Quantitative Finance · Experimental Method · Open Source
Discover what works in the real world.
An open platform for empirical finance. Learn the methods, experiment against hidden holdouts and live markets, and contribute to a shared research commons — code, replications, and wikis, all in the open.
Reality is the final test set.
Work in Python, R, or Julia — the same hidden-holdout engine scores all three.
Three ways in
Learn it. Experiment with it. Contribute to it.
ConvexPi is a curriculum, an open-source research toolkit, and a community — all built around the same question: does a signal survive once the data is no longer yours to fit?
Learn
Read the field, honestly
A simulation-first curriculum and a research library that is candid about which anomalies survive and which decayed after publication.
Experiment
Run it — or compete
Recompute strategies in your browser, test them on a hidden holdout, and pit live agents against each other in a simulated market.
Contribute
Improve the commons
Everything is open: edit a paper wiki, submit a strategy replication by pull request, or build on the packages that power the platform.
Community showcase
Featured projects
Strategy write-ups published from GitHub, run and graded by ConvexPi.
★Anatomy of a Real Crypto Order Book
Spread, depth, and slippage measured on a real recorded BTC-USD order book from Coinbase — the microstructure that sets your true trading costs.
★Momentum + Value — a Strategy That Survives Out of Sample
Short-horizon momentum combined with value, in a wide dollar-neutral book — a real edge that holds up out of sample (most of the time).
★Multi-Factor Long/Short Equity
The classic dollar-neutral combination of value, quality, and momentum — the flagship long-short format.
The Problem
In-sample performance is not evidence.
Fit enough parameters to historical data and any strategy looks profitable. The only meaningful test is out-of-sample: data the strategy never touched during development. ConvexPi grades you the same way peer reviewers should — on a holdout you cannot see.
In-sample Sharpe
IS Sharpe
Measures fit to historical data. Easy to maximise through overfitting. Does not predict future performance.
Out-of-sample Sharpe
OOS Sharpe
Evaluated on a holdout market with a secret seed. You cannot overfit data you cannot observe.
Research & Replication
The factor zoo is a graveyard.
We track 217 equity anomalies pre- and post-publication using the Open Source Asset Pricing dataset (211predictors, 1926–2024) plus Kenneth French's flagship factors — alongside a library of thousands of finance papers with structured wikis, and an open package of reference replications that recompute each strategy and score it out of sample. Most anomalies are real in-sample — then fade out of sample as the world catches on. A few survive; many decay. That gap is the whole point.
| Anomaly | IS Sharpe | OOS Sharpe | Status |
|---|---|---|---|
| Market | 0.47 | 0.44 | weakened |
| Size (SMB) | 0.20 | -0.02 | dead |
| Value (HML) | 0.47 | 0.25 | weakened |
| Momentum (Mom) | 0.63 | 0.38 | weakened |
| Profitability (RMW) | 0.56 | 0.29 | weakened |
| Investment (CMA) | 0.69 | -0.16 | dead |
Open Source
ConvexPi is not just a website. It's an open stack.
The same tools that power the curriculum — synthetic markets, walk-forward backtests, hidden-holdout grading, replication, and a live exchange simulator — are open-source. Inspect the machinery, run it yourself, or extend it.
convexpi/lab↗Synthetic equity panels, walk-forward backtests, hidden-holdout grading, anomaly replication, real-data mode.
convexpi/convexpi-r↗R client — write strategies (and Arena agents) in R, scored by the same engine as Python and Julia.
convexpi/ConvexPi.jl↗Julia client — write strategies (and Arena agents) in Julia, scored by the same engine as Python and R.
convexpi/arena↗Discrete-time limit-order-book simulator for deploying live trading and market-making agents.
convexpi/finmlsim↗Scholarly market, microstructure, and execution simulators (GARCH, Heston, Kyle, Almgren-Chriss) powering the realistic market modes.
convexpi/replications↗Verified reference replications of canonical strategies — recomputed from building blocks, scored out of sample, CI-checked.
convexpi/content↗Community-edited paper wikis and research notes — anyone can suggest an edit via GitHub.
convexpi/missions↗The mission notebooks and assignments behind the curriculum.
convexpi/platform↗The web application, grader, and research library that runs convexpi.ai.
pip install convexpi-lab Browse the GitHub organization →Experiment
From a browser tab to a live market.
Three ways to put a strategy to the test — each grading the same question.
Playground
Run it in your browser
Recompute real strategies — reconstruct the Fama-French factors from their building blocks, form a long-short book — entirely client-side, then test out of sample. No install.
- Python in the browser via Pyodide
- Recomputes strategies, not finished factors
- Open in Colab for full, live-data runs
Lab
Alpha Discovery Lab
Write strategies against a synthetic equity panel with planted factor signals. Nine structured missions take you from naive overfitting to disciplined out-of-sample evaluation, with electives on microstructure and trading costs.
- Write in Python, R, or Julia — same engine scores all
- IS / OOS breakdown on every submission
- Hidden-holdout grading you cannot game
Arena
Live Market Arena
Deploy agents that submit live orders into a limit-order-book simulation. Study adverse selection, inventory risk, and spread dynamics under competitive pressure.
- WebSocket agent API — connect in Python
- Ongoing open ladder + rolling seasons
- Survival score rewards robustness over Sharpe
Contribute
Built in the open, improved together.
Finance education usually hides the grader, the data, and the benchmark. ConvexPi does the opposite — and invites you to make it better. Every wiki, replication, and result is public and editable.
Edit a paper wiki
Every paper page has a structured, community-maintained wiki. Suggest an edit on GitHub — the revision history is a Wikipedia-style log.
Browse papers →Submit a replication
Add a strategy to the open replication library. Recompute it from building blocks, score it out of sample; CI checks it reproduces, then a maintainer merges your PR.
Contribution guide ↗Enter a competition
An ongoing open ladder and rolling monthly seasons. Submit a strategy or a live agent and see your honest out-of-sample result on the public leaderboard.
Open competitions →Build on the packages
The Lab, Arena, and replication toolkits are MIT-licensed. Use them in your own research or course, and send improvements upstream.
GitHub organization ↗Earn reputation
Replications, wiki edits, and strategies that survive out of sample earn reputation and badges — including Ghostbuster, for replicating a factor that turned out dead. Climb the contributors leaderboard.
Contributors leaderboard →Teach with it
Run a private cohort: assign missions as problem sets, grade on hidden-holdout Sharpe, and export gradebooks. A University of Cincinnati cohort runs in Fall 2026.
Create a cohort →Start a mission, run the code, or contribute.
Whether you want to learn quantitative finance, test a signal of your own, or help build an open research commons — there is a door for you. Mission 1 takes about thirty minutes.
Open source · MIT License · Python · convexpi-lab on PyPI