pretium

A market simulator you can run a strategy against. Rust core, Python API.

Give it a seed and a set of companies and it runs a market forward: prices, a limit order book, fills, and an economy that advances itself daily. Your orders match against real depth, so trading moves the price the way it would anywhere else.

It exists because historical data can't answer two questions. What would have happened if I'd traded differently, and what actually caused this price move.

Install

pip install pretium

Wheels are prebuilt for Linux, macOS and Windows on CPython 3.11+, and the core package has no dependencies.

The engine is also published as a Rust crate, so you can use it without Python at all:

cargo add pretium

Both are built from the same source at the same version, and run the same market: https://crates.io/crates/pretium.

The API may still move before 1.0. What will not move is a published result: anything that changes the simulated trajectory arrives as a new model preset rather than an edit to an existing one, so a run you cited last month reproduces this month.

In thirty seconds

import pretium as pt

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

spec = pt.StrategySpec.momentum(lookback_days=1.0, top_k=5)
scores = pt.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

That's one market draw, which tells you as much about the seed as about the strategy. pt.rank runs many seeds and compares them with a paired sign test.

What makes it different

Determinism that's checked. The crate ships its own exp, log, sin and cos instead of calling the platform libm. Every release builds five targets -- linux-x86_64, linux-aarch64, macos-arm64, macos-x86_64 and windows-x86_64 -- runs one fixed simulation inside each, and compares digests. A disagreement fails the release.

The WebAssembly build is checked by hand rather than by that gate, because it needs a toolchain CI does not carry: on 2026-08-24 the native and wasm32-unknown-unknown builds hashed the same fixed simulation -- twelve instruments, five days, pt-v3 -- to 2b2f3141..., and tests/test_wasm_parity.py pins the Python half of that pair so the two cannot drift apart unnoticed.

Ground truth you can read. The simulator computed every price, so it can tell you why. One row per instrument per tick, with nine factor contributions that sum to the move and a residual around 1e-16. No historical dataset carries those labels. You can see that a stock fell, never that 60% of the fall was order-flow pressure.

Counterfactuals you can run. The same seed runs with and without your orders, so every fill is priced against the market where you never traded. That's the number TCA vendors approximate.

Results someone else can check. A StrategySpec is a declarative, hashable document rather than an arbitrary callable. A RunManifest carries the seed, universe, macro conditions, scenario, model and strategy together with the expected digest. reproduce() refuses on a mismatch and names the culprit, because a manifest that quietly reproduced a different market would be worse than no manifest.

Driving it from an agent

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

Eleven read-only tools over the simulator, so a coding agent can ask whether a momentum strategy beats buy-and-hold here, and whether the difference is real. Strategies, universes and scenarios are composed as data. There's no path from a tool argument to code execution.

Every result carries computed caveats and full provenance. A model summarising a result has the tool output and nothing else, and will otherwise report return_pct: 88.7 as "the strategy made 88.7%". The MCP server has the tool list and the client configuration.

What it's bad at

Good results here don't predict real returns. The price process comes from a known model, so a strategy that fits that model's structure will look excellent and teach you nothing. A strategy that fails here is more informative, because it broke against a live order book under honest impact costs.

Realism is a stated envelope, not a score. pt.facts.measure() reports fourteen statistics against real-market bands. At 252 days the shipped pt-v12 preset holds all fourteen in band on thirty calibration seeds, all fourteen again at 504 days against bands re-derived at that window, and all fourteen again on a 60-name universe it never saw. On thirty seeds it never used it holds thirteen, the single miss being corr_persistence_acf1. Read that as fourteen in-band verdicts rather than fourteen independent validations: five of the fourteen were live calibration targets, the bands were used both to tune against and to grade against, and "held out" means unseen simulation seeds and another roster from the same generator rather than withheld market data. That is a market with the right volatility, the right tails, the right co-movement, industries that co-move more than strangers, correlation that stays elevated after a panic, volume that behaves, and volatility episodes it produces itself.

pt-v12 became the default on 2026-08-26, an era boundary: every trajectory that came from the default changed. Runs recorded before it are not comparable to runs after it unless they name their preset, and every earlier preset from pt-v1 onward stays selectable and reproduces bit for bit. Checked rather than asserted: pt-v1, pt-v2 and pt-v3 give identical market digests under 0.1.4 and 0.2.0. A fully specified run is still (package version, model, universe fingerprint, seed), since the preset pins the coefficients and the package version pins the implementation.

At 504 days pt-v12 holds all fourteen, which is the first two-year clean sheet this project has measured -- pt-v3 held seven there and pt-v10 held thirteen, missing volume-change autocorrelation. Read the headroom rather than the count: annualised volatility reads 33.89 against a band ending at 34.0, so that row is genuine but thin, and the certified horizon stays 252 days because that is where the certification is measured.

Beyond two years the model has now been measured rather than assumed. At 2520 days it holds ten of the fourteen against the (horizon-mismatched) two-year bands, and nothing runs away or drifts: measured year by year over ten years, annualised volatility reads 31.5, 35.6, 30.2, 33.5, 33.0, 33.1, 31.3, 32.4, 32.4, 31.6 percent -- flat, with the year-to-year variation a real market has.

Five gaps are measured and named rather than assumed. The certified horizon is 252 days; two and five-year panels are measured but not certified. Volatility memory decays exponentially where real markets decay hyperbolically. Scenario response has the right size, measured as a regression gain, but too much noise around it for one run to size a scenario from -- the driven-window noise ratio is 1.57x real. The endogenous MACRO economy cannot reach its own crisis regimes, so an inflation regime or a policy crisis has to be driven through a scenario; a volatility crisis does not: measured on pt-v10, the preset's own VIX crossed its crisis threshold on 10.2% of days against a real 12.5%, and that figure has not been re-measured on pt-v12. And certification was measured on a sector-balanced roster, which no real index is: re-measured on pt-v12, an S&P-like or technology-heavy roster holds all fourteen at one year, so the envelope transfers at the certified horizon, and the cost of concentration shows up in the second year instead. The realism envelope states each gap and what it forbids. pt.envelope.check() refuses to certify a question that falls outside one.

Single venue, no latency, no strategic counterparties. Orders arrive instantly, there's one book per name, and you trade against a market maker and aggregate flow rather than agents that adapt to you.

Worked examples

Eight notebooks and two scripts in examples/, numbered in reading order and executed as part of the test suite:

The two scripts run from the command line rather than a notebook. The whole study takes about five seconds:

python examples/07-research-workflow.py

Documentation

Full docs: simoncoombes.github.io/pretium

Getting started and core concepts. Agents, evaluation and the RL environment. Scenarios and macro paths, strategy specs, model presets, checkpointing and forking, sharing a run, Arrow output and streaming sweeps, SEC EDGAR loading, transaction cost analysis, the MCP server, running in the browser, the realism envelope and metrics, and the conventions worth reading before you hit them.

Contributing

See CONTRIBUTING.md, and RELEASING.md for how a version reaches PyPI and crates.io. One rule shapes the rest: a change to the simulated trajectory is a breaking change, whatever its size. A market that runs differently from the same seed invalidates every published result that cited it.

Citing this

See CITATION.cff, or cite a specific result by its manifest. A citation identifies the software; it does not identify a run. A RunManifest does: it carries the package version and preset it was written by, the seed, the universe fingerprint, the macro conditions, the scenario and the strategy fingerprint.

Licence

MIT OR Apache-2.0, at your option. See LICENSE-MIT and LICENSE-APACHE.