the model's fair value into twelve factors, and engine.explain(ticker, day) breaks down the move in the traded price, two records no historical dataset carries.

Documentation is at https://docs.tradefloor.dev.

Install

pip install tradefloor

There are wheels for Linux, macOS and Windows on CPython 3.11+, and no dependencies. The same engine is a Rust crate (cargo add tradefloor). Optional extras add the MCP server (tradefloor[mcp]), Arrow output (tradefloor[arrow]), the Gymnasium environment (tradefloor[rl]) and one extra per agent framework.

The API may change before 1.0. Model changes ship as new presets, so a market with no agent orders in it replays exactly on its named preset in later releases. tradefloor was called pretium until 0.5.0. Versions up to 0.4.3 still install under that name, and results recorded with them still replay.

A first run

import tradefloor as tf

universe = tf.Universe.random(40, seed=111)

spec = tf.StrategySpec.momentum(lookback_days=1.0, top_k=5)
scores = tf.evaluate({"mine": spec}, seed=7, universe=universe, days=10)

scores["mine"].return_pct            # what it made
scores["mine"].impact_bps            # what its own footprint cost
scores["mine"].strategy_fingerprint  # sha256, cite this
scores["mine"].errors                # each step that raised or was refused
scores["mine"].sharpe                # annualised, from the daily closes
scores["mine"].time_in_market        # share of steps holding a position

That result comes from one random market, so it says as much about the seed as about the strategy. tf.rank runs many seeds and compares strategies with a paired sign test. Add tf.baselines.reference_agents() to the entrants and tf.versus_buy_and_hold(scores) reads each score against buy-and-hold on the same market.

A Python agent is any object with act(obs) that returns orders: a number of shares for a market order, tf.Limit(quantity, price) or tf.Cancel(). It sees a read-only view of the market and its own portfolio, and obs.history holds a daily bar per name. A bar's close is the day's last print. On pt-v21 the close then re-marks every name and the next session opens after an overnight move, so the next day starts away from that print: the close sits 11 bp from it at the median on a 20-name roster, and the open 25 bp from the close. docs/AGENTS.md covers what the view holds, how trades are charged, the framework adapters (OpenAI Agents SDK, PydanticAI, LangGraph, FinRobot) and how scoring works.

The demo

The examples are in this repository and not in the package, so clone it first:

git clone https://github.com/simoncoombes/tradefloor
cd tradefloor
python examples/rate-shock/counterfactual.py

It runs an agent in a controlled market, checkpoints the world and forks it, raises rates by 200 bp in one branch, and compares what the same agent does next. It prints nine checks that the two branches started identical, the step at which the agent's behavior changed, and the two branches side by side. The run takes under five seconds of CPU and needs no keys and no network. The walkthrough is Your first counterfactual experiment.

Contents

Drive it from an agent

pip install "tradefloor[mcp]"
claude mcp add tradefloor -- tradefloor-mcp

tradefloor-mcp speaks MCP over stdio, and tradefloor mcp starts the same server. Strategies, universes and scenarios are data, so a tool argument cannot reach code. Each result carries its own caveats. A session keeps one market in the server between calls, so an agent can place orders a step at a time, fork the market to try two actions, and rewind. See the MCP page.

Scenarios

engine = tf.Engine(seed=42, universe=universe)
engine.run_days(20)                                # a shared history first
scenario = tf.Scenario.load("liquidity_crisis")    # ships with the package

control, stress = tf.branch(engine, 2)
for day in range(80):
    scenario.apply(stress, day)
    ...                                  # run both branches

A scenario is a file of changes to the market and the assumptions behind them. Each change targets a field the engine reads, and the file keeps the shock apart from the knock-on effects you assume follow it:

tradefloor scenario show oil_price_spike

Exogenous shocks
----------------------------------------------------------
  day 50+            commodity.oil            x1.4

Assumed transmission
----------------------------------------------------------
  day 55..74 ramp    macro.inflation          +1.50pp
  day 55+            macro.corporate_yield    +0.50pp

at in a scenario counts days from the first day it is applied, so on a branch it counts from the branch. Most packaged files first fire on day 50. To fire one on the first day after a fork, use scenario.starting_at(0), or world.apply(scenario, at=0) on a World. The gaps between its events stay the same, and the run's record keeps the packaged file's fingerprint and the days each event fired.

tradefloor does not predict what a war, an election, an oil shock or a recession will do to markets. You state the assumptions and it measures how an agent behaves under them. tradefloor scenario list names the seven packaged scenarios, and tradefloor scenario targets lists every field a scenario can change.

Reproducibility

The same seed gives the same market on every platform. tradefloor ships its own exp, log, pow, sin and cos, so the system's math library cannot change a result, and each release runs a fixed simulation on five platforms and stops if any result differs.

A shipped preset never changes, so a market with no agent orders in it replays exactly on its named preset in every later release. Each release checks that with a digest per preset. A run with agent orders in it replays exactly on the same release. Across releases the promise is narrower. 0.8.5 changed how an agent's fills reach the market, on every preset, so a traded run recorded before 0.8.5 matches up to its first trade and differs after it. The default preset is pt-v21, and any earlier one can be named:

eng = tf.Engine(seed=42, universe=u, model="pt-v10")

To let a reader rerun a result, publish its RunManifest. It records the version, preset, seed, universe, macro state and scenario, and reproduce() stops on a mismatch. A manifest checks the market and carries no score. Its result block holds the market's digest, the number of days and draws_consumed. tf.evaluate and tf.rank write no manifest, so a published score has to be rerun to be checked. docs/REPRODUCIBILITY.md has the full contract, including what a saved engine state promises when it is restored, and docs/SUPPORT.md says which release to pin for a long study.

Realism

tradefloor checks its market against real ones with three named sets of statistics, listed in docs/STATISTICS.md. On pt-v21, the default from 0.10.0, all 19 statistics of the one-year table (volatility, fat tails, how much stocks move together, how far the VIX jumps after a fall) are inside the range real markets show over a year, on the ruled bands. The index's return and the share of days it falls 3% or more are read over 360 seeds, as the long-run criteria read them: the index falls 3% or more on 0.98% of days, against a range of 0.64 to 2.34. All 15 statistics of the two-year panel are inside their two-year ranges. The long-run criteria are 40 rows over 21 years, covering crash depth, how long fear lasts, bear markets per decade, the 2008 and 2020 replays, the rate indices and the cost of size in the book. pt-v21 meets all 40, read on 270 histories.

Read those claims narrowly:

• The one-year count is a verdict on figures pooled over 30 seeds (360 for the index's return and 3% falls). One seed's year often misses some of its 14 shape statistics. On seeds 101 to 116, all 14 were in range on 8 of the 16, and one seed had 11 of 14. If you run one market per condition, read tf.envelope.intervals() for each statistic's spread across seeds.
• A shape statistic's range is the median of 35 real one-year windows plus or minus 2.1 trimmed standard deviations, so passing one is weak evidence. Volatility clustering is one case. abs_return_acf5 reads 0.020, below every real 2015 to 2025 window (the lowest is 0.034), and it passes because its range reaches lower than those windows do.
• The VIX is stickier than real: its day-to-day persistence reads 0.956 against the tape's 0.930, and the sign test that grades it refuses it.
• The one-year table helped choose many of the coefficients pt-v21 keeps, so the held-out checks are the fresh seeds and the fresh set of companies the panel is repeated on.
• One year is the certified horizon. Two years is graded on the two-year panel, and longer runs only by the long-run criteria.
• A driven scenario moves prices at about half to four-fifths of the real size, in the right direction. Use a scenario to detect a response, and do not read its size as a forecast.
• Volatility memory reads 0.095 at lag 1 against a real 0.107 and fades much faster after it: at lag 20 it is under a twentieth of real. Nothing below the 65-minute step is calibrated.
• An order sliced over a day costs a median 0.13 of a daily standard deviation for 10% of a day's volume in 36 slices, at the low end of the 0.15 to 0.3 published studies find.
• By default no other trader reacts to you, so no liquidity spiral or predatory trading can arise. Populated mode adds background traders that trade your signals and front-run predictable flow, but they charge a predictable programme about 2.4% more, far below what real markets show.

tf.envelope.check(horizon_days=...) refuses a question that falls outside a measured limit. docs/REALISM.md has every number behind these claims and the full table of limits.

Before you publish a result

• An agent scored on naming the factor behind each day's move gets an explanation_accuracy. On pt-v21 a constant answer scores 0.70 to 0.95, so quote explanation_edge, the accuracy minus that baseline, and never the accuracy alone.
• Agents in one tf.evaluate or tf.rank call run one after another in one Python process, on one seed. An earlier agent can leave the price path in a class variable for a later one. An agent written to cheat can read the seed from the harness's frames through sys._getframe and run a copy of the market ahead. Nothing flags either. The read-only market view guards only against accidents, so run each agent you did not write in its own process, through the MCP server.
• Every evaluate and rank run starts at day 0, so a rule that needs 20 days of prices sits out the first 20 while buy-and-hold is invested. Pass history_days=20 to run the market 20 days first with nobody trading.
• In a World with several agents, orders placed at the same step execute in label order, alphabetical, for the whole run. Rotate the labels across runs when you compare different agents in one market.
• There are no commissions, no borrow fee on a short and no stop orders. A stop you check at each step fills a median 26.5 bp past its level at six steps a day. Uninvested cash earns nothing unless you pass cash_interest=True, and a negative cash balance pays the policy rate, which is below a broker's margin rate.

docs/AGENTS.md has the measurements behind each of these.

Examples

The twelve numbered examples/ are in reading order, and the test suite runs them:

The rate-shock/ study is the demo above, and integrations/ runs the same kind of experiment through each agent framework, offline and without an API key.

Documentation

https://docs.tradefloor.dev covers install, the API, the guides and how the model is measured. These pages in this repository have the detail behind the sections above:

• docs/MODEL.md: the model as equations, with every coefficient's value on the default preset and where it came from
• docs/STATISTICS.md: the named sets of realism statistics
• docs/REALISM.md: the realism results and every measured limit
• docs/REPRODUCIBILITY.md: what replays exactly, across platforms, releases and restored state
• docs/AGENTS.md: writing, scoring and comparing agents
• docs/SUPPORT.md: which release lines get fixes, and for how long
• docs/EMBEDDING.md: driving the Rust engine from a game or another host, and how much flow of its own the model takes

Contributing and support

CONTRIBUTING.md explains how to build and test the project. Its main rule is that any change to the simulated trajectory is a breaking change, however small, so a model change ships as a new preset. RELEASING.md is the release checklist.

Report a vulnerability through GitHub's security advisory form, not a public issue. SECURITY.md says what is in scope. Bugs and questions go to GitHub issues.

0.8.5 and the 0.8 patches after it are the long-term support line, with bug and security fixes for 24 months. docs/SUPPORT.md says what a support line promises and which release to pin for a long study.

Citing tradefloor

Cite the version you ran and name the preset. The same version can run several presets, and results depend on the preset.

@software{tradefloor,
  author  = {Coombes, Simon},
  title   = {tradefloor: a deterministic market simulator with a limit order book},
  version = {0.10.2},
  year    = {2026},
  url     = {https://github.com/simoncoombes/tradefloor},
  note    = {Model preset pt-v21}
}

CITATION.cff carries the same details, and GitHub's "Cite this repository" button reads it.

In the text, say which model you used, for example: "tradefloor 0.10.2, preset pt-v21, specified in its docs/MODEL.md". docs/REPRODUCIBILITY.md says how to publish a result so a reader can rerun it, and how to show a score was not tuned to its seeds.

License

tradefloor is licensed under MIT OR Apache-2.0, at your option. See LICENSE-MIT and LICENSE-APACHE. GitHub's sidebar reads Apache-2.0 because its license detection picks one file and stops. The grant that applies is the dual one, stated in pyproject.toml, rust/Cargo.toml and this section.