~kris/dots

srice

ref: 4471d164e01750db50f2d73d2a7c8af0e2318591 srice/doc/trading/research-quant-oss.md -rw-r--r-- 7.3 KiB
4471d164 — Kris Yotam wallpapermenu: pass -no-config so ambient rofi theme can't override the picker 2 months ago

#title: "Open-Source Repos from Quant Funds and Banks" date: 2026-05-21 source: agent research

#Open-Source from Quant Funds and Banks

106 repositories cataloged across 9 major firms. Sorted by practical utility for a solo systematic trader running VectorBT + NautilusTrader + TimescaleDB.

#Tier 1: Immediately Useful

Repo Stars Language License What It Does
man-group/ArcticDB 2,320 C++/Python BSL (free non-commercial) Serverless time-series database. Billions of rows in S3. Pandas in, pandas out. Bloomberg licensed this.
man-group/dtale 5,144 Python Apache 2.0 Interactive pandas DataFrame explorer in browser. One line of code. Filters, correlations, charts.
goldmansachs/gs-quant 10,358 Python Apache 2.0 Quantitative finance toolkit. Market data, risk models, analytics. Most complete firm-backed trading library.
man-group/notebooker 900 Python Apache 2.0 Turns Jupyter into scheduled reporting engine. Daily analytics runs.
deshaw/pyflyby 409 Python BSD Auto-import for IPython/Jupyter. Saves hours across 50 notebooks/day.

#Tier 2: Useful for Specific Workflows

Repo Stars Language License What It Does When to Use
man-group/PyBloqs 183 Python Apache 2.0 HTML reports from Python. Tables, charts, layout in blocks. Generating tearsheets/reports for review.
deshaw/versioned-hdf5 89 Python BSD Version control for HDF5 files (git for scientific data). Rolling back to "yesterday's version" of a dataset.
deshaw/nbstripout-fast 29 Rust BSD Strips outputs from Jupyter notebooks before git commits. Fast. Git hygiene for research notebooks.
optiver/timestamp9 65 Python Apache 2.0 Nanosecond timestamps for Python. When tick data needs sub-millisecond precision.
optiver/optiver-asyncpg N/A Python Apache 2.0 Production fork of asyncpg (fast Postgres client for asyncio). Writing ticks at latencies psycopg2 can't reach.
yli188/WorldQuant_alpha101_code 748 Python MIT Community implementation of all 101 alpha formulas from WorldQuant's "101 Formulaic Alphas" paper. Factor research starting point.

#Tier 3: Interesting but Not Directly Actionable

#Two Sigma

Repo Stars Language What It Does Why Not Now
twosigma/beakerx 2,800 Java/Python Multi-language Jupyter (Python, Scala, Groovy, Kotlin, Clojure, Java). Standard Jupyter suffices.
twosigma/flint 1,000 Scala Time-series joins on Apache Spark with temporal tolerance. Overkill; DuckDB handles your scale.
twosigma/cook 338 Clojure Batch job scheduler on Mesos/Kubernetes. systemd is fine at single-server scale.
twosigma/marbles 116 Python Unit tests that explain failures in plain English. Nice-to-have, not critical path.
twosigma/frost N/A SystemVerilog FPGA RISC-V core. FPGA hardware design; not relevant to retail.

#Jane Street (OCaml Ecosystem)

Repo Stars What It Does Why Not Now
janestreet/core 1,200 Alternative OCaml standard library. OCaml-only. Not useful for Python stack.
janestreet/magic-trace 5,300 High-resolution process tracer via Intel Processor Trace. Useful if you ever profile at CPU-instruction level.
janestreet/async 233 Cooperative concurrency in OCaml. OCaml-only.
janestreet/hardcaml 1,000 OCaml library for hardware design (FPGA, ASIC). FPGA design; academic interest only.

#Hudson River Trading

Repo Stars What It Does Why Not Now
hudson-trading/corral 175 Structured concurrency for C++20. C++ infrastructure; not relevant to Python stack.
hudson-trading/slang-server 224 SystemVerilog language server. FPGA tooling.
hudson-trading/heracles-ql 27 Python DSL for alerts. Too small; custom Prometheus alerts are better.

#D.E. Shaw

Repo Stars What It Does Why Not Now
deshaw/pjrmi 47 RPC between Python and Java. No Java in your stack.

#Banks

#Goldman Sachs

#JPMorgan Chase

#Morgan Stanley

  • No significant public repositories for trading.

#The Silent Ones (Zero Public Repos)

Firm AUM Status
Renaissance Technologies ~$130B Zero. NDAs prevent even detailed resumes.
Citadel / Citadel Securities ~$65B No official GitHub account.
Bridgewater Associates ~$100B Zero.
Millennium Management ~$70B Zero.
Point72 / Cubist Systems ~$35B Zero.
AQR Capital ~$99B Has an account; contains only a pandas fork. Effectively nothing. 
Data Storage:     ArcticDB (serverless, Man Group, optimized for time-series)
Data Exploration: dtale (Man Group, interactive DataFrames)
Analytics:        gs-quant (Goldman Sachs, risk models + market data)
Reporting:        notebooker (Man Group, scheduled Jupyter reports)
Factor Research:  WorldQuant alpha101 (community, 101 formulaic alphas)
Jupyter DX:       pyflyby (D.E. Shaw, auto-imports)
Git Hygiene:      nbstripout-fast (D.E. Shaw, Rust, fast notebook stripping)

#Why Some Firms Share and Others Don't

Three patterns:

  1. Recruiting brand (Two Sigma, D.E. Shaw, Jane Street, HRT): Public code attracts PhDs. magic-trace's 5,300 stars = every infrastructure engineer now knows where to send a CV.

  2. Marketing after IP is already licensed (Man Group): Bloomberg already paid for ArcticDB. The rest is marketing for a publicly listed company.

  3. Any public code = hint to competitors (Renaissance, Citadel, Bridgewater): Renaissance doesn't even allow former employees to write detailed resumes about their work.

For solo traders: The "recruiting brand" firms produce the most useful tools because they're designed to be general-purpose infrastructure, not alpha-generating code. Alpha stays private everywhere.